{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 119, 97, 492, 117 ], "lines": [ { "bbox": [ 118, 97, 492, 120 ], "spans": [ { "bbox": [ 118, 97, 492, 120 ], "score": 1.0, "content": "Any-to-Any Generation via Composable Diffusion", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 113, 159, 175, 171 ], "lines": [ { "bbox": [ 111, 157, 178, 173 ], "spans": [ { "bbox": [ 111, 157, 178, 173 ], "score": 1.0, "content": "Zineng Tang1∗", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "title", "bbox": [ 432, 159, 498, 171 ], "lines": [ { "bbox": [ 431, 158, 500, 171 ], "spans": [ { "bbox": [ 431, 158, 500, 171 ], "score": 1.0, "content": "Mohit Bansal1†", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 190, 158, 414, 205 ], "lines": [ { "bbox": [ 189, 158, 414, 172 ], "spans": [ { "bbox": [ 189, 158, 244, 171 ], "score": 1.0, "content": "Ziyi Yang2†", "type": "text" }, { "bbox": [ 255, 158, 307, 171 ], "score": 1.0, "content": "Chenguang", "type": "text" }, { "bbox": [ 307, 158, 335, 171 ], "score": 0.42, "content": "{ \\bf Z } { \\bf h } { \\bf u } ^ { 2 \\ddagger }", "type": "inline_equation" }, { "bbox": [ 345, 158, 414, 172 ], "score": 1.0, "content": "Michael Zeng2", "type": "text" } ], "index": 2 }, { "bbox": [ 215, 169, 397, 184 ], "spans": [ { "bbox": [ 215, 169, 397, 184 ], "score": 1.0, "content": "1University of North Carolina at Chapel Hill", "type": "text" } ], "index": 4 }, { "bbox": [ 210, 180, 401, 195 ], "spans": [ { "bbox": [ 210, 180, 401, 195 ], "score": 1.0, "content": "2Microsoft Azure Cognitive Services Research", "type": "text" } ], "index": 5 }, { "bbox": [ 236, 193, 375, 205 ], "spans": [ { "bbox": [ 236, 193, 375, 205 ], "score": 1.0, "content": "https://codi-gen.github.io", "type": "text" } ], "index": 6 } ], "index": 4.5 }, { "type": "title", "bbox": [ 283, 233, 328, 246 ], "lines": [ { "bbox": [ 281, 232, 331, 248 ], "spans": [ { "bbox": [ 281, 232, 331, 248 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 142, 261, 469, 424 ], "lines": [ { "bbox": [ 142, 260, 470, 273 ], "spans": [ { "bbox": [ 142, 260, 470, 273 ], "score": 1.0, "content": "We present Composable Diffusion (CoDi), a novel generative model capable of", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 271, 470, 284 ], "spans": [ { "bbox": [ 141, 271, 470, 284 ], "score": 1.0, "content": "generating any combination of output modalities, such as language, image, video,", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 282, 469, 295 ], "spans": [ { "bbox": [ 141, 282, 469, 295 ], "score": 1.0, "content": "or audio, from any combination of input modalities. Unlike existing generative", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 293, 470, 305 ], "spans": [ { "bbox": [ 141, 293, 470, 305 ], "score": 1.0, "content": "AI systems, CoDi can generate multiple modalities in parallel and its input is not", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 303, 469, 316 ], "spans": [ { "bbox": [ 141, 303, 469, 316 ], "score": 1.0, "content": "limited to a subset of modalities like text or image. Despite the absence of training", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 315, 469, 327 ], "spans": [ { "bbox": [ 141, 315, 469, 327 ], "score": 1.0, "content": "datasets for many combinations of modalities, we propose to align modalities in", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 326, 469, 338 ], "spans": [ { "bbox": [ 141, 326, 469, 338 ], "score": 1.0, "content": "both the input and output space. This allows CoDi to freely condition on any input", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 336, 469, 349 ], "spans": [ { "bbox": [ 141, 336, 469, 349 ], "score": 1.0, "content": "combination and generate any group of modalities, even if they are not present in", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 348, 470, 360 ], "spans": [ { "bbox": [ 141, 348, 470, 360 ], "score": 1.0, "content": "the training data. CoDi employs a novel composable generation strategy which", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 358, 470, 371 ], "spans": [ { "bbox": [ 141, 358, 470, 371 ], "score": 1.0, "content": "involves building a shared multimodal space by bridging alignment in the diffusion", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 370, 469, 381 ], "spans": [ { "bbox": [ 141, 370, 469, 381 ], "score": 1.0, "content": "process, enabling the synchronized generation of intertwined modalities, such", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 380, 469, 392 ], "spans": [ { "bbox": [ 141, 380, 469, 392 ], "score": 1.0, "content": "as temporally aligned video and audio. Highly customizable and flexible, CoDi", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 391, 470, 404 ], "spans": [ { "bbox": [ 141, 391, 470, 404 ], "score": 1.0, "content": "achieves strong joint-modality generation quality, and outperforms or is on par", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 401, 469, 415 ], "spans": [ { "bbox": [ 141, 401, 469, 415 ], "score": 1.0, "content": "with the unimodal state-of-the-art for single-modality synthesis. The project page", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 412, 425, 426 ], "spans": [ { "bbox": [ 141, 412, 425, 426 ], "score": 1.0, "content": "with demonstrations and code is at https://codi-gen.github.io/", "type": "text" } ], "index": 22 } ], "index": 15 }, { "type": "image", "bbox": [ 108, 448, 502, 613 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 448, 502, 613 ], "group_id": 0, "lines": [ { "bbox": [ 108, 448, 502, 613 ], "spans": [ { "bbox": [ 108, 448, 502, 613 ], "score": 0.97, "type": "image", "image_path": "f6d7a60ecba1b89ddec9c2b8b8bfea37a1e8ff89e67fccf04a7b8dcb4427834d.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 108, 448, 502, 503.0 ], "spans": [], "index": 23 }, { "bbox": [ 108, 503.0, 502, 558.0 ], "spans": [], "index": 24 }, { "bbox": [ 108, 558.0, 502, 613.0 ], "spans": [], "index": 25 } ] }, { "type": "image_caption", "bbox": [ 106, 633, 506, 655 ], "group_id": 0, "lines": [ { "bbox": [ 106, 633, 506, 645 ], "spans": [ { "bbox": [ 106, 633, 506, 645 ], "score": 1.0, "content": "Figure 1: CoDi can generate various (joint) combinations of output modalities from diverse (joint)", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 643, 506, 656 ], "spans": [ { "bbox": [ 106, 643, 506, 656 ], "score": 1.0, "content": "sets of inputs: video, image, audio, and text (example combinations depicted by the colored arrows).", "type": "text" } ], "index": 27 } ], "index": 26.5 } ], "index": 25.25 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 118, 677, 384, 710 ], "lines": [ { "bbox": [ 117, 676, 287, 689 ], "spans": [ { "bbox": [ 117, 676, 287, 689 ], "score": 1.0, "content": "∗Work done at Microsoft internship and UNC.", "type": "text" } ] }, { "bbox": [ 118, 687, 384, 700 ], "spans": [ { "bbox": [ 118, 687, 384, 700 ], "score": 1.0, "content": "†Corresponding authors: ziyiyang@microsoft.com, mbansal@cs.unc.edu", "type": "text" } ] }, { "bbox": [ 118, 698, 235, 711 ], "spans": [ { "bbox": [ 118, 698, 235, 711 ], "score": 1.0, "content": "‡Work done while at Microsoft.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 731, 385, 742 ], "lines": [ { "bbox": [ 105, 730, 386, 744 ], "spans": [ { "bbox": [ 105, 730, 386, 744 ], "score": 1.0, "content": "37th Conference on Neural Information Processing Systems (NeurIPS 2023).", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 119, 97, 492, 117 ], "lines": [ { "bbox": [ 118, 97, 492, 120 ], "spans": [ { "bbox": [ 118, 97, 492, 120 ], "score": 1.0, "content": "Any-to-Any Generation via Composable Diffusion", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 113, 159, 175, 171 ], "lines": [ { "bbox": [ 111, 157, 178, 173 ], "spans": [ { "bbox": [ 111, 157, 178, 173 ], "score": 1.0, "content": "Zineng Tang1∗", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "title", "bbox": [ 432, 159, 498, 171 ], "lines": [ { "bbox": [ 431, 158, 500, 171 ], "spans": [ { "bbox": [ 431, 158, 500, 171 ], "score": 1.0, "content": "Mohit Bansal1†", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 190, 158, 414, 205 ], "lines": [ { "bbox": [ 189, 158, 414, 172 ], "spans": [ { "bbox": [ 189, 158, 244, 171 ], "score": 1.0, "content": "Ziyi Yang2†", "type": "text" }, { "bbox": [ 255, 158, 307, 171 ], "score": 1.0, "content": "Chenguang", "type": "text" }, { "bbox": [ 307, 158, 335, 171 ], "score": 0.42, "content": "{ \\bf Z } { \\bf h } { \\bf u } ^ { 2 \\ddagger }", "type": "inline_equation" }, { "bbox": [ 345, 158, 414, 172 ], "score": 1.0, "content": "Michael Zeng2", "type": "text" } ], "index": 2 }, { "bbox": [ 215, 169, 397, 184 ], "spans": [ { "bbox": [ 215, 169, 397, 184 ], "score": 1.0, "content": "1University of North Carolina at Chapel Hill", "type": "text" } ], "index": 4 }, { "bbox": [ 210, 180, 401, 195 ], "spans": [ { "bbox": [ 210, 180, 401, 195 ], "score": 1.0, "content": "2Microsoft Azure Cognitive Services Research", "type": "text" } ], "index": 5 }, { "bbox": [ 236, 193, 375, 205 ], "spans": [ { "bbox": [ 236, 193, 375, 205 ], "score": 1.0, "content": "https://codi-gen.github.io", "type": "text" } ], "index": 6 } ], "index": 4.5, "bbox_fs": [ 189, 158, 414, 205 ] }, { "type": "title", "bbox": [ 283, 233, 328, 246 ], "lines": [ { "bbox": [ 281, 232, 331, 248 ], "spans": [ { "bbox": [ 281, 232, 331, 248 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 142, 261, 469, 424 ], "lines": [ { "bbox": [ 142, 260, 470, 273 ], "spans": [ { "bbox": [ 142, 260, 470, 273 ], "score": 1.0, "content": "We present Composable Diffusion (CoDi), a novel generative model capable of", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 271, 470, 284 ], "spans": [ { "bbox": [ 141, 271, 470, 284 ], "score": 1.0, "content": "generating any combination of output modalities, such as language, image, video,", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 282, 469, 295 ], "spans": [ { "bbox": [ 141, 282, 469, 295 ], "score": 1.0, "content": "or audio, from any combination of input modalities. Unlike existing generative", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 293, 470, 305 ], "spans": [ { "bbox": [ 141, 293, 470, 305 ], "score": 1.0, "content": "AI systems, CoDi can generate multiple modalities in parallel and its input is not", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 303, 469, 316 ], "spans": [ { "bbox": [ 141, 303, 469, 316 ], "score": 1.0, "content": "limited to a subset of modalities like text or image. Despite the absence of training", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 315, 469, 327 ], "spans": [ { "bbox": [ 141, 315, 469, 327 ], "score": 1.0, "content": "datasets for many combinations of modalities, we propose to align modalities in", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 326, 469, 338 ], "spans": [ { "bbox": [ 141, 326, 469, 338 ], "score": 1.0, "content": "both the input and output space. This allows CoDi to freely condition on any input", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 336, 469, 349 ], "spans": [ { "bbox": [ 141, 336, 469, 349 ], "score": 1.0, "content": "combination and generate any group of modalities, even if they are not present in", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 348, 470, 360 ], "spans": [ { "bbox": [ 141, 348, 470, 360 ], "score": 1.0, "content": "the training data. CoDi employs a novel composable generation strategy which", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 358, 470, 371 ], "spans": [ { "bbox": [ 141, 358, 470, 371 ], "score": 1.0, "content": "involves building a shared multimodal space by bridging alignment in the diffusion", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 370, 469, 381 ], "spans": [ { "bbox": [ 141, 370, 469, 381 ], "score": 1.0, "content": "process, enabling the synchronized generation of intertwined modalities, such", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 380, 469, 392 ], "spans": [ { "bbox": [ 141, 380, 469, 392 ], "score": 1.0, "content": "as temporally aligned video and audio. Highly customizable and flexible, CoDi", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 391, 470, 404 ], "spans": [ { "bbox": [ 141, 391, 470, 404 ], "score": 1.0, "content": "achieves strong joint-modality generation quality, and outperforms or is on par", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 401, 469, 415 ], "spans": [ { "bbox": [ 141, 401, 469, 415 ], "score": 1.0, "content": "with the unimodal state-of-the-art for single-modality synthesis. The project page", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 412, 425, 426 ], "spans": [ { "bbox": [ 141, 412, 425, 426 ], "score": 1.0, "content": "with demonstrations and code is at https://codi-gen.github.io/", "type": "text" } ], "index": 22 } ], "index": 15, "bbox_fs": [ 141, 260, 470, 426 ] }, { "type": "image", "bbox": [ 108, 448, 502, 613 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 448, 502, 613 ], "group_id": 0, "lines": [ { "bbox": [ 108, 448, 502, 613 ], "spans": [ { "bbox": [ 108, 448, 502, 613 ], "score": 0.97, "type": "image", "image_path": "f6d7a60ecba1b89ddec9c2b8b8bfea37a1e8ff89e67fccf04a7b8dcb4427834d.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 108, 448, 502, 503.0 ], "spans": [], "index": 23 }, { "bbox": [ 108, 503.0, 502, 558.0 ], "spans": [], "index": 24 }, { "bbox": [ 108, 558.0, 502, 613.0 ], "spans": [], "index": 25 } ] }, { "type": "image_caption", "bbox": [ 106, 633, 506, 655 ], "group_id": 0, "lines": [ { "bbox": [ 106, 633, 506, 645 ], "spans": [ { "bbox": [ 106, 633, 506, 645 ], "score": 1.0, "content": "Figure 1: CoDi can generate various (joint) combinations of output modalities from diverse (joint)", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 643, 506, 656 ], "spans": [ { "bbox": [ 106, 643, 506, 656 ], "score": 1.0, "content": "sets of inputs: video, image, audio, and text (example combinations depicted by the colored arrows).", "type": "text" } ], "index": 27 } ], "index": 26.5 } ], "index": 25.25 } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 71, 190, 84 ], "lines": [ { "bbox": [ 105, 69, 192, 87 ], "spans": [ { "bbox": [ 105, 69, 192, 87 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 97, 505, 229 ], "lines": [ { "bbox": [ 106, 98, 505, 110 ], "spans": [ { "bbox": [ 106, 98, 505, 110 ], "score": 1.0, "content": "Recent years have seen the rise of powerful cross-modal models that can generate one modality", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 108, 506, 121 ], "spans": [ { "bbox": [ 105, 108, 506, 121 ], "score": 1.0, "content": "from another, e.g. text-to-text [6, 37], text-to-image [13, 19, 22, 41, 44], or text-to-audio [23, 33].", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 119, 505, 132 ], "spans": [ { "bbox": [ 106, 119, 505, 132 ], "score": 1.0, "content": "However, these models are restricted in their real-world applicability where multiple modalities", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 130, 505, 143 ], "spans": [ { "bbox": [ 105, 130, 505, 143 ], "score": 1.0, "content": "coexist and interact. While one can chain together modality-specific generative models in a multi-step", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 141, 506, 154 ], "spans": [ { "bbox": [ 105, 141, 506, 154 ], "score": 1.0, "content": "generation setting, the generation power of each step remains inherently limited, and a serial, multi-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 153, 505, 165 ], "spans": [ { "bbox": [ 106, 153, 505, 165 ], "score": 1.0, "content": "step process can be cumbersome and slow. Moreover, independently generated unimodal streams", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 163, 506, 176 ], "spans": [ { "bbox": [ 105, 163, 506, 176 ], "score": 1.0, "content": "will not be consistent and aligned when stitched together in a post-processing way (e.g., synchronized", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 172, 505, 187 ], "spans": [ { "bbox": [ 105, 172, 505, 187 ], "score": 1.0, "content": "video and audio). The development of a comprehensive and versatile model that can generate any", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 186, 505, 197 ], "spans": [ { "bbox": [ 106, 186, 505, 197 ], "score": 1.0, "content": "combination of modalities from any set of input conditions has been eagerly anticipated, as it would", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 196, 505, 208 ], "spans": [ { "bbox": [ 106, 196, 505, 208 ], "score": 1.0, "content": "more accurately capture the multimodal nature of the world and human comprehension, seamlessly", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 207, 506, 219 ], "spans": [ { "bbox": [ 106, 207, 506, 219 ], "score": 1.0, "content": "consolidate information from a wide range of sources, and enable strong immersion in human-AI", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 218, 506, 230 ], "spans": [ { "bbox": [ 106, 218, 506, 230 ], "score": 1.0, "content": "interactions (for example, by generating coherent video, audio, and text description at the same time).", "type": "text" } ], "index": 12 } ], "index": 6.5 }, { "type": "text", "bbox": [ 107, 234, 505, 343 ], "lines": [ { "bbox": [ 105, 234, 506, 246 ], "spans": [ { "bbox": [ 105, 234, 506, 246 ], "score": 1.0, "content": "In pursuit of this goal, we propose Composable Diffusion, or CoDi, the first model capable of", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 245, 507, 257 ], "spans": [ { "bbox": [ 105, 245, 507, 257 ], "score": 1.0, "content": "simultaneously processing and generating arbitrary combinations of modalities as shown in Fig. 1.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 257, 505, 268 ], "spans": [ { "bbox": [ 106, 257, 505, 268 ], "score": 1.0, "content": "Training a model to take any mixture of input modalities and flexibly generate any mixture of outputs", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 267, 506, 279 ], "spans": [ { "bbox": [ 105, 267, 506, 279 ], "score": 1.0, "content": "presents significant computational and data requirements, as the number of combinations for the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 277, 506, 291 ], "spans": [ { "bbox": [ 105, 277, 506, 291 ], "score": 1.0, "content": "input and output modalities scales exponentially. Also aligned training data for many groups of", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 289, 506, 302 ], "spans": [ { "bbox": [ 105, 289, 506, 302 ], "score": 1.0, "content": "modalities is scarce or even non-existent, making it infeasible to train with all possible input-output", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 299, 506, 312 ], "spans": [ { "bbox": [ 105, 299, 506, 312 ], "score": 1.0, "content": "combinations. To address this challenge, we propose to align multiple modalities in both the input", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 310, 506, 323 ], "spans": [ { "bbox": [ 105, 310, 506, 323 ], "score": 1.0, "content": "conditioning (Section 3.2) and generation diffusion step (Section 3.4). Furthermore, a proposed", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 320, 505, 334 ], "spans": [ { "bbox": [ 105, 320, 505, 334 ], "score": 1.0, "content": "“Bridging Alignment” strategy for contrastive learning (Section 3.2) allows us to efficiently model the", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 333, 481, 345 ], "spans": [ { "bbox": [ 106, 333, 481, 345 ], "score": 1.0, "content": "exponential number of input-output combinations with a linear number of training objectives.", "type": "text" } ], "index": 22 } ], "index": 17.5 }, { "type": "text", "bbox": [ 107, 348, 506, 458 ], "lines": [ { "bbox": [ 106, 349, 506, 361 ], "spans": [ { "bbox": [ 106, 349, 506, 361 ], "score": 1.0, "content": "Building a model with any-to-any generation capacity with exceptional generation quality requires", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 360, 506, 372 ], "spans": [ { "bbox": [ 105, 360, 506, 372 ], "score": 1.0, "content": "comprehensive model design and training on diverse data resources. Therefore, we build CoDi in an", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 369, 507, 385 ], "spans": [ { "bbox": [ 105, 369, 507, 385 ], "score": 1.0, "content": "integrative way. First, we train a latent diffusion model (LDM) for each modality, e.g., text, image,", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 381, 506, 394 ], "spans": [ { "bbox": [ 106, 381, 506, 394 ], "score": 1.0, "content": "video, and audio. These models can be trained in parallel independently, ensuring exceptional single-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 392, 505, 405 ], "spans": [ { "bbox": [ 106, 392, 505, 405 ], "score": 1.0, "content": "modality generation quality using widely available modality-specific training data (i.e., data with one", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 402, 506, 417 ], "spans": [ { "bbox": [ 105, 402, 506, 417 ], "score": 1.0, "content": "or more modalities as input and one modality as output). For conditional cross-modality generation,", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "such as generating images using audio+language prompts, the input modalities are projected into a", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 425, 507, 437 ], "spans": [ { "bbox": [ 106, 425, 507, 437 ], "score": 1.0, "content": "shared feature space (Section 3.2), and the output LDM attends to the combination of input features.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 435, 505, 449 ], "spans": [ { "bbox": [ 105, 435, 505, 449 ], "score": 1.0, "content": "This multimodal conditioning mechanism prepares the diffusion model to condition on any modality", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 446, 394, 460 ], "spans": [ { "bbox": [ 105, 446, 394, 460 ], "score": 1.0, "content": "or combination of modalities without directly training for such settings.", "type": "text" } ], "index": 32 } ], "index": 27.5 }, { "type": "text", "bbox": [ 107, 463, 505, 572 ], "lines": [ { "bbox": [ 106, 463, 506, 475 ], "spans": [ { "bbox": [ 106, 463, 506, 475 ], "score": 1.0, "content": "The second stage of training enables the model to handle many-to-many generation strategies that", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 474, 506, 487 ], "spans": [ { "bbox": [ 106, 474, 506, 487 ], "score": 1.0, "content": "involve simultaneously generating arbitrary combinations of output modalities. To the best of our", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 485, 506, 497 ], "spans": [ { "bbox": [ 106, 485, 506, 497 ], "score": 1.0, "content": "knowledge, CoDi is the first AI model with this capability. This is achieved by adding a cross-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 496, 506, 508 ], "spans": [ { "bbox": [ 105, 496, 365, 508 ], "score": 1.0, "content": "attention module to each diffuser, and an environment encoder", "type": "text" }, { "bbox": [ 366, 496, 375, 506 ], "score": 0.7, "content": "V", "type": "inline_equation" }, { "bbox": [ 375, 496, 506, 508 ], "score": 1.0, "content": "to project the latent variable of", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 506, 506, 519 ], "spans": [ { "bbox": [ 106, 506, 506, 519 ], "score": 1.0, "content": "different LDMs into a shared latent space (Section 3.4). Next, we freeze the parameters of the LDM,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 518, 506, 529 ], "spans": [ { "bbox": [ 105, 518, 308, 529 ], "score": 1.0, "content": "training only the cross-attention parameters and", "type": "text" }, { "bbox": [ 308, 518, 317, 528 ], "score": 0.64, "content": "V", "type": "inline_equation" }, { "bbox": [ 317, 518, 506, 529 ], "score": 1.0, "content": ". Since the environment encoder of different", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 529, 505, 541 ], "spans": [ { "bbox": [ 106, 529, 505, 541 ], "score": 1.0, "content": "modalities are aligned, an LDM can cross-attend with any group of co-generated modalities by", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 538, 506, 554 ], "spans": [ { "bbox": [ 105, 538, 280, 554 ], "score": 1.0, "content": "interpolating the representation’s output by", "type": "text" }, { "bbox": [ 280, 540, 289, 550 ], "score": 0.6, "content": "V", "type": "inline_equation" }, { "bbox": [ 290, 538, 506, 554 ], "score": 1.0, "content": ". This enables CoDi to seamlessly generate any group", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 550, 505, 563 ], "spans": [ { "bbox": [ 106, 550, 505, 563 ], "score": 1.0, "content": "of modalities, without training on all possible generation combinations. This reduces the number of", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 562, 290, 573 ], "spans": [ { "bbox": [ 106, 562, 290, 573 ], "score": 1.0, "content": "training objectives from exponential to linear.", "type": "text" } ], "index": 42 } ], "index": 37.5 }, { "type": "text", "bbox": [ 107, 578, 505, 655 ], "lines": [ { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "We demonstrate the any-to-any generation capability of CoDi, including single-to-single modality", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 588, 507, 602 ], "spans": [ { "bbox": [ 105, 588, 507, 602 ], "score": 1.0, "content": "generation, multi-condition generation, and the novel capacity of joint generation of multiple modali-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 598, 505, 613 ], "spans": [ { "bbox": [ 105, 598, 505, 613 ], "score": 1.0, "content": "ties. For example, generating synchronized video and audio given the text input prompt; or generating", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 611, 505, 623 ], "spans": [ { "bbox": [ 106, 611, 505, 623 ], "score": 1.0, "content": "video given a prompt image and audio. We also provide a quantitative evaluation of CoDi using eight", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 621, 506, 635 ], "spans": [ { "bbox": [ 105, 621, 506, 635 ], "score": 1.0, "content": "multimodal datasets. As the latest work from Project i-Code [55] towards Composable AI, CoDi", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 633, 506, 645 ], "spans": [ { "bbox": [ 105, 633, 506, 645 ], "score": 1.0, "content": "exhibits exceptional generation quality across assorted scenarios, with synthesis quality on par or", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 644, 480, 657 ], "spans": [ { "bbox": [ 105, 644, 480, 657 ], "score": 1.0, "content": "even better than single to single modality SOTA, e.g., audio generation and audio captioning.", "type": "text" } ], "index": 49 } ], "index": 46 }, { "type": "title", "bbox": [ 108, 673, 201, 686 ], "lines": [ { "bbox": [ 105, 672, 203, 688 ], "spans": [ { "bbox": [ 105, 672, 203, 688 ], "score": 1.0, "content": "2 Related Works", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 107, 699, 504, 722 ], "lines": [ { "bbox": [ 107, 700, 504, 712 ], "spans": [ { "bbox": [ 107, 700, 504, 712 ], "score": 1.0, "content": "Diffusion models (DMs) learn the data distribution by denoising and recovering the original data.", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 709, 505, 725 ], "spans": [ { "bbox": [ 105, 709, 505, 725 ], "score": 1.0, "content": "Deep Diffusion Process (DDP) [45] adopts a sequence of reversible diffusion steps to model image", "type": "text" } ], "index": 52 } ], "index": 51.5 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 309, 750 ], "lines": [ { "bbox": [ 301, 741, 310, 753 ], "spans": [ { "bbox": [ 301, 741, 310, 753 ], "score": 1.0, "content": "2", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 71, 190, 84 ], "lines": [ { "bbox": [ 105, 69, 192, 87 ], "spans": [ { "bbox": [ 105, 69, 192, 87 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 97, 505, 229 ], "lines": [ { "bbox": [ 106, 98, 505, 110 ], "spans": [ { "bbox": [ 106, 98, 505, 110 ], "score": 1.0, "content": "Recent years have seen the rise of powerful cross-modal models that can generate one modality", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 108, 506, 121 ], "spans": [ { "bbox": [ 105, 108, 506, 121 ], "score": 1.0, "content": "from another, e.g. text-to-text [6, 37], text-to-image [13, 19, 22, 41, 44], or text-to-audio [23, 33].", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 119, 505, 132 ], "spans": [ { "bbox": [ 106, 119, 505, 132 ], "score": 1.0, "content": "However, these models are restricted in their real-world applicability where multiple modalities", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 130, 505, 143 ], "spans": [ { "bbox": [ 105, 130, 505, 143 ], "score": 1.0, "content": "coexist and interact. While one can chain together modality-specific generative models in a multi-step", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 141, 506, 154 ], "spans": [ { "bbox": [ 105, 141, 506, 154 ], "score": 1.0, "content": "generation setting, the generation power of each step remains inherently limited, and a serial, multi-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 153, 505, 165 ], "spans": [ { "bbox": [ 106, 153, 505, 165 ], "score": 1.0, "content": "step process can be cumbersome and slow. Moreover, independently generated unimodal streams", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 163, 506, 176 ], "spans": [ { "bbox": [ 105, 163, 506, 176 ], "score": 1.0, "content": "will not be consistent and aligned when stitched together in a post-processing way (e.g., synchronized", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 172, 505, 187 ], "spans": [ { "bbox": [ 105, 172, 505, 187 ], "score": 1.0, "content": "video and audio). The development of a comprehensive and versatile model that can generate any", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 186, 505, 197 ], "spans": [ { "bbox": [ 106, 186, 505, 197 ], "score": 1.0, "content": "combination of modalities from any set of input conditions has been eagerly anticipated, as it would", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 196, 505, 208 ], "spans": [ { "bbox": [ 106, 196, 505, 208 ], "score": 1.0, "content": "more accurately capture the multimodal nature of the world and human comprehension, seamlessly", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 207, 506, 219 ], "spans": [ { "bbox": [ 106, 207, 506, 219 ], "score": 1.0, "content": "consolidate information from a wide range of sources, and enable strong immersion in human-AI", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 218, 506, 230 ], "spans": [ { "bbox": [ 106, 218, 506, 230 ], "score": 1.0, "content": "interactions (for example, by generating coherent video, audio, and text description at the same time).", "type": "text" } ], "index": 12 } ], "index": 6.5, "bbox_fs": [ 105, 98, 506, 230 ] }, { "type": "text", "bbox": [ 107, 234, 505, 343 ], "lines": [ { "bbox": [ 105, 234, 506, 246 ], "spans": [ { "bbox": [ 105, 234, 506, 246 ], "score": 1.0, "content": "In pursuit of this goal, we propose Composable Diffusion, or CoDi, the first model capable of", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 245, 507, 257 ], "spans": [ { "bbox": [ 105, 245, 507, 257 ], "score": 1.0, "content": "simultaneously processing and generating arbitrary combinations of modalities as shown in Fig. 1.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 257, 505, 268 ], "spans": [ { "bbox": [ 106, 257, 505, 268 ], "score": 1.0, "content": "Training a model to take any mixture of input modalities and flexibly generate any mixture of outputs", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 267, 506, 279 ], "spans": [ { "bbox": [ 105, 267, 506, 279 ], "score": 1.0, "content": "presents significant computational and data requirements, as the number of combinations for the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 277, 506, 291 ], "spans": [ { "bbox": [ 105, 277, 506, 291 ], "score": 1.0, "content": "input and output modalities scales exponentially. Also aligned training data for many groups of", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 289, 506, 302 ], "spans": [ { "bbox": [ 105, 289, 506, 302 ], "score": 1.0, "content": "modalities is scarce or even non-existent, making it infeasible to train with all possible input-output", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 299, 506, 312 ], "spans": [ { "bbox": [ 105, 299, 506, 312 ], "score": 1.0, "content": "combinations. To address this challenge, we propose to align multiple modalities in both the input", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 310, 506, 323 ], "spans": [ { "bbox": [ 105, 310, 506, 323 ], "score": 1.0, "content": "conditioning (Section 3.2) and generation diffusion step (Section 3.4). Furthermore, a proposed", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 320, 505, 334 ], "spans": [ { "bbox": [ 105, 320, 505, 334 ], "score": 1.0, "content": "“Bridging Alignment” strategy for contrastive learning (Section 3.2) allows us to efficiently model the", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 333, 481, 345 ], "spans": [ { "bbox": [ 106, 333, 481, 345 ], "score": 1.0, "content": "exponential number of input-output combinations with a linear number of training objectives.", "type": "text" } ], "index": 22 } ], "index": 17.5, "bbox_fs": [ 105, 234, 507, 345 ] }, { "type": "text", "bbox": [ 107, 348, 506, 458 ], "lines": [ { "bbox": [ 106, 349, 506, 361 ], "spans": [ { "bbox": [ 106, 349, 506, 361 ], "score": 1.0, "content": "Building a model with any-to-any generation capacity with exceptional generation quality requires", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 360, 506, 372 ], "spans": [ { "bbox": [ 105, 360, 506, 372 ], "score": 1.0, "content": "comprehensive model design and training on diverse data resources. Therefore, we build CoDi in an", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 369, 507, 385 ], "spans": [ { "bbox": [ 105, 369, 507, 385 ], "score": 1.0, "content": "integrative way. First, we train a latent diffusion model (LDM) for each modality, e.g., text, image,", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 381, 506, 394 ], "spans": [ { "bbox": [ 106, 381, 506, 394 ], "score": 1.0, "content": "video, and audio. These models can be trained in parallel independently, ensuring exceptional single-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 392, 505, 405 ], "spans": [ { "bbox": [ 106, 392, 505, 405 ], "score": 1.0, "content": "modality generation quality using widely available modality-specific training data (i.e., data with one", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 402, 506, 417 ], "spans": [ { "bbox": [ 105, 402, 506, 417 ], "score": 1.0, "content": "or more modalities as input and one modality as output). For conditional cross-modality generation,", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "such as generating images using audio+language prompts, the input modalities are projected into a", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 425, 507, 437 ], "spans": [ { "bbox": [ 106, 425, 507, 437 ], "score": 1.0, "content": "shared feature space (Section 3.2), and the output LDM attends to the combination of input features.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 435, 505, 449 ], "spans": [ { "bbox": [ 105, 435, 505, 449 ], "score": 1.0, "content": "This multimodal conditioning mechanism prepares the diffusion model to condition on any modality", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 446, 394, 460 ], "spans": [ { "bbox": [ 105, 446, 394, 460 ], "score": 1.0, "content": "or combination of modalities without directly training for such settings.", "type": "text" } ], "index": 32 } ], "index": 27.5, "bbox_fs": [ 105, 349, 507, 460 ] }, { "type": "text", "bbox": [ 107, 463, 505, 572 ], "lines": [ { "bbox": [ 106, 463, 506, 475 ], "spans": [ { "bbox": [ 106, 463, 506, 475 ], "score": 1.0, "content": "The second stage of training enables the model to handle many-to-many generation strategies that", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 474, 506, 487 ], "spans": [ { "bbox": [ 106, 474, 506, 487 ], "score": 1.0, "content": "involve simultaneously generating arbitrary combinations of output modalities. To the best of our", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 485, 506, 497 ], "spans": [ { "bbox": [ 106, 485, 506, 497 ], "score": 1.0, "content": "knowledge, CoDi is the first AI model with this capability. This is achieved by adding a cross-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 496, 506, 508 ], "spans": [ { "bbox": [ 105, 496, 365, 508 ], "score": 1.0, "content": "attention module to each diffuser, and an environment encoder", "type": "text" }, { "bbox": [ 366, 496, 375, 506 ], "score": 0.7, "content": "V", "type": "inline_equation" }, { "bbox": [ 375, 496, 506, 508 ], "score": 1.0, "content": "to project the latent variable of", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 506, 506, 519 ], "spans": [ { "bbox": [ 106, 506, 506, 519 ], "score": 1.0, "content": "different LDMs into a shared latent space (Section 3.4). Next, we freeze the parameters of the LDM,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 518, 506, 529 ], "spans": [ { "bbox": [ 105, 518, 308, 529 ], "score": 1.0, "content": "training only the cross-attention parameters and", "type": "text" }, { "bbox": [ 308, 518, 317, 528 ], "score": 0.64, "content": "V", "type": "inline_equation" }, { "bbox": [ 317, 518, 506, 529 ], "score": 1.0, "content": ". Since the environment encoder of different", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 529, 505, 541 ], "spans": [ { "bbox": [ 106, 529, 505, 541 ], "score": 1.0, "content": "modalities are aligned, an LDM can cross-attend with any group of co-generated modalities by", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 538, 506, 554 ], "spans": [ { "bbox": [ 105, 538, 280, 554 ], "score": 1.0, "content": "interpolating the representation’s output by", "type": "text" }, { "bbox": [ 280, 540, 289, 550 ], "score": 0.6, "content": "V", "type": "inline_equation" }, { "bbox": [ 290, 538, 506, 554 ], "score": 1.0, "content": ". This enables CoDi to seamlessly generate any group", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 550, 505, 563 ], "spans": [ { "bbox": [ 106, 550, 505, 563 ], "score": 1.0, "content": "of modalities, without training on all possible generation combinations. This reduces the number of", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 562, 290, 573 ], "spans": [ { "bbox": [ 106, 562, 290, 573 ], "score": 1.0, "content": "training objectives from exponential to linear.", "type": "text" } ], "index": 42 } ], "index": 37.5, "bbox_fs": [ 105, 463, 506, 573 ] }, { "type": "text", "bbox": [ 107, 578, 505, 655 ], "lines": [ { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "We demonstrate the any-to-any generation capability of CoDi, including single-to-single modality", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 588, 507, 602 ], "spans": [ { "bbox": [ 105, 588, 507, 602 ], "score": 1.0, "content": "generation, multi-condition generation, and the novel capacity of joint generation of multiple modali-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 598, 505, 613 ], "spans": [ { "bbox": [ 105, 598, 505, 613 ], "score": 1.0, "content": "ties. For example, generating synchronized video and audio given the text input prompt; or generating", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 611, 505, 623 ], "spans": [ { "bbox": [ 106, 611, 505, 623 ], "score": 1.0, "content": "video given a prompt image and audio. We also provide a quantitative evaluation of CoDi using eight", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 621, 506, 635 ], "spans": [ { "bbox": [ 105, 621, 506, 635 ], "score": 1.0, "content": "multimodal datasets. As the latest work from Project i-Code [55] towards Composable AI, CoDi", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 633, 506, 645 ], "spans": [ { "bbox": [ 105, 633, 506, 645 ], "score": 1.0, "content": "exhibits exceptional generation quality across assorted scenarios, with synthesis quality on par or", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 644, 480, 657 ], "spans": [ { "bbox": [ 105, 644, 480, 657 ], "score": 1.0, "content": "even better than single to single modality SOTA, e.g., audio generation and audio captioning.", "type": "text" } ], "index": 49 } ], "index": 46, "bbox_fs": [ 105, 577, 507, 657 ] }, { "type": "title", "bbox": [ 108, 673, 201, 686 ], "lines": [ { "bbox": [ 105, 672, 203, 688 ], "spans": [ { "bbox": [ 105, 672, 203, 688 ], "score": 1.0, "content": "2 Related Works", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 107, 699, 504, 722 ], "lines": [ { "bbox": [ 107, 700, 504, 712 ], "spans": [ { "bbox": [ 107, 700, 504, 712 ], "score": 1.0, "content": "Diffusion models (DMs) learn the data distribution by denoising and recovering the original data.", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 709, 505, 725 ], "spans": [ { "bbox": [ 105, 709, 505, 725 ], "score": 1.0, "content": "Deep Diffusion Process (DDP) [45] adopts a sequence of reversible diffusion steps to model image", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "probability distribution. It uses a reversible encoder to map the input image to a latent space and", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 104, 403, 505, 416 ], "spans": [ { "bbox": [ 104, 403, 505, 416 ], "score": 1.0, "content": "a decoder to map the latent variables to an output image. Denoising diffusion probabilistic model", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 106, 415, 505, 426 ], "spans": [ { "bbox": [ 106, 415, 505, 426 ], "score": 1.0, "content": "(DDPM) [20] uses a cascade of diffusion processes to gradually increase the complexity of the", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 426, 505, 438 ], "spans": [ { "bbox": [ 105, 426, 505, 438 ], "score": 1.0, "content": "probability density function model. At each step, the model adds noise to the input image and", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 106, 437, 505, 449 ], "spans": [ { "bbox": [ 106, 437, 505, 449 ], "score": 1.0, "content": "estimates the corresponding noise level using an autoregressive model. This allows the model to", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 447, 505, 459 ], "spans": [ { "bbox": [ 105, 447, 505, 459 ], "score": 1.0, "content": "capture the dependencies between adjacent pixels and generate high-quality images. Score-based", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 104, 458, 506, 471 ], "spans": [ { "bbox": [ 104, 458, 506, 471 ], "score": 1.0, "content": "generative models (SOG) [46] use the score function to model the diffusion process. [40] generates", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 105, 469, 505, 481 ], "spans": [ { "bbox": [ 105, 469, 505, 481 ], "score": 1.0, "content": "high-fidelity images conditioned on CLIP representations of text prompts. Latent diffusion model", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 105, 480, 506, 492 ], "spans": [ { "bbox": [ 105, 480, 506, 492 ], "score": 1.0, "content": "(LDM) [41] uses a VAE to encode inputs into latent space to reduce modeling dimension and", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 105, 491, 505, 504 ], "spans": [ { "bbox": [ 105, 491, 505, 504 ], "score": 1.0, "content": "improves efficiency. The motivation is that image compression can be separated into semantic space", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 105, 501, 506, 516 ], "spans": [ { "bbox": [ 105, 501, 506, 516 ], "score": 1.0, "content": "by a diffusion model and perceptual space by an autoencoder. By incorporating temporal modeling", "type": "text", "cross_page": true } ], "index": 16 }, { "bbox": [ 105, 512, 505, 526 ], "spans": [ { "bbox": [ 105, 512, 505, 526 ], "score": 1.0, "content": "modules and cascading model architectures, video diffusion models have been built upon image", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 106, 523, 505, 536 ], "spans": [ { "bbox": [ 106, 523, 505, 536 ], "score": 1.0, "content": "diffusers to generate temporally consistent and inherent frames[14, 19, 21, 44]. Diffusion models have", "type": "text", "cross_page": true } ], "index": 18 }, { "bbox": [ 106, 535, 502, 547 ], "spans": [ { "bbox": [ 106, 535, 502, 547 ], "score": 1.0, "content": "also been applied to other domains, such as generating audio from text and vision prompts[23, 33].", "type": "text", "cross_page": true } ], "index": 19 } ], "index": 51.5, "bbox_fs": [ 105, 700, 505, 725 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 69, 503, 329 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 69, 503, 329 ], "group_id": 0, "lines": [ { "bbox": [ 106, 69, 503, 329 ], "spans": [ { "bbox": [ 106, 69, 503, 329 ], "score": 0.974, "type": "image", "image_path": "fd34f6f54041b6f9329e8b6d950176f23a2bd0ce1819fa8d9711a4b7d903ec66.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 69, 503, 155.66666666666669 ], "spans": [], "index": 0 }, { "bbox": [ 106, 155.66666666666669, 503, 242.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 106, 242.33333333333337, 503, 329.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 348, 505, 381 ], "group_id": 0, "lines": [ { "bbox": [ 105, 347, 506, 361 ], "spans": [ { "bbox": [ 105, 347, 506, 361 ], "score": 1.0, "content": "Figure 2: CoDi model architecture: (a) We first train individual diffusion model with aligned prompt", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 360, 505, 371 ], "spans": [ { "bbox": [ 106, 360, 505, 371 ], "score": 1.0, "content": "encoder by “Bridging Alignment”; (b) Diffusion models learn to attend with each other via “Latent", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 370, 497, 383 ], "spans": [ { "bbox": [ 106, 370, 497, 383 ], "score": 1.0, "content": "Alignment”; (c) CoDi achieves any-to-any generation with a linear number of training objectives.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 393, 505, 546 ], "lines": [ { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "probability distribution. It uses a reversible encoder to map the input image to a latent space and", "type": "text" } ], "index": 6 }, { "bbox": [ 104, 403, 505, 416 ], "spans": [ { "bbox": [ 104, 403, 505, 416 ], "score": 1.0, "content": "a decoder to map the latent variables to an output image. Denoising diffusion probabilistic model", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 415, 505, 426 ], "spans": [ { "bbox": [ 106, 415, 505, 426 ], "score": 1.0, "content": "(DDPM) [20] uses a cascade of diffusion processes to gradually increase the complexity of the", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 426, 505, 438 ], "spans": [ { "bbox": [ 105, 426, 505, 438 ], "score": 1.0, "content": "probability density function model. At each step, the model adds noise to the input image and", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 437, 505, 449 ], "spans": [ { "bbox": [ 106, 437, 505, 449 ], "score": 1.0, "content": "estimates the corresponding noise level using an autoregressive model. This allows the model to", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 447, 505, 459 ], "spans": [ { "bbox": [ 105, 447, 505, 459 ], "score": 1.0, "content": "capture the dependencies between adjacent pixels and generate high-quality images. Score-based", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 458, 506, 471 ], "spans": [ { "bbox": [ 104, 458, 506, 471 ], "score": 1.0, "content": "generative models (SOG) [46] use the score function to model the diffusion process. [40] generates", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 469, 505, 481 ], "spans": [ { "bbox": [ 105, 469, 505, 481 ], "score": 1.0, "content": "high-fidelity images conditioned on CLIP representations of text prompts. Latent diffusion model", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 480, 506, 492 ], "spans": [ { "bbox": [ 105, 480, 506, 492 ], "score": 1.0, "content": "(LDM) [41] uses a VAE to encode inputs into latent space to reduce modeling dimension and", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 491, 505, 504 ], "spans": [ { "bbox": [ 105, 491, 505, 504 ], "score": 1.0, "content": "improves efficiency. The motivation is that image compression can be separated into semantic space", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 501, 506, 516 ], "spans": [ { "bbox": [ 105, 501, 506, 516 ], "score": 1.0, "content": "by a diffusion model and perceptual space by an autoencoder. By incorporating temporal modeling", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 512, 505, 526 ], "spans": [ { "bbox": [ 105, 512, 505, 526 ], "score": 1.0, "content": "modules and cascading model architectures, video diffusion models have been built upon image", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 523, 505, 536 ], "spans": [ { "bbox": [ 106, 523, 505, 536 ], "score": 1.0, "content": "diffusers to generate temporally consistent and inherent frames[14, 19, 21, 44]. Diffusion models have", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 535, 502, 547 ], "spans": [ { "bbox": [ 106, 535, 502, 547 ], "score": 1.0, "content": "also been applied to other domains, such as generating audio from text and vision prompts[23, 33].", "type": "text" } ], "index": 19 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 551, 505, 639 ], "lines": [ { "bbox": [ 106, 551, 505, 563 ], "spans": [ { "bbox": [ 106, 551, 505, 563 ], "score": 1.0, "content": "Multimodal modeling has experienced rapid advancement recently, with researchers striving to build", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 563, 505, 573 ], "spans": [ { "bbox": [ 106, 563, 505, 573 ], "score": 1.0, "content": "uniform representations of multiple modalities using a single model to achieve more comprehensive", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 573, 505, 585 ], "spans": [ { "bbox": [ 106, 573, 505, 585 ], "score": 1.0, "content": "cross-modal understanding. Vision transformers [11], featuring diverse model architectures and", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 584, 505, 596 ], "spans": [ { "bbox": [ 106, 584, 505, 596 ], "score": 1.0, "content": "training techniques, have been applied to various downstream tasks such as vision Q&A and image", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 595, 506, 606 ], "spans": [ { "bbox": [ 105, 595, 506, 606 ], "score": 1.0, "content": "captioning. Multimodal encoders have also proven successful in vision-language [1, 8, 57], video-", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 605, 505, 618 ], "spans": [ { "bbox": [ 106, 605, 505, 618 ], "score": 1.0, "content": "audio [47] and video-speech-language [55, 56] domains. Aligning data from different modalities is", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 615, 505, 630 ], "spans": [ { "bbox": [ 105, 615, 505, 630 ], "score": 1.0, "content": "an active research area [12, 38], with promising applications in cross-modality retrieval and building", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 627, 303, 639 ], "spans": [ { "bbox": [ 106, 627, 303, 639 ], "score": 1.0, "content": "uniform multimodal representations [33, 35, 41].", "type": "text" } ], "index": 27 } ], "index": 23.5 }, { "type": "title", "bbox": [ 107, 654, 192, 668 ], "lines": [ { "bbox": [ 104, 651, 194, 672 ], "spans": [ { "bbox": [ 104, 651, 194, 672 ], "score": 1.0, "content": "3 Methodology", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "title", "bbox": [ 107, 679, 288, 691 ], "lines": [ { "bbox": [ 105, 678, 289, 693 ], "spans": [ { "bbox": [ 105, 678, 289, 693 ], "score": 1.0, "content": "3.1 Preliminary: Latent Diffusion Model", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 700, 503, 722 ], "lines": [ { "bbox": [ 105, 698, 505, 713 ], "spans": [ { "bbox": [ 105, 698, 471, 713 ], "score": 1.0, "content": "Diffusion models (DM) represent a class of generative models that learn data distributions", "type": "text" }, { "bbox": [ 471, 700, 491, 712 ], "score": 0.92, "content": "p ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 492, 698, 505, 713 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 505, 723 ], "score": 1.0, "content": "simulating the diffusion of information over time. During training, random noise is iteratively added", "type": "text" } ], "index": 31 } ], "index": 30.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 301, 740, 309, 752 ], "spans": [ { "bbox": [ 301, 740, 309, 752 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 69, 503, 329 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 69, 503, 329 ], "group_id": 0, "lines": [ { "bbox": [ 106, 69, 503, 329 ], "spans": [ { "bbox": [ 106, 69, 503, 329 ], "score": 0.974, "type": "image", "image_path": "fd34f6f54041b6f9329e8b6d950176f23a2bd0ce1819fa8d9711a4b7d903ec66.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 69, 503, 155.66666666666669 ], "spans": [], "index": 0 }, { "bbox": [ 106, 155.66666666666669, 503, 242.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 106, 242.33333333333337, 503, 329.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 348, 505, 381 ], "group_id": 0, "lines": [ { "bbox": [ 105, 347, 506, 361 ], "spans": [ { "bbox": [ 105, 347, 506, 361 ], "score": 1.0, "content": "Figure 2: CoDi model architecture: (a) We first train individual diffusion model with aligned prompt", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 360, 505, 371 ], "spans": [ { "bbox": [ 106, 360, 505, 371 ], "score": 1.0, "content": "encoder by “Bridging Alignment”; (b) Diffusion models learn to attend with each other via “Latent", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 370, 497, 383 ], "spans": [ { "bbox": [ 106, 370, 497, 383 ], "score": 1.0, "content": "Alignment”; (c) CoDi achieves any-to-any generation with a linear number of training objectives.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 393, 505, 546 ], "lines": [], "index": 12.5, "bbox_fs": [ 104, 392, 506, 547 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 551, 505, 639 ], "lines": [ { "bbox": [ 106, 551, 505, 563 ], "spans": [ { "bbox": [ 106, 551, 505, 563 ], "score": 1.0, "content": "Multimodal modeling has experienced rapid advancement recently, with researchers striving to build", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 563, 505, 573 ], "spans": [ { "bbox": [ 106, 563, 505, 573 ], "score": 1.0, "content": "uniform representations of multiple modalities using a single model to achieve more comprehensive", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 573, 505, 585 ], "spans": [ { "bbox": [ 106, 573, 505, 585 ], "score": 1.0, "content": "cross-modal understanding. Vision transformers [11], featuring diverse model architectures and", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 584, 505, 596 ], "spans": [ { "bbox": [ 106, 584, 505, 596 ], "score": 1.0, "content": "training techniques, have been applied to various downstream tasks such as vision Q&A and image", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 595, 506, 606 ], "spans": [ { "bbox": [ 105, 595, 506, 606 ], "score": 1.0, "content": "captioning. Multimodal encoders have also proven successful in vision-language [1, 8, 57], video-", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 605, 505, 618 ], "spans": [ { "bbox": [ 106, 605, 505, 618 ], "score": 1.0, "content": "audio [47] and video-speech-language [55, 56] domains. Aligning data from different modalities is", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 615, 505, 630 ], "spans": [ { "bbox": [ 105, 615, 505, 630 ], "score": 1.0, "content": "an active research area [12, 38], with promising applications in cross-modality retrieval and building", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 627, 303, 639 ], "spans": [ { "bbox": [ 106, 627, 303, 639 ], "score": 1.0, "content": "uniform multimodal representations [33, 35, 41].", "type": "text" } ], "index": 27 } ], "index": 23.5, "bbox_fs": [ 105, 551, 506, 639 ] }, { "type": "title", "bbox": [ 107, 654, 192, 668 ], "lines": [ { "bbox": [ 104, 651, 194, 672 ], "spans": [ { "bbox": [ 104, 651, 194, 672 ], "score": 1.0, "content": "3 Methodology", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "title", "bbox": [ 107, 679, 288, 691 ], "lines": [ { "bbox": [ 105, 678, 289, 693 ], "spans": [ { "bbox": [ 105, 678, 289, 693 ], "score": 1.0, "content": "3.1 Preliminary: Latent Diffusion Model", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 700, 503, 722 ], "lines": [ { "bbox": [ 105, 698, 505, 713 ], "spans": [ { "bbox": [ 105, 698, 471, 713 ], "score": 1.0, "content": "Diffusion models (DM) represent a class of generative models that learn data distributions", "type": "text" }, { "bbox": [ 471, 700, 491, 712 ], "score": 0.92, "content": "p ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 492, 698, 505, 713 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 505, 723 ], "score": 1.0, "content": "simulating the diffusion of information over time. During training, random noise is iteratively added", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 72, 505, 85 ], "spans": [ { "bbox": [ 105, 72, 117, 85 ], "score": 1.0, "content": "to", "type": "text", "cross_page": true }, { "bbox": [ 117, 75, 124, 83 ], "score": 0.69, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 125, 72, 505, 85 ], "score": 1.0, "content": ", while the model learns to denoise the examples. For inference, the model denoises data points", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "sampled from simple distributions such as Gaussian. Latent diffusion models (LDM) [41] learn the", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 94, 505, 108 ], "spans": [ { "bbox": [ 105, 94, 238, 108 ], "score": 1.0, "content": "distribution of the latent variable", "type": "text", "cross_page": true }, { "bbox": [ 238, 96, 245, 104 ], "score": 0.76, "content": "_ { z }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 246, 94, 316, 108 ], "score": 1.0, "content": "corresponding to", "type": "text", "cross_page": true }, { "bbox": [ 316, 96, 323, 104 ], "score": 0.75, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 324, 94, 505, 108 ], "score": 1.0, "content": ", significantly reducing computational cost by", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 105, 231, 118 ], "spans": [ { "bbox": [ 105, 105, 231, 118 ], "score": 1.0, "content": "decreasing the data dimension.", "type": "text", "cross_page": true } ], "index": 3 } ], "index": 30.5, "bbox_fs": [ 105, 698, 505, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 117 ], "lines": [ { "bbox": [ 105, 72, 505, 85 ], "spans": [ { "bbox": [ 105, 72, 117, 85 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 117, 75, 124, 83 ], "score": 0.69, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 125, 72, 505, 85 ], "score": 1.0, "content": ", while the model learns to denoise the examples. For inference, the model denoises data points", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "sampled from simple distributions such as Gaussian. Latent diffusion models (LDM) [41] learn the", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 94, 505, 108 ], "spans": [ { "bbox": [ 105, 94, 238, 108 ], "score": 1.0, "content": "distribution of the latent variable", "type": "text" }, { "bbox": [ 238, 96, 245, 104 ], "score": 0.76, "content": "_ { z }", "type": "inline_equation" }, { "bbox": [ 246, 94, 316, 108 ], "score": 1.0, "content": "corresponding to", "type": "text" }, { "bbox": [ 316, 96, 323, 104 ], "score": 0.75, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 324, 94, 505, 108 ], "score": 1.0, "content": ", significantly reducing computational cost by", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 105, 231, 118 ], "spans": [ { "bbox": [ 105, 105, 231, 118 ], "score": 1.0, "content": "decreasing the data dimension.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 106, 121, 505, 166 ], "lines": [ { "bbox": [ 105, 121, 506, 134 ], "spans": [ { "bbox": [ 105, 121, 319, 134 ], "score": 1.0, "content": "In LDM, an autoencoder is first trained to reconstruct", "type": "text" }, { "bbox": [ 320, 124, 327, 132 ], "score": 0.77, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 328, 121, 348, 134 ], "score": 1.0, "content": ", i.e.,", "type": "text" }, { "bbox": [ 348, 122, 407, 133 ], "score": 0.91, "content": "\\hat { \\pmb { x } } = D ( E ( \\pmb { x } ) )", "type": "inline_equation" }, { "bbox": [ 408, 121, 438, 134 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 438, 122, 447, 132 ], "score": 0.83, "content": "E", "type": "inline_equation" }, { "bbox": [ 448, 121, 465, 134 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 465, 122, 475, 132 ], "score": 0.8, "content": "D", "type": "inline_equation" }, { "bbox": [ 475, 121, 506, 134 ], "score": 1.0, "content": "denote", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 132, 506, 144 ], "spans": [ { "bbox": [ 106, 132, 348, 144 ], "score": 1.0, "content": "the encoder and decoder, respectively. The latent variable", "type": "text" }, { "bbox": [ 348, 133, 393, 144 ], "score": 0.9, "content": "z = E ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 393, 132, 506, 144 ], "score": 1.0, "content": "is iteratively diffused over", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 143, 505, 157 ], "spans": [ { "bbox": [ 105, 143, 150, 157 ], "score": 1.0, "content": "time steps", "type": "text" }, { "bbox": [ 150, 145, 155, 154 ], "score": 0.75, "content": "t", "type": "inline_equation" }, { "bbox": [ 155, 143, 276, 157 ], "score": 1.0, "content": "based on a variance schedule", "type": "text" }, { "bbox": [ 276, 144, 321, 155 ], "score": 0.93, "content": "\\beta _ { 1 } , \\ldots , \\beta _ { T }", "type": "inline_equation" }, { "bbox": [ 322, 143, 343, 157 ], "score": 1.0, "content": ", i.e.,", "type": "text" }, { "bbox": [ 343, 143, 505, 156 ], "score": 0.93, "content": "q ( z _ { t } | z _ { t - 1 } ) = \\mathcal { N } ( z _ { t } ; \\sqrt { 1 - \\beta _ { t } } z _ { t - 1 } , \\beta _ { t } I )", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 153, 144, 167 ], "spans": [ { "bbox": [ 105, 153, 144, 167 ], "score": 1.0, "content": "[20, 45].", "type": "text" } ], "index": 7 } ], "index": 5.5 }, { "type": "text", "bbox": [ 106, 171, 505, 217 ], "lines": [ { "bbox": [ 105, 171, 506, 184 ], "spans": [ { "bbox": [ 105, 171, 322, 184 ], "score": 1.0, "content": "The forward process allows the random sampling of", "type": "text" }, { "bbox": [ 323, 173, 333, 182 ], "score": 0.86, "content": "{ \\boldsymbol { z } } _ { t }", "type": "inline_equation" }, { "bbox": [ 333, 171, 506, 184 ], "score": 1.0, "content": "at any timestep in a closed form [20, 45]:", "type": "text" } ], "index": 8 }, { "bbox": [ 107, 182, 506, 198 ], "spans": [ { "bbox": [ 107, 184, 172, 195 ], "score": 0.89, "content": "\\boldsymbol { z } _ { t } = \\alpha _ { t } \\boldsymbol { z } + \\sigma _ { t } \\boldsymbol { \\epsilon }", "type": "inline_equation" }, { "bbox": [ 172, 182, 201, 198 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 202, 183, 252, 195 ], "score": 0.82, "content": "\\epsilon \\sim \\mathcal { N } ( 0 , I )", "type": "inline_equation" }, { "bbox": [ 253, 182, 256, 198 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 257, 183, 309, 195 ], "score": 0.86, "content": "\\alpha _ { t } : = 1 - \\beta _ { t }", "type": "inline_equation" }, { "bbox": [ 309, 182, 326, 198 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 327, 182, 406, 196 ], "score": 0.92, "content": "\\begin{array} { r } { \\sigma _ { t } : = \\dot { 1 } - \\prod _ { s = 1 } ^ { t } \\dot { \\alpha } _ { s } } \\end{array}", "type": "inline_equation" }, { "bbox": [ 407, 182, 506, 198 ], "score": 1.0, "content": ". The diffuser learns how", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 194, 505, 207 ], "spans": [ { "bbox": [ 105, 194, 172, 207 ], "score": 1.0, "content": "to denoise from", "type": "text" }, { "bbox": [ 173, 194, 192, 206 ], "score": 0.92, "content": "\\left\\{ { z } _ { t } \\right\\}", "type": "inline_equation" }, { "bbox": [ 193, 194, 237, 207 ], "score": 1.0, "content": "to recover", "type": "text" }, { "bbox": [ 237, 196, 244, 204 ], "score": 0.71, "content": "_ z", "type": "inline_equation" }, { "bbox": [ 244, 194, 505, 207 ], "score": 1.0, "content": ". Following the reparameterization method proposed in [20], the", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 205, 322, 217 ], "spans": [ { "bbox": [ 106, 205, 322, 217 ], "score": 1.0, "content": "denoising training objective can be expressed as [41]:", "type": "text" } ], "index": 11 } ], "index": 9.5 }, { "type": "interline_equation", "bbox": [ 231, 221, 380, 236 ], "lines": [ { "bbox": [ 231, 221, 380, 236 ], "spans": [ { "bbox": [ 231, 221, 380, 236 ], "score": 0.91, "content": "\\begin{array} { r } { \\mathcal { L } _ { D } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta } ( z _ { t } , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } . } \\end{array}", "type": "interline_equation", "image_path": "c0b6a42a7dddc81d5d528ed8a349c84f1a8e8e52d97b6eb0c26878eaf69606c7.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 231, 221, 380, 236 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 107, 241, 506, 254 ], "lines": [ { "bbox": [ 105, 239, 507, 256 ], "spans": [ { "bbox": [ 105, 239, 507, 256 ], "score": 1.0, "content": "In data generation, the denoising process can be realized through reparameterized Gaussian sampling:", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "interline_equation", "bbox": [ 194, 261, 416, 289 ], "lines": [ { "bbox": [ 194, 261, 416, 289 ], "spans": [ { "bbox": [ 194, 261, 416, 289 ], "score": 0.94, "content": "p ( z _ { t - 1 } | z _ { t } ) = \\mathcal { N } \\left( z _ { t - 1 } ; \\frac { 1 } { \\sqrt { \\alpha _ { t } } } \\left( z _ { t } - \\frac { \\beta _ { t } } { \\sqrt { \\sigma _ { t } } } \\epsilon _ { \\theta } \\right) , \\beta _ { t } I \\right) .", "type": "interline_equation", "image_path": "cd64023bfb25f4aca1f1f00e845397403302ab368fd5956a291fe57286f62be3.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 194, 261, 416, 275.0 ], "spans": [], "index": 14 }, { "bbox": [ 194, 275.0, 416, 289.0 ], "spans": [], "index": 15 } ] }, { "type": "text", "bbox": [ 106, 292, 506, 358 ], "lines": [ { "bbox": [ 106, 292, 506, 304 ], "spans": [ { "bbox": [ 106, 293, 117, 304 ], "score": 1.0, "content": "In", "type": "text" }, { "bbox": [ 118, 293, 133, 304 ], "score": 0.89, "content": "\\mathcal { L } _ { D }", "type": "inline_equation" }, { "bbox": [ 133, 293, 231, 304 ], "score": 1.0, "content": ", the diffusion time step", "type": "text" }, { "bbox": [ 232, 292, 280, 304 ], "score": 0.86, "content": "t \\sim \\mathcal { U } [ 1 , T ]", "type": "inline_equation" }, { "bbox": [ 280, 293, 284, 304 ], "score": 1.0, "content": ";", "type": "text" }, { "bbox": [ 284, 294, 295, 304 ], "score": 0.65, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 295, 293, 506, 304 ], "score": 1.0, "content": "is a denoising model with UNet backbone parame-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 149, 316 ], "score": 1.0, "content": "terized by", "type": "text" }, { "bbox": [ 150, 304, 168, 315 ], "score": 0.36, "content": "\\theta ; { \\boldsymbol { y } }", "type": "inline_equation" }, { "bbox": [ 169, 303, 470, 316 ], "score": 1.0, "content": "represents the conditional variable that can be used to control generation;", "type": "text" }, { "bbox": [ 470, 304, 479, 313 ], "score": 0.83, "content": "C", "type": "inline_equation" }, { "bbox": [ 479, 303, 505, 316 ], "score": 1.0, "content": "is the", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 315, 505, 327 ], "spans": [ { "bbox": [ 105, 315, 432, 327 ], "score": 1.0, "content": "prompt encoder. The conditioning mechanism is implemented by first featurizing", "type": "text" }, { "bbox": [ 432, 316, 440, 326 ], "score": 0.78, "content": "\\textbf { { y } }", "type": "inline_equation" }, { "bbox": [ 440, 315, 459, 327 ], "score": 1.0, "content": "into", "type": "text" }, { "bbox": [ 459, 315, 481, 326 ], "score": 0.91, "content": "C ( \\boldsymbol { y } )", "type": "inline_equation" }, { "bbox": [ 482, 315, 505, 327 ], "score": 1.0, "content": ", then", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 325, 505, 338 ], "spans": [ { "bbox": [ 106, 325, 146, 338 ], "score": 1.0, "content": "the UNet", "type": "text" }, { "bbox": [ 146, 327, 157, 336 ], "score": 0.85, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 158, 325, 217, 338 ], "score": 1.0, "content": "conditions on", "type": "text" }, { "bbox": [ 218, 326, 240, 338 ], "score": 0.91, "content": "C ( \\boldsymbol { y } )", "type": "inline_equation" }, { "bbox": [ 240, 325, 505, 338 ], "score": 1.0, "content": "via cross-attention, as described in [41]. Distinct from previous", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 337, 506, 348 ], "spans": [ { "bbox": [ 106, 337, 506, 348 ], "score": 1.0, "content": "works, our model can condition on any combinations of modalities of text, image, video and audio.", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 347, 291, 359 ], "spans": [ { "bbox": [ 106, 347, 291, 359 ], "score": 1.0, "content": "Details are presented in the following section.", "type": "text" } ], "index": 21 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 372, 294, 384 ], "lines": [ { "bbox": [ 105, 370, 296, 387 ], "spans": [ { "bbox": [ 105, 370, 296, 387 ], "score": 1.0, "content": "3.2 Composable Multimodal Conditioning", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 392, 505, 470 ], "lines": [ { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "To enable our model to condition on any combination of input/prompt modalities, we align the prompt", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 404, 505, 416 ], "spans": [ { "bbox": [ 106, 404, 315, 416 ], "score": 1.0, "content": "encoder of text, image, video and audio (denoted by", "type": "text" }, { "bbox": [ 316, 404, 327, 415 ], "score": 0.73, "content": "C _ { t }", "type": "inline_equation" }, { "bbox": [ 327, 404, 331, 416 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 331, 404, 342, 415 ], "score": 0.74, "content": "C _ { i }", "type": "inline_equation" }, { "bbox": [ 343, 404, 346, 416 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 346, 404, 359, 415 ], "score": 0.78, "content": "C _ { v }", "type": "inline_equation" }, { "bbox": [ 360, 404, 380, 416 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 380, 404, 393, 415 ], "score": 0.89, "content": "C _ { a }", "type": "inline_equation" }, { "bbox": [ 393, 404, 505, 416 ], "score": 1.0, "content": ", respectively) to project the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 415, 505, 427 ], "spans": [ { "bbox": [ 106, 415, 505, 427 ], "score": 1.0, "content": "input from any modality into the same space. Multimodal conditioning can then be conveniently", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 424, 505, 441 ], "spans": [ { "bbox": [ 104, 424, 349, 441 ], "score": 1.0, "content": "achieved by interpolating the representations of each modality", "type": "text" }, { "bbox": [ 350, 426, 360, 435 ], "score": 0.55, "content": "m", "type": "inline_equation" }, { "bbox": [ 360, 424, 365, 441 ], "score": 1.0, "content": ":", "type": "text" }, { "bbox": [ 365, 425, 505, 438 ], "score": 0.91, "content": "\\begin{array} { r } { C ( x _ { t } , \\bar { x _ { i } } , x _ { v } , x _ { a } ) = \\sum _ { m } \\alpha _ { m } C ( \\bar { m } ) } \\end{array}", "type": "inline_equation" } ], "index": 26 }, { "bbox": [ 104, 435, 506, 452 ], "spans": [ { "bbox": [ 104, 435, 122, 452 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 122, 437, 201, 448 ], "score": 0.9, "content": "m \\in \\ b { x } _ { t } , \\ b { x } _ { i } , \\ b { x } _ { v } , \\ b { x } _ { a }", "type": "inline_equation" }, { "bbox": [ 201, 435, 228, 452 ], "score": 1.0, "content": ", with", "type": "text" }, { "bbox": [ 228, 436, 286, 449 ], "score": 0.91, "content": "\\textstyle \\sum _ { m } \\alpha _ { m } = 1", "type": "inline_equation" }, { "bbox": [ 286, 435, 506, 452 ], "score": 1.0, "content": ". Through simple weighted interpolation of aligned", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 447, 505, 460 ], "spans": [ { "bbox": [ 106, 447, 505, 460 ], "score": 1.0, "content": "embeddings, we enable models trained with single-conditioning (i.e., with only one input) to perform", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 458, 501, 471 ], "spans": [ { "bbox": [ 106, 458, 501, 471 ], "score": 1.0, "content": "zero-shot multi-conditioning (i.e., with multiple inputs). This process is illustrated in Fig. 2 (a)(2).", "type": "text" } ], "index": 29 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 474, 505, 573 ], "lines": [ { "bbox": [ 106, 474, 505, 487 ], "spans": [ { "bbox": [ 106, 474, 505, 487 ], "score": 1.0, "content": "Optimizing all four prompt encoders simultaneously in a combinatorial manner is computationally", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 484, 506, 500 ], "spans": [ { "bbox": [ 105, 484, 155, 500 ], "score": 1.0, "content": "heavy, with", "type": "text" }, { "bbox": [ 155, 485, 182, 497 ], "score": 0.93, "content": "\\mathcal { O } ( n ^ { 2 } )", "type": "inline_equation" }, { "bbox": [ 183, 484, 506, 500 ], "score": 1.0, "content": "pairs. Additionally, for certain dual modalities, well-aligned paired datasets are", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 495, 505, 509 ], "spans": [ { "bbox": [ 105, 495, 505, 509 ], "score": 1.0, "content": "limited or unavailable e.g., image-audio pairs. To address this challenge, we propose a simple and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 506, 505, 520 ], "spans": [ { "bbox": [ 105, 506, 505, 520 ], "score": 1.0, "content": "effective technique called \"Bridging Alignment\" to efficiently align conditional encoders. As shown", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 518, 506, 532 ], "spans": [ { "bbox": [ 105, 518, 506, 532 ], "score": 1.0, "content": "in Fig. 2 (a)(1), we choose the text modality as the \"bridging\" modality due to its ubiquitous presence", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 529, 505, 542 ], "spans": [ { "bbox": [ 105, 529, 505, 542 ], "score": 1.0, "content": "in paired data, such as text-image, text-video, and text-audio pairs. We begin with a pretrained", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 540, 505, 552 ], "spans": [ { "bbox": [ 105, 540, 505, 552 ], "score": 1.0, "content": "text-image paired encoder, i.e., CLIP [38]. We then train audio and video prompt encoders on", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 551, 506, 564 ], "spans": [ { "bbox": [ 106, 551, 506, 564 ], "score": 1.0, "content": "audio-text and video-text paired datasets using contrastive learning, with text and image encoder", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 561, 169, 575 ], "spans": [ { "bbox": [ 105, 561, 169, 575 ], "score": 1.0, "content": "weights frozen.", "type": "text" } ], "index": 38 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 578, 505, 644 ], "lines": [ { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "In this way, all four modalities are aligned in the feature space. As shown in Section 5.2, CoDi", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 589, 505, 602 ], "spans": [ { "bbox": [ 105, 589, 505, 602 ], "score": 1.0, "content": "can effectively leverage and combine the complementary information present in any combination", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "of modalities to generate more accurate and comprehensive outputs. The high generation quality", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 611, 505, 622 ], "spans": [ { "bbox": [ 106, 611, 505, 622 ], "score": 1.0, "content": "remains unaffected with respect to the number of prompt modalities. As we will discuss in subsequent", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "sections, we continue to apply Bridging Alignment to align the latent space of LDMs with different", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 632, 309, 645 ], "spans": [ { "bbox": [ 105, 632, 309, 645 ], "score": 1.0, "content": "modalities to achieve joint multimodal generation.", "type": "text" } ], "index": 44 } ], "index": 41.5 }, { "type": "title", "bbox": [ 107, 657, 225, 669 ], "lines": [ { "bbox": [ 104, 656, 226, 672 ], "spans": [ { "bbox": [ 104, 656, 226, 672 ], "score": 1.0, "content": "3.3 Composable Diffusion", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 678, 505, 722 ], "lines": [ { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 505, 691 ], "score": 1.0, "content": "Training an end-to-end anything-to-anything model requires extensive learning on various data", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 689, 505, 701 ], "spans": [ { "bbox": [ 105, 689, 505, 701 ], "score": 1.0, "content": "resources. The model also needs to maintain generation quality for all synthesis flows. To address", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 506, 712 ], "spans": [ { "bbox": [ 105, 700, 506, 712 ], "score": 1.0, "content": "these challenges, CoDi is designed to be composable and integrative, allowing individual modality-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 711, 505, 724 ], "spans": [ { "bbox": [ 105, 711, 505, 724 ], "score": 1.0, "content": "specific models to be built independently and then smoothly integrated later. Specifically, we start by", "type": "text" } ], "index": 49 } ], "index": 47.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 308, 750 ], "lines": [ { "bbox": [ 301, 741, 310, 752 ], "spans": [ { "bbox": [ 301, 741, 310, 752 ], "score": 1.0, "content": "", "type": "text", "height": 11, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 117 ], "lines": [], "index": 1.5, "bbox_fs": [ 105, 72, 505, 118 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 121, 505, 166 ], "lines": [ { "bbox": [ 105, 121, 506, 134 ], "spans": [ { "bbox": [ 105, 121, 319, 134 ], "score": 1.0, "content": "In LDM, an autoencoder is first trained to reconstruct", "type": "text" }, { "bbox": [ 320, 124, 327, 132 ], "score": 0.77, "content": "_ { \\textbf { \\em x } }", "type": "inline_equation" }, { "bbox": [ 328, 121, 348, 134 ], "score": 1.0, "content": ", i.e.,", "type": "text" }, { "bbox": [ 348, 122, 407, 133 ], "score": 0.91, "content": "\\hat { \\pmb { x } } = D ( E ( \\pmb { x } ) )", "type": "inline_equation" }, { "bbox": [ 408, 121, 438, 134 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 438, 122, 447, 132 ], "score": 0.83, "content": "E", "type": "inline_equation" }, { "bbox": [ 448, 121, 465, 134 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 465, 122, 475, 132 ], "score": 0.8, "content": "D", "type": "inline_equation" }, { "bbox": [ 475, 121, 506, 134 ], "score": 1.0, "content": "denote", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 132, 506, 144 ], "spans": [ { "bbox": [ 106, 132, 348, 144 ], "score": 1.0, "content": "the encoder and decoder, respectively. The latent variable", "type": "text" }, { "bbox": [ 348, 133, 393, 144 ], "score": 0.9, "content": "z = E ( { \\pmb x } )", "type": "inline_equation" }, { "bbox": [ 393, 132, 506, 144 ], "score": 1.0, "content": "is iteratively diffused over", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 143, 505, 157 ], "spans": [ { "bbox": [ 105, 143, 150, 157 ], "score": 1.0, "content": "time steps", "type": "text" }, { "bbox": [ 150, 145, 155, 154 ], "score": 0.75, "content": "t", "type": "inline_equation" }, { "bbox": [ 155, 143, 276, 157 ], "score": 1.0, "content": "based on a variance schedule", "type": "text" }, { "bbox": [ 276, 144, 321, 155 ], "score": 0.93, "content": "\\beta _ { 1 } , \\ldots , \\beta _ { T }", "type": "inline_equation" }, { "bbox": [ 322, 143, 343, 157 ], "score": 1.0, "content": ", i.e.,", "type": "text" }, { "bbox": [ 343, 143, 505, 156 ], "score": 0.93, "content": "q ( z _ { t } | z _ { t - 1 } ) = \\mathcal { N } ( z _ { t } ; \\sqrt { 1 - \\beta _ { t } } z _ { t - 1 } , \\beta _ { t } I )", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 153, 144, 167 ], "spans": [ { "bbox": [ 105, 153, 144, 167 ], "score": 1.0, "content": "[20, 45].", "type": "text" } ], "index": 7 } ], "index": 5.5, "bbox_fs": [ 105, 121, 506, 167 ] }, { "type": "text", "bbox": [ 106, 171, 505, 217 ], "lines": [ { "bbox": [ 105, 171, 506, 184 ], "spans": [ { "bbox": [ 105, 171, 322, 184 ], "score": 1.0, "content": "The forward process allows the random sampling of", "type": "text" }, { "bbox": [ 323, 173, 333, 182 ], "score": 0.86, "content": "{ \\boldsymbol { z } } _ { t }", "type": "inline_equation" }, { "bbox": [ 333, 171, 506, 184 ], "score": 1.0, "content": "at any timestep in a closed form [20, 45]:", "type": "text" } ], "index": 8 }, { "bbox": [ 107, 182, 506, 198 ], "spans": [ { "bbox": [ 107, 184, 172, 195 ], "score": 0.89, "content": "\\boldsymbol { z } _ { t } = \\alpha _ { t } \\boldsymbol { z } + \\sigma _ { t } \\boldsymbol { \\epsilon }", "type": "inline_equation" }, { "bbox": [ 172, 182, 201, 198 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 202, 183, 252, 195 ], "score": 0.82, "content": "\\epsilon \\sim \\mathcal { N } ( 0 , I )", "type": "inline_equation" }, { "bbox": [ 253, 182, 256, 198 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 257, 183, 309, 195 ], "score": 0.86, "content": "\\alpha _ { t } : = 1 - \\beta _ { t }", "type": "inline_equation" }, { "bbox": [ 309, 182, 326, 198 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 327, 182, 406, 196 ], "score": 0.92, "content": "\\begin{array} { r } { \\sigma _ { t } : = \\dot { 1 } - \\prod _ { s = 1 } ^ { t } \\dot { \\alpha } _ { s } } \\end{array}", "type": "inline_equation" }, { "bbox": [ 407, 182, 506, 198 ], "score": 1.0, "content": ". The diffuser learns how", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 194, 505, 207 ], "spans": [ { "bbox": [ 105, 194, 172, 207 ], "score": 1.0, "content": "to denoise from", "type": "text" }, { "bbox": [ 173, 194, 192, 206 ], "score": 0.92, "content": "\\left\\{ { z } _ { t } \\right\\}", "type": "inline_equation" }, { "bbox": [ 193, 194, 237, 207 ], "score": 1.0, "content": "to recover", "type": "text" }, { "bbox": [ 237, 196, 244, 204 ], "score": 0.71, "content": "_ z", "type": "inline_equation" }, { "bbox": [ 244, 194, 505, 207 ], "score": 1.0, "content": ". Following the reparameterization method proposed in [20], the", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 205, 322, 217 ], "spans": [ { "bbox": [ 106, 205, 322, 217 ], "score": 1.0, "content": "denoising training objective can be expressed as [41]:", "type": "text" } ], "index": 11 } ], "index": 9.5, "bbox_fs": [ 105, 171, 506, 217 ] }, { "type": "interline_equation", "bbox": [ 231, 221, 380, 236 ], "lines": [ { "bbox": [ 231, 221, 380, 236 ], "spans": [ { "bbox": [ 231, 221, 380, 236 ], "score": 0.91, "content": "\\begin{array} { r } { \\mathcal { L } _ { D } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta } ( z _ { t } , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } . } \\end{array}", "type": "interline_equation", "image_path": "c0b6a42a7dddc81d5d528ed8a349c84f1a8e8e52d97b6eb0c26878eaf69606c7.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 231, 221, 380, 236 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 107, 241, 506, 254 ], "lines": [ { "bbox": [ 105, 239, 507, 256 ], "spans": [ { "bbox": [ 105, 239, 507, 256 ], "score": 1.0, "content": "In data generation, the denoising process can be realized through reparameterized Gaussian sampling:", "type": "text" } ], "index": 13 } ], "index": 13, "bbox_fs": [ 105, 239, 507, 256 ] }, { "type": "interline_equation", "bbox": [ 194, 261, 416, 289 ], "lines": [ { "bbox": [ 194, 261, 416, 289 ], "spans": [ { "bbox": [ 194, 261, 416, 289 ], "score": 0.94, "content": "p ( z _ { t - 1 } | z _ { t } ) = \\mathcal { N } \\left( z _ { t - 1 } ; \\frac { 1 } { \\sqrt { \\alpha _ { t } } } \\left( z _ { t } - \\frac { \\beta _ { t } } { \\sqrt { \\sigma _ { t } } } \\epsilon _ { \\theta } \\right) , \\beta _ { t } I \\right) .", "type": "interline_equation", "image_path": "cd64023bfb25f4aca1f1f00e845397403302ab368fd5956a291fe57286f62be3.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 194, 261, 416, 275.0 ], "spans": [], "index": 14 }, { "bbox": [ 194, 275.0, 416, 289.0 ], "spans": [], "index": 15 } ] }, { "type": "text", "bbox": [ 106, 292, 506, 358 ], "lines": [ { "bbox": [ 106, 292, 506, 304 ], "spans": [ { "bbox": [ 106, 293, 117, 304 ], "score": 1.0, "content": "In", "type": "text" }, { "bbox": [ 118, 293, 133, 304 ], "score": 0.89, "content": "\\mathcal { L } _ { D }", "type": "inline_equation" }, { "bbox": [ 133, 293, 231, 304 ], "score": 1.0, "content": ", the diffusion time step", "type": "text" }, { "bbox": [ 232, 292, 280, 304 ], "score": 0.86, "content": "t \\sim \\mathcal { U } [ 1 , T ]", "type": "inline_equation" }, { "bbox": [ 280, 293, 284, 304 ], "score": 1.0, "content": ";", "type": "text" }, { "bbox": [ 284, 294, 295, 304 ], "score": 0.65, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 295, 293, 506, 304 ], "score": 1.0, "content": "is a denoising model with UNet backbone parame-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 149, 316 ], "score": 1.0, "content": "terized by", "type": "text" }, { "bbox": [ 150, 304, 168, 315 ], "score": 0.36, "content": "\\theta ; { \\boldsymbol { y } }", "type": "inline_equation" }, { "bbox": [ 169, 303, 470, 316 ], "score": 1.0, "content": "represents the conditional variable that can be used to control generation;", "type": "text" }, { "bbox": [ 470, 304, 479, 313 ], "score": 0.83, "content": "C", "type": "inline_equation" }, { "bbox": [ 479, 303, 505, 316 ], "score": 1.0, "content": "is the", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 315, 505, 327 ], "spans": [ { "bbox": [ 105, 315, 432, 327 ], "score": 1.0, "content": "prompt encoder. The conditioning mechanism is implemented by first featurizing", "type": "text" }, { "bbox": [ 432, 316, 440, 326 ], "score": 0.78, "content": "\\textbf { { y } }", "type": "inline_equation" }, { "bbox": [ 440, 315, 459, 327 ], "score": 1.0, "content": "into", "type": "text" }, { "bbox": [ 459, 315, 481, 326 ], "score": 0.91, "content": "C ( \\boldsymbol { y } )", "type": "inline_equation" }, { "bbox": [ 482, 315, 505, 327 ], "score": 1.0, "content": ", then", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 325, 505, 338 ], "spans": [ { "bbox": [ 106, 325, 146, 338 ], "score": 1.0, "content": "the UNet", "type": "text" }, { "bbox": [ 146, 327, 157, 336 ], "score": 0.85, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 158, 325, 217, 338 ], "score": 1.0, "content": "conditions on", "type": "text" }, { "bbox": [ 218, 326, 240, 338 ], "score": 0.91, "content": "C ( \\boldsymbol { y } )", "type": "inline_equation" }, { "bbox": [ 240, 325, 505, 338 ], "score": 1.0, "content": "via cross-attention, as described in [41]. Distinct from previous", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 337, 506, 348 ], "spans": [ { "bbox": [ 106, 337, 506, 348 ], "score": 1.0, "content": "works, our model can condition on any combinations of modalities of text, image, video and audio.", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 347, 291, 359 ], "spans": [ { "bbox": [ 106, 347, 291, 359 ], "score": 1.0, "content": "Details are presented in the following section.", "type": "text" } ], "index": 21 } ], "index": 18.5, "bbox_fs": [ 105, 292, 506, 359 ] }, { "type": "title", "bbox": [ 108, 372, 294, 384 ], "lines": [ { "bbox": [ 105, 370, 296, 387 ], "spans": [ { "bbox": [ 105, 370, 296, 387 ], "score": 1.0, "content": "3.2 Composable Multimodal Conditioning", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 392, 505, 470 ], "lines": [ { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "To enable our model to condition on any combination of input/prompt modalities, we align the prompt", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 404, 505, 416 ], "spans": [ { "bbox": [ 106, 404, 315, 416 ], "score": 1.0, "content": "encoder of text, image, video and audio (denoted by", "type": "text" }, { "bbox": [ 316, 404, 327, 415 ], "score": 0.73, "content": "C _ { t }", "type": "inline_equation" }, { "bbox": [ 327, 404, 331, 416 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 331, 404, 342, 415 ], "score": 0.74, "content": "C _ { i }", "type": "inline_equation" }, { "bbox": [ 343, 404, 346, 416 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 346, 404, 359, 415 ], "score": 0.78, "content": "C _ { v }", "type": "inline_equation" }, { "bbox": [ 360, 404, 380, 416 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 380, 404, 393, 415 ], "score": 0.89, "content": "C _ { a }", "type": "inline_equation" }, { "bbox": [ 393, 404, 505, 416 ], "score": 1.0, "content": ", respectively) to project the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 415, 505, 427 ], "spans": [ { "bbox": [ 106, 415, 505, 427 ], "score": 1.0, "content": "input from any modality into the same space. Multimodal conditioning can then be conveniently", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 424, 505, 441 ], "spans": [ { "bbox": [ 104, 424, 349, 441 ], "score": 1.0, "content": "achieved by interpolating the representations of each modality", "type": "text" }, { "bbox": [ 350, 426, 360, 435 ], "score": 0.55, "content": "m", "type": "inline_equation" }, { "bbox": [ 360, 424, 365, 441 ], "score": 1.0, "content": ":", "type": "text" }, { "bbox": [ 365, 425, 505, 438 ], "score": 0.91, "content": "\\begin{array} { r } { C ( x _ { t } , \\bar { x _ { i } } , x _ { v } , x _ { a } ) = \\sum _ { m } \\alpha _ { m } C ( \\bar { m } ) } \\end{array}", "type": "inline_equation" } ], "index": 26 }, { "bbox": [ 104, 435, 506, 452 ], "spans": [ { "bbox": [ 104, 435, 122, 452 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 122, 437, 201, 448 ], "score": 0.9, "content": "m \\in \\ b { x } _ { t } , \\ b { x } _ { i } , \\ b { x } _ { v } , \\ b { x } _ { a }", "type": "inline_equation" }, { "bbox": [ 201, 435, 228, 452 ], "score": 1.0, "content": ", with", "type": "text" }, { "bbox": [ 228, 436, 286, 449 ], "score": 0.91, "content": "\\textstyle \\sum _ { m } \\alpha _ { m } = 1", "type": "inline_equation" }, { "bbox": [ 286, 435, 506, 452 ], "score": 1.0, "content": ". Through simple weighted interpolation of aligned", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 447, 505, 460 ], "spans": [ { "bbox": [ 106, 447, 505, 460 ], "score": 1.0, "content": "embeddings, we enable models trained with single-conditioning (i.e., with only one input) to perform", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 458, 501, 471 ], "spans": [ { "bbox": [ 106, 458, 501, 471 ], "score": 1.0, "content": "zero-shot multi-conditioning (i.e., with multiple inputs). This process is illustrated in Fig. 2 (a)(2).", "type": "text" } ], "index": 29 } ], "index": 26, "bbox_fs": [ 104, 392, 506, 471 ] }, { "type": "text", "bbox": [ 106, 474, 505, 573 ], "lines": [ { "bbox": [ 106, 474, 505, 487 ], "spans": [ { "bbox": [ 106, 474, 505, 487 ], "score": 1.0, "content": "Optimizing all four prompt encoders simultaneously in a combinatorial manner is computationally", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 484, 506, 500 ], "spans": [ { "bbox": [ 105, 484, 155, 500 ], "score": 1.0, "content": "heavy, with", "type": "text" }, { "bbox": [ 155, 485, 182, 497 ], "score": 0.93, "content": "\\mathcal { O } ( n ^ { 2 } )", "type": "inline_equation" }, { "bbox": [ 183, 484, 506, 500 ], "score": 1.0, "content": "pairs. Additionally, for certain dual modalities, well-aligned paired datasets are", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 495, 505, 509 ], "spans": [ { "bbox": [ 105, 495, 505, 509 ], "score": 1.0, "content": "limited or unavailable e.g., image-audio pairs. To address this challenge, we propose a simple and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 506, 505, 520 ], "spans": [ { "bbox": [ 105, 506, 505, 520 ], "score": 1.0, "content": "effective technique called \"Bridging Alignment\" to efficiently align conditional encoders. As shown", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 518, 506, 532 ], "spans": [ { "bbox": [ 105, 518, 506, 532 ], "score": 1.0, "content": "in Fig. 2 (a)(1), we choose the text modality as the \"bridging\" modality due to its ubiquitous presence", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 529, 505, 542 ], "spans": [ { "bbox": [ 105, 529, 505, 542 ], "score": 1.0, "content": "in paired data, such as text-image, text-video, and text-audio pairs. We begin with a pretrained", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 540, 505, 552 ], "spans": [ { "bbox": [ 105, 540, 505, 552 ], "score": 1.0, "content": "text-image paired encoder, i.e., CLIP [38]. We then train audio and video prompt encoders on", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 551, 506, 564 ], "spans": [ { "bbox": [ 106, 551, 506, 564 ], "score": 1.0, "content": "audio-text and video-text paired datasets using contrastive learning, with text and image encoder", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 561, 169, 575 ], "spans": [ { "bbox": [ 105, 561, 169, 575 ], "score": 1.0, "content": "weights frozen.", "type": "text" } ], "index": 38 } ], "index": 34, "bbox_fs": [ 105, 474, 506, 575 ] }, { "type": "text", "bbox": [ 107, 578, 505, 644 ], "lines": [ { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "In this way, all four modalities are aligned in the feature space. As shown in Section 5.2, CoDi", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 589, 505, 602 ], "spans": [ { "bbox": [ 105, 589, 505, 602 ], "score": 1.0, "content": "can effectively leverage and combine the complementary information present in any combination", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "of modalities to generate more accurate and comprehensive outputs. The high generation quality", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 611, 505, 622 ], "spans": [ { "bbox": [ 106, 611, 505, 622 ], "score": 1.0, "content": "remains unaffected with respect to the number of prompt modalities. As we will discuss in subsequent", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "sections, we continue to apply Bridging Alignment to align the latent space of LDMs with different", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 632, 309, 645 ], "spans": [ { "bbox": [ 105, 632, 309, 645 ], "score": 1.0, "content": "modalities to achieve joint multimodal generation.", "type": "text" } ], "index": 44 } ], "index": 41.5, "bbox_fs": [ 105, 578, 506, 645 ] }, { "type": "title", "bbox": [ 107, 657, 225, 669 ], "lines": [ { "bbox": [ 104, 656, 226, 672 ], "spans": [ { "bbox": [ 104, 656, 226, 672 ], "score": 1.0, "content": "3.3 Composable Diffusion", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 678, 505, 722 ], "lines": [ { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 505, 691 ], "score": 1.0, "content": "Training an end-to-end anything-to-anything model requires extensive learning on various data", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 689, 505, 701 ], "spans": [ { "bbox": [ 105, 689, 505, 701 ], "score": 1.0, "content": "resources. The model also needs to maintain generation quality for all synthesis flows. To address", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 506, 712 ], "spans": [ { "bbox": [ 105, 700, 506, 712 ], "score": 1.0, "content": "these challenges, CoDi is designed to be composable and integrative, allowing individual modality-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 711, 505, 724 ], "spans": [ { "bbox": [ 105, 711, 505, 724 ], "score": 1.0, "content": "specific models to be built independently and then smoothly integrated later. Specifically, we start by", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 72, 505, 85 ], "spans": [ { "bbox": [ 105, 72, 505, 85 ], "score": 1.0, "content": "independently training image, video, audio, and text LDMs. These diffusion models then efficiently", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 84, 505, 96 ], "spans": [ { "bbox": [ 106, 84, 505, 96 ], "score": 1.0, "content": "learn to attend across modalities for joint multimodal generation (Section 3.4) by a novel mechanism", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 106, 94, 213, 106 ], "spans": [ { "bbox": [ 106, 94, 213, 106 ], "score": 1.0, "content": "named “latent alignment”.", "type": "text", "cross_page": true } ], "index": 2 } ], "index": 47.5, "bbox_fs": [ 105, 678, 506, 724 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 72, 504, 105 ], "lines": [ { "bbox": [ 105, 72, 505, 85 ], "spans": [ { "bbox": [ 105, 72, 505, 85 ], "score": 1.0, "content": "independently training image, video, audio, and text LDMs. These diffusion models then efficiently", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 84, 505, 96 ], "spans": [ { "bbox": [ 106, 84, 505, 96 ], "score": 1.0, "content": "learn to attend across modalities for joint multimodal generation (Section 3.4) by a novel mechanism", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 94, 213, 106 ], "spans": [ { "bbox": [ 106, 94, 213, 106 ], "score": 1.0, "content": "named “latent alignment”.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 107, 116, 505, 150 ], "lines": [ { "bbox": [ 105, 116, 505, 129 ], "spans": [ { "bbox": [ 105, 116, 505, 129 ], "score": 1.0, "content": "Image Diffusion Model. The image LDM follows the same structure as Stable Diffusion 1.5 [41]", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 128, 505, 140 ], "spans": [ { "bbox": [ 106, 128, 505, 140 ], "score": 1.0, "content": "and is initialized with the same weights. Reusing the weights transfers the knowledge and exceptional", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 138, 497, 151 ], "spans": [ { "bbox": [ 105, 138, 497, 151 ], "score": 1.0, "content": "generation fidelity of Stable Diffusion trained on large-scale high-quality image datasets to CoDi.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 106, 161, 505, 292 ], "lines": [ { "bbox": [ 106, 160, 505, 174 ], "spans": [ { "bbox": [ 106, 160, 505, 174 ], "score": 1.0, "content": "Video Diffusion Model. To model the temporal properties of videos and simultaneously maintain", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 172, 505, 185 ], "spans": [ { "bbox": [ 106, 172, 505, 185 ], "score": 1.0, "content": "vision generation quality, we construct the video diffuser by extending the image diffuser with", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "temporal modules. Specifically, we insert pseudo-temporal attention before the residual block [13].", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 194, 505, 206 ], "spans": [ { "bbox": [ 106, 194, 505, 206 ], "score": 1.0, "content": "However, we argue that pseudo-temporal attention only enables video frames to globally attend", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 205, 505, 217 ], "spans": [ { "bbox": [ 105, 205, 505, 217 ], "score": 1.0, "content": "to each other by flattening the pixels (height, width dimension) to batch dimension, resulting in", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 216, 505, 228 ], "spans": [ { "bbox": [ 105, 216, 505, 228 ], "score": 1.0, "content": "a lack of cross-frame interaction between local pixels. We argue that this results in the common", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 226, 505, 239 ], "spans": [ { "bbox": [ 105, 226, 505, 239 ], "score": 1.0, "content": "temporal-inconsistency issue in video generation that locations, shapes, colors, etc. of objects can", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 237, 505, 250 ], "spans": [ { "bbox": [ 106, 237, 505, 250 ], "score": 1.0, "content": "be inconsistent across generated frames. To address this problem, we propose adapting the latent", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 249, 506, 261 ], "spans": [ { "bbox": [ 106, 249, 506, 261 ], "score": 1.0, "content": "shift method [2] that performs temporal-spatial shifts on latent features in accordance with temporal", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 259, 505, 271 ], "spans": [ { "bbox": [ 106, 259, 359, 271 ], "score": 1.0, "content": "attention. We divide the video by the hidden dimension into", "type": "text" }, { "bbox": [ 360, 259, 387, 270 ], "score": 0.9, "content": "k = 8", "type": "inline_equation" }, { "bbox": [ 388, 259, 505, 271 ], "score": 1.0, "content": "chunks, and for each chunk", "type": "text" } ], "index": 15 }, { "bbox": [ 107, 270, 506, 282 ], "spans": [ { "bbox": [ 107, 271, 129, 280 ], "score": 0.88, "content": "i = 0", "type": "inline_equation" }, { "bbox": [ 130, 270, 329, 282 ], "score": 1.0, "content": "to 7, we shift the temporal dimension forward by", "type": "text" }, { "bbox": [ 329, 271, 334, 280 ], "score": 0.71, "content": "i", "type": "inline_equation" }, { "bbox": [ 334, 270, 506, 282 ], "score": 1.0, "content": "positions. Further details will be provided", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 281, 173, 294 ], "spans": [ { "bbox": [ 106, 281, 173, 294 ], "score": 1.0, "content": "in the appendix.", "type": "text" } ], "index": 17 } ], "index": 11.5 }, { "type": "text", "bbox": [ 107, 303, 505, 369 ], "lines": [ { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "Audio Diffusion Model. To enable flexible cross-modality attention in joint generation, the audio", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 313, 506, 327 ], "spans": [ { "bbox": [ 105, 313, 506, 327 ], "score": 1.0, "content": "diffuser is designed to have a similar architecture to vision diffusers, where the mel-spectrogram", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 325, 506, 336 ], "spans": [ { "bbox": [ 106, 325, 506, 336 ], "score": 1.0, "content": "can be naturally viewed as an image with 1 channel. We use a VAE encoder to encode the mel-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 336, 506, 348 ], "spans": [ { "bbox": [ 105, 336, 506, 348 ], "score": 1.0, "content": "spectrogram of audio to a compressed latent space. In audio synthesis, a VAE decoder maps the latent", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 347, 506, 360 ], "spans": [ { "bbox": [ 105, 347, 506, 360 ], "score": 1.0, "content": "variable to the mel-spectrogram, and a vocoder generates the audio sample from the mel-spectrogram.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 358, 365, 370 ], "spans": [ { "bbox": [ 106, 358, 365, 370 ], "score": 1.0, "content": "We employ the audio VAE from [33] and the vocoder from [27].", "type": "text" } ], "index": 23 } ], "index": 20.5 }, { "type": "text", "bbox": [ 107, 380, 504, 414 ], "lines": [ { "bbox": [ 106, 380, 505, 392 ], "spans": [ { "bbox": [ 106, 380, 505, 392 ], "score": 1.0, "content": "Text Diffusion Model. The VAE of the text LDM is OPTIMUS [29], and its encoder and decoder", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 392, 505, 404 ], "spans": [ { "bbox": [ 105, 392, 505, 404 ], "score": 1.0, "content": "are [9] and GPT-2 [39], respectively. For the denoising UNet, unlike the one in image diffusion, the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 402, 398, 415 ], "spans": [ { "bbox": [ 105, 402, 398, 415 ], "score": 1.0, "content": "2D convolution in residual blocks is replaced with 1D convolution [53].", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "title", "bbox": [ 107, 425, 346, 437 ], "lines": [ { "bbox": [ 105, 424, 348, 441 ], "spans": [ { "bbox": [ 105, 424, 348, 441 ], "score": 1.0, "content": "3.4 Joint Multimodal Generation by Latent Alignment", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 106, 445, 505, 534 ], "lines": [ { "bbox": [ 105, 446, 505, 459 ], "spans": [ { "bbox": [ 105, 446, 505, 459 ], "score": 1.0, "content": "The final step is to enable cross-attention between diffusion flows in joint generation, i.e., generating", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 458, 505, 469 ], "spans": [ { "bbox": [ 106, 458, 505, 469 ], "score": 1.0, "content": "two or more modalities simultaneously. This is achieved by adding cross-modal attention sublayers to", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 468, 505, 480 ], "spans": [ { "bbox": [ 106, 468, 145, 480 ], "score": 1.0, "content": "the UNet", "type": "text" }, { "bbox": [ 145, 469, 155, 479 ], "score": 0.85, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 156, 468, 424, 480 ], "score": 1.0, "content": "(Fig. 2 (b)(2)). Specifically, consider a diffusion model of modality", "type": "text" }, { "bbox": [ 424, 468, 433, 478 ], "score": 0.78, "content": "A", "type": "inline_equation" }, { "bbox": [ 433, 468, 505, 480 ], "score": 1.0, "content": "that cross-attends", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 478, 505, 491 ], "spans": [ { "bbox": [ 105, 478, 199, 491 ], "score": 1.0, "content": "with another modality", "type": "text" }, { "bbox": [ 199, 479, 208, 489 ], "score": 0.81, "content": "B", "type": "inline_equation" }, { "bbox": [ 208, 478, 365, 491 ], "score": 1.0, "content": ". Let the latent variables of modalities", "type": "text" }, { "bbox": [ 365, 480, 381, 490 ], "score": 0.86, "content": "m _ { A }", "type": "inline_equation" }, { "bbox": [ 381, 478, 400, 491 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 400, 480, 417, 490 ], "score": 0.87, "content": "m _ { B }", "type": "inline_equation" }, { "bbox": [ 417, 478, 486, 491 ], "score": 1.0, "content": "at diffusion step", "type": "text" }, { "bbox": [ 487, 480, 492, 489 ], "score": 0.72, "content": "t", "type": "inline_equation" }, { "bbox": [ 492, 478, 505, 491 ], "score": 1.0, "content": "be", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 488, 507, 503 ], "spans": [ { "bbox": [ 105, 488, 153, 503 ], "score": 1.0, "content": "denoted as", "type": "text" }, { "bbox": [ 153, 489, 167, 502 ], "score": 0.91, "content": " { \\boldsymbol { z } } _ { t } ^ { A }", "type": "inline_equation" }, { "bbox": [ 168, 488, 186, 503 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 187, 489, 200, 501 ], "score": 0.91, "content": "\\hat { z _ { t } ^ { B } }", "type": "inline_equation" }, { "bbox": [ 201, 488, 507, 503 ], "score": 1.0, "content": ", respectively. The proposed “Latent Alignment” technique is such that a", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 500, 506, 514 ], "spans": [ { "bbox": [ 105, 500, 267, 514 ], "score": 1.0, "content": "modality-specific environment encoder", "type": "text" }, { "bbox": [ 267, 501, 280, 511 ], "score": 0.89, "content": "V _ { B }", "type": "inline_equation" }, { "bbox": [ 281, 500, 334, 514 ], "score": 1.0, "content": "first projects", "type": "text" }, { "bbox": [ 335, 500, 348, 512 ], "score": 0.9, "content": " { \\boldsymbol { z } } _ { t } ^ { B }", "type": "inline_equation" }, { "bbox": [ 349, 500, 506, 514 ], "score": 1.0, "content": "into a shared latent space for different", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 511, 505, 524 ], "spans": [ { "bbox": [ 105, 511, 343, 524 ], "score": 1.0, "content": "modalities. Then, in each layer of the UNet for modality", "type": "text" }, { "bbox": [ 343, 513, 351, 521 ], "score": 0.79, "content": "A", "type": "inline_equation" }, { "bbox": [ 352, 511, 505, 524 ], "score": 1.0, "content": ", a cross-attention sublayer attends to", "type": "text" } ], "index": 34 }, { "bbox": [ 107, 522, 490, 535 ], "spans": [ { "bbox": [ 107, 522, 141, 534 ], "score": 0.93, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 141, 522, 290, 535 ], "score": 1.0, "content": ". For the diffusion model of modality", "type": "text" }, { "bbox": [ 291, 523, 299, 532 ], "score": 0.79, "content": "A", "type": "inline_equation" }, { "bbox": [ 299, 522, 490, 535 ], "score": 1.0, "content": ", the training objective in Eq. (1) now becomes:", "type": "text" } ], "index": 35 } ], "index": 31.5 }, { "type": "interline_equation", "bbox": [ 201, 547, 409, 563 ], "lines": [ { "bbox": [ 201, 547, 409, 563 ], "spans": [ { "bbox": [ 201, 547, 409, 563 ], "score": 0.9, "content": "\\mathcal { L } _ { C r o s s } ^ { A } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta _ { c } } ( z _ { t } ^ { A } , V _ { B } ( z _ { t } ^ { B } ) , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } ,", "type": "interline_equation", "image_path": "543f8049db9dbb90f598c7dea6ef4aa9f7c34c5bc297bb830ab29bb93064c182.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 201, 547, 409, 563 ], "spans": [], "index": 36 } ] }, { "type": "text", "bbox": [ 104, 568, 386, 579 ], "lines": [ { "bbox": [ 106, 567, 385, 581 ], "spans": [ { "bbox": [ 106, 567, 133, 581 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 568, 143, 579 ], "score": 0.88, "content": "\\theta _ { c }", "type": "inline_equation" }, { "bbox": [ 143, 567, 385, 581 ], "score": 1.0, "content": "denotes the weights of cross-attention modules in the UNet.", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 106, 583, 505, 695 ], "lines": [ { "bbox": [ 103, 576, 505, 603 ], "spans": [ { "bbox": [ 103, 576, 205, 603 ], "score": 1.0, "content": "The training objective of", "type": "text" }, { "bbox": [ 205, 584, 232, 595 ], "score": 0.88, "content": "A + B", "type": "inline_equation" }, { "bbox": [ 233, 576, 306, 603 ], "score": 1.0, "content": "joint generation is", "type": "text" }, { "bbox": [ 307, 583, 376, 596 ], "score": 0.85, "content": "\\mathcal { L } _ { C r o s s } ^ { A } + \\mathcal { L } _ { C r o s s } ^ { B }", "type": "inline_equation" }, { "bbox": [ 376, 576, 380, 603 ], "score": 1.0, "content": ".", "type": "text" }, { "bbox": [ 380, 583, 400, 596 ], "score": 0.8, "content": "V ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 401, 576, 505, 603 ], "score": 1.0, "content": "of different modalities are", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 594, 506, 611 ], "spans": [ { "bbox": [ 104, 594, 315, 611 ], "score": 1.0, "content": "trained to be aligned with contrastive learning. Since", "type": "text" }, { "bbox": [ 315, 596, 327, 608 ], "score": 0.9, "content": "z _ { t } ^ { A }", "type": "inline_equation" }, { "bbox": [ 327, 594, 345, 611 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 345, 596, 358, 608 ], "score": 0.9, "content": "z _ { t } ^ { B }", "type": "inline_equation" }, { "bbox": [ 358, 594, 506, 611 ], "score": 1.0, "content": "at any time step can be sampled with", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 606, 505, 621 ], "spans": [ { "bbox": [ 105, 606, 505, 621 ], "score": 1.0, "content": "closed form in the diffusion process Section 3.1, one can conveniently train the contrastive learning", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 617, 506, 632 ], "spans": [ { "bbox": [ 105, 617, 160, 632 ], "score": 1.0, "content": "together with", "type": "text" }, { "bbox": [ 160, 618, 190, 629 ], "score": 0.91, "content": "\\mathcal { L } _ { C r o s s }", "type": "inline_equation" }, { "bbox": [ 191, 617, 255, 632 ], "score": 1.0, "content": ". The purpose of", "type": "text" }, { "bbox": [ 256, 618, 265, 628 ], "score": 0.78, "content": "V", "type": "inline_equation" }, { "bbox": [ 265, 617, 506, 632 ], "score": 1.0, "content": "is to achieve the generation of any combination of modalities", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 629, 505, 642 ], "spans": [ { "bbox": [ 106, 629, 505, 642 ], "score": 1.0, "content": "(in polynomial) by training on a linear number of joint-generation tasks. For example, if we have", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 638, 507, 654 ], "spans": [ { "bbox": [ 104, 638, 278, 654 ], "score": 1.0, "content": "trained the joint generation of modalities", "type": "text" }, { "bbox": [ 279, 640, 301, 650 ], "score": 0.38, "content": "A , B", "type": "inline_equation" }, { "bbox": [ 301, 638, 323, 654 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 324, 640, 333, 650 ], "score": 0.69, "content": "B", "type": "inline_equation" }, { "bbox": [ 333, 638, 337, 654 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 338, 640, 347, 650 ], "score": 0.68, "content": "C", "type": "inline_equation" }, { "bbox": [ 347, 638, 468, 654 ], "score": 1.0, "content": "independently, then we have", "type": "text" }, { "bbox": [ 469, 639, 502, 652 ], "score": 0.92, "content": "V _ { A } ( z _ { t } ^ { A } )", "type": "inline_equation" }, { "bbox": [ 502, 638, 507, 654 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 43 }, { "bbox": [ 107, 649, 506, 664 ], "spans": [ { "bbox": [ 107, 650, 141, 662 ], "score": 0.92, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 141, 649, 160, 664 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 161, 650, 195, 663 ], "score": 0.93, "content": "V _ { C } ( z _ { t } ^ { C } )", "type": "inline_equation" }, { "bbox": [ 195, 649, 506, 664 ], "score": 1.0, "content": "aligned. Therefore, CoDi can seamlessly achieve joint generation of modalities", "type": "text" } ], "index": 44 }, { "bbox": [ 107, 661, 505, 675 ], "spans": [ { "bbox": [ 107, 663, 115, 672 ], "score": 0.77, "content": "A", "type": "inline_equation" }, { "bbox": [ 116, 661, 133, 675 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 134, 663, 142, 672 ], "score": 0.84, "content": "C", "type": "inline_equation" }, { "bbox": [ 143, 661, 505, 675 ], "score": 1.0, "content": "without any additional training. Moreover, such design automatically effortlessly enables", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 672, 505, 685 ], "spans": [ { "bbox": [ 104, 672, 229, 685 ], "score": 1.0, "content": "joint generation of modalities", "type": "text" }, { "bbox": [ 230, 674, 237, 682 ], "score": 0.66, "content": "A", "type": "inline_equation" }, { "bbox": [ 237, 672, 241, 685 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 242, 674, 251, 682 ], "score": 0.71, "content": "B", "type": "inline_equation" }, { "bbox": [ 251, 672, 272, 685 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 272, 674, 281, 682 ], "score": 0.86, "content": "C", "type": "inline_equation" }, { "bbox": [ 281, 672, 426, 685 ], "score": 1.0, "content": "concurrently. Specifically, UNet of", "type": "text" }, { "bbox": [ 427, 673, 436, 682 ], "score": 0.76, "content": "A", "type": "inline_equation" }, { "bbox": [ 436, 672, 505, 685 ], "score": 1.0, "content": "can cross-attend", "type": "text" } ], "index": 46 }, { "bbox": [ 104, 682, 506, 698 ], "spans": [ { "bbox": [ 104, 682, 205, 698 ], "score": 1.0, "content": "with the interpolation of", "type": "text" }, { "bbox": [ 205, 684, 239, 695 ], "score": 0.93, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 240, 682, 261, 698 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 261, 684, 294, 695 ], "score": 0.92, "content": "V _ { C } ( z _ { t } ^ { C } )", "type": "inline_equation" }, { "bbox": [ 294, 682, 506, 698 ], "score": 1.0, "content": ", although CoDi has not been trained with such task.", "type": "text" } ], "index": 47 } ], "index": 42.5 }, { "type": "text", "bbox": [ 106, 699, 503, 722 ], "lines": [ { "bbox": [ 105, 699, 505, 713 ], "spans": [ { "bbox": [ 105, 699, 505, 713 ], "score": 1.0, "content": "As shown in Fig. 2(b)(3), we follow similar designs to the \"Bridging Alignment\" in training joint", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "generation: (1) We first train the cross-attention weights in the image and text diffusers, as well", "type": "text" } ], "index": 49 } ], "index": 48.5 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 308, 750 ], "lines": [ { "bbox": [ 302, 740, 309, 753 ], "spans": [ { "bbox": [ 302, 740, 309, 753 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 72, 504, 105 ], "lines": [], "index": 1, "bbox_fs": [ 105, 72, 505, 106 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 116, 505, 150 ], "lines": [ { "bbox": [ 105, 116, 505, 129 ], "spans": [ { "bbox": [ 105, 116, 505, 129 ], "score": 1.0, "content": "Image Diffusion Model. The image LDM follows the same structure as Stable Diffusion 1.5 [41]", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 128, 505, 140 ], "spans": [ { "bbox": [ 106, 128, 505, 140 ], "score": 1.0, "content": "and is initialized with the same weights. Reusing the weights transfers the knowledge and exceptional", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 138, 497, 151 ], "spans": [ { "bbox": [ 105, 138, 497, 151 ], "score": 1.0, "content": "generation fidelity of Stable Diffusion trained on large-scale high-quality image datasets to CoDi.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 105, 116, 505, 151 ] }, { "type": "text", "bbox": [ 106, 161, 505, 292 ], "lines": [ { "bbox": [ 106, 160, 505, 174 ], "spans": [ { "bbox": [ 106, 160, 505, 174 ], "score": 1.0, "content": "Video Diffusion Model. To model the temporal properties of videos and simultaneously maintain", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 172, 505, 185 ], "spans": [ { "bbox": [ 106, 172, 505, 185 ], "score": 1.0, "content": "vision generation quality, we construct the video diffuser by extending the image diffuser with", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "temporal modules. Specifically, we insert pseudo-temporal attention before the residual block [13].", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 194, 505, 206 ], "spans": [ { "bbox": [ 106, 194, 505, 206 ], "score": 1.0, "content": "However, we argue that pseudo-temporal attention only enables video frames to globally attend", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 205, 505, 217 ], "spans": [ { "bbox": [ 105, 205, 505, 217 ], "score": 1.0, "content": "to each other by flattening the pixels (height, width dimension) to batch dimension, resulting in", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 216, 505, 228 ], "spans": [ { "bbox": [ 105, 216, 505, 228 ], "score": 1.0, "content": "a lack of cross-frame interaction between local pixels. We argue that this results in the common", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 226, 505, 239 ], "spans": [ { "bbox": [ 105, 226, 505, 239 ], "score": 1.0, "content": "temporal-inconsistency issue in video generation that locations, shapes, colors, etc. of objects can", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 237, 505, 250 ], "spans": [ { "bbox": [ 106, 237, 505, 250 ], "score": 1.0, "content": "be inconsistent across generated frames. To address this problem, we propose adapting the latent", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 249, 506, 261 ], "spans": [ { "bbox": [ 106, 249, 506, 261 ], "score": 1.0, "content": "shift method [2] that performs temporal-spatial shifts on latent features in accordance with temporal", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 259, 505, 271 ], "spans": [ { "bbox": [ 106, 259, 359, 271 ], "score": 1.0, "content": "attention. We divide the video by the hidden dimension into", "type": "text" }, { "bbox": [ 360, 259, 387, 270 ], "score": 0.9, "content": "k = 8", "type": "inline_equation" }, { "bbox": [ 388, 259, 505, 271 ], "score": 1.0, "content": "chunks, and for each chunk", "type": "text" } ], "index": 15 }, { "bbox": [ 107, 270, 506, 282 ], "spans": [ { "bbox": [ 107, 271, 129, 280 ], "score": 0.88, "content": "i = 0", "type": "inline_equation" }, { "bbox": [ 130, 270, 329, 282 ], "score": 1.0, "content": "to 7, we shift the temporal dimension forward by", "type": "text" }, { "bbox": [ 329, 271, 334, 280 ], "score": 0.71, "content": "i", "type": "inline_equation" }, { "bbox": [ 334, 270, 506, 282 ], "score": 1.0, "content": "positions. Further details will be provided", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 281, 173, 294 ], "spans": [ { "bbox": [ 106, 281, 173, 294 ], "score": 1.0, "content": "in the appendix.", "type": "text" } ], "index": 17 } ], "index": 11.5, "bbox_fs": [ 105, 160, 506, 294 ] }, { "type": "text", "bbox": [ 107, 303, 505, 369 ], "lines": [ { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "Audio Diffusion Model. To enable flexible cross-modality attention in joint generation, the audio", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 313, 506, 327 ], "spans": [ { "bbox": [ 105, 313, 506, 327 ], "score": 1.0, "content": "diffuser is designed to have a similar architecture to vision diffusers, where the mel-spectrogram", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 325, 506, 336 ], "spans": [ { "bbox": [ 106, 325, 506, 336 ], "score": 1.0, "content": "can be naturally viewed as an image with 1 channel. We use a VAE encoder to encode the mel-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 336, 506, 348 ], "spans": [ { "bbox": [ 105, 336, 506, 348 ], "score": 1.0, "content": "spectrogram of audio to a compressed latent space. In audio synthesis, a VAE decoder maps the latent", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 347, 506, 360 ], "spans": [ { "bbox": [ 105, 347, 506, 360 ], "score": 1.0, "content": "variable to the mel-spectrogram, and a vocoder generates the audio sample from the mel-spectrogram.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 358, 365, 370 ], "spans": [ { "bbox": [ 106, 358, 365, 370 ], "score": 1.0, "content": "We employ the audio VAE from [33] and the vocoder from [27].", "type": "text" } ], "index": 23 } ], "index": 20.5, "bbox_fs": [ 105, 303, 506, 370 ] }, { "type": "text", "bbox": [ 107, 380, 504, 414 ], "lines": [ { "bbox": [ 106, 380, 505, 392 ], "spans": [ { "bbox": [ 106, 380, 505, 392 ], "score": 1.0, "content": "Text Diffusion Model. The VAE of the text LDM is OPTIMUS [29], and its encoder and decoder", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 392, 505, 404 ], "spans": [ { "bbox": [ 105, 392, 505, 404 ], "score": 1.0, "content": "are [9] and GPT-2 [39], respectively. For the denoising UNet, unlike the one in image diffusion, the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 402, 398, 415 ], "spans": [ { "bbox": [ 105, 402, 398, 415 ], "score": 1.0, "content": "2D convolution in residual blocks is replaced with 1D convolution [53].", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 105, 380, 505, 415 ] }, { "type": "title", "bbox": [ 107, 425, 346, 437 ], "lines": [ { "bbox": [ 105, 424, 348, 441 ], "spans": [ { "bbox": [ 105, 424, 348, 441 ], "score": 1.0, "content": "3.4 Joint Multimodal Generation by Latent Alignment", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 106, 445, 505, 534 ], "lines": [ { "bbox": [ 105, 446, 505, 459 ], "spans": [ { "bbox": [ 105, 446, 505, 459 ], "score": 1.0, "content": "The final step is to enable cross-attention between diffusion flows in joint generation, i.e., generating", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 458, 505, 469 ], "spans": [ { "bbox": [ 106, 458, 505, 469 ], "score": 1.0, "content": "two or more modalities simultaneously. This is achieved by adding cross-modal attention sublayers to", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 468, 505, 480 ], "spans": [ { "bbox": [ 106, 468, 145, 480 ], "score": 1.0, "content": "the UNet", "type": "text" }, { "bbox": [ 145, 469, 155, 479 ], "score": 0.85, "content": "\\epsilon _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 156, 468, 424, 480 ], "score": 1.0, "content": "(Fig. 2 (b)(2)). Specifically, consider a diffusion model of modality", "type": "text" }, { "bbox": [ 424, 468, 433, 478 ], "score": 0.78, "content": "A", "type": "inline_equation" }, { "bbox": [ 433, 468, 505, 480 ], "score": 1.0, "content": "that cross-attends", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 478, 505, 491 ], "spans": [ { "bbox": [ 105, 478, 199, 491 ], "score": 1.0, "content": "with another modality", "type": "text" }, { "bbox": [ 199, 479, 208, 489 ], "score": 0.81, "content": "B", "type": "inline_equation" }, { "bbox": [ 208, 478, 365, 491 ], "score": 1.0, "content": ". Let the latent variables of modalities", "type": "text" }, { "bbox": [ 365, 480, 381, 490 ], "score": 0.86, "content": "m _ { A }", "type": "inline_equation" }, { "bbox": [ 381, 478, 400, 491 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 400, 480, 417, 490 ], "score": 0.87, "content": "m _ { B }", "type": "inline_equation" }, { "bbox": [ 417, 478, 486, 491 ], "score": 1.0, "content": "at diffusion step", "type": "text" }, { "bbox": [ 487, 480, 492, 489 ], "score": 0.72, "content": "t", "type": "inline_equation" }, { "bbox": [ 492, 478, 505, 491 ], "score": 1.0, "content": "be", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 488, 507, 503 ], "spans": [ { "bbox": [ 105, 488, 153, 503 ], "score": 1.0, "content": "denoted as", "type": "text" }, { "bbox": [ 153, 489, 167, 502 ], "score": 0.91, "content": " { \\boldsymbol { z } } _ { t } ^ { A }", "type": "inline_equation" }, { "bbox": [ 168, 488, 186, 503 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 187, 489, 200, 501 ], "score": 0.91, "content": "\\hat { z _ { t } ^ { B } }", "type": "inline_equation" }, { "bbox": [ 201, 488, 507, 503 ], "score": 1.0, "content": ", respectively. The proposed “Latent Alignment” technique is such that a", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 500, 506, 514 ], "spans": [ { "bbox": [ 105, 500, 267, 514 ], "score": 1.0, "content": "modality-specific environment encoder", "type": "text" }, { "bbox": [ 267, 501, 280, 511 ], "score": 0.89, "content": "V _ { B }", "type": "inline_equation" }, { "bbox": [ 281, 500, 334, 514 ], "score": 1.0, "content": "first projects", "type": "text" }, { "bbox": [ 335, 500, 348, 512 ], "score": 0.9, "content": " { \\boldsymbol { z } } _ { t } ^ { B }", "type": "inline_equation" }, { "bbox": [ 349, 500, 506, 514 ], "score": 1.0, "content": "into a shared latent space for different", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 511, 505, 524 ], "spans": [ { "bbox": [ 105, 511, 343, 524 ], "score": 1.0, "content": "modalities. Then, in each layer of the UNet for modality", "type": "text" }, { "bbox": [ 343, 513, 351, 521 ], "score": 0.79, "content": "A", "type": "inline_equation" }, { "bbox": [ 352, 511, 505, 524 ], "score": 1.0, "content": ", a cross-attention sublayer attends to", "type": "text" } ], "index": 34 }, { "bbox": [ 107, 522, 490, 535 ], "spans": [ { "bbox": [ 107, 522, 141, 534 ], "score": 0.93, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 141, 522, 290, 535 ], "score": 1.0, "content": ". For the diffusion model of modality", "type": "text" }, { "bbox": [ 291, 523, 299, 532 ], "score": 0.79, "content": "A", "type": "inline_equation" }, { "bbox": [ 299, 522, 490, 535 ], "score": 1.0, "content": ", the training objective in Eq. (1) now becomes:", "type": "text" } ], "index": 35 } ], "index": 31.5, "bbox_fs": [ 105, 446, 507, 535 ] }, { "type": "interline_equation", "bbox": [ 201, 547, 409, 563 ], "lines": [ { "bbox": [ 201, 547, 409, 563 ], "spans": [ { "bbox": [ 201, 547, 409, 563 ], "score": 0.9, "content": "\\mathcal { L } _ { C r o s s } ^ { A } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta _ { c } } ( z _ { t } ^ { A } , V _ { B } ( z _ { t } ^ { B } ) , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } ,", "type": "interline_equation", "image_path": "543f8049db9dbb90f598c7dea6ef4aa9f7c34c5bc297bb830ab29bb93064c182.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 201, 547, 409, 563 ], "spans": [], "index": 36 } ] }, { "type": "text", "bbox": [ 104, 568, 386, 579 ], "lines": [ { "bbox": [ 106, 567, 385, 581 ], "spans": [ { "bbox": [ 106, 567, 133, 581 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 568, 143, 579 ], "score": 0.88, "content": "\\theta _ { c }", "type": "inline_equation" }, { "bbox": [ 143, 567, 385, 581 ], "score": 1.0, "content": "denotes the weights of cross-attention modules in the UNet.", "type": "text" } ], "index": 37 } ], "index": 37, "bbox_fs": [ 106, 567, 385, 581 ] }, { "type": "text", "bbox": [ 106, 583, 505, 695 ], "lines": [ { "bbox": [ 103, 576, 505, 603 ], "spans": [ { "bbox": [ 103, 576, 205, 603 ], "score": 1.0, "content": "The training objective of", "type": "text" }, { "bbox": [ 205, 584, 232, 595 ], "score": 0.88, "content": "A + B", "type": "inline_equation" }, { "bbox": [ 233, 576, 306, 603 ], "score": 1.0, "content": "joint generation is", "type": "text" }, { "bbox": [ 307, 583, 376, 596 ], "score": 0.85, "content": "\\mathcal { L } _ { C r o s s } ^ { A } + \\mathcal { L } _ { C r o s s } ^ { B }", "type": "inline_equation" }, { "bbox": [ 376, 576, 380, 603 ], "score": 1.0, "content": ".", "type": "text" }, { "bbox": [ 380, 583, 400, 596 ], "score": 0.8, "content": "V ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 401, 576, 505, 603 ], "score": 1.0, "content": "of different modalities are", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 594, 506, 611 ], "spans": [ { "bbox": [ 104, 594, 315, 611 ], "score": 1.0, "content": "trained to be aligned with contrastive learning. Since", "type": "text" }, { "bbox": [ 315, 596, 327, 608 ], "score": 0.9, "content": "z _ { t } ^ { A }", "type": "inline_equation" }, { "bbox": [ 327, 594, 345, 611 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 345, 596, 358, 608 ], "score": 0.9, "content": "z _ { t } ^ { B }", "type": "inline_equation" }, { "bbox": [ 358, 594, 506, 611 ], "score": 1.0, "content": "at any time step can be sampled with", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 606, 505, 621 ], "spans": [ { "bbox": [ 105, 606, 505, 621 ], "score": 1.0, "content": "closed form in the diffusion process Section 3.1, one can conveniently train the contrastive learning", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 617, 506, 632 ], "spans": [ { "bbox": [ 105, 617, 160, 632 ], "score": 1.0, "content": "together with", "type": "text" }, { "bbox": [ 160, 618, 190, 629 ], "score": 0.91, "content": "\\mathcal { L } _ { C r o s s }", "type": "inline_equation" }, { "bbox": [ 191, 617, 255, 632 ], "score": 1.0, "content": ". The purpose of", "type": "text" }, { "bbox": [ 256, 618, 265, 628 ], "score": 0.78, "content": "V", "type": "inline_equation" }, { "bbox": [ 265, 617, 506, 632 ], "score": 1.0, "content": "is to achieve the generation of any combination of modalities", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 629, 505, 642 ], "spans": [ { "bbox": [ 106, 629, 505, 642 ], "score": 1.0, "content": "(in polynomial) by training on a linear number of joint-generation tasks. For example, if we have", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 638, 507, 654 ], "spans": [ { "bbox": [ 104, 638, 278, 654 ], "score": 1.0, "content": "trained the joint generation of modalities", "type": "text" }, { "bbox": [ 279, 640, 301, 650 ], "score": 0.38, "content": "A , B", "type": "inline_equation" }, { "bbox": [ 301, 638, 323, 654 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 324, 640, 333, 650 ], "score": 0.69, "content": "B", "type": "inline_equation" }, { "bbox": [ 333, 638, 337, 654 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 338, 640, 347, 650 ], "score": 0.68, "content": "C", "type": "inline_equation" }, { "bbox": [ 347, 638, 468, 654 ], "score": 1.0, "content": "independently, then we have", "type": "text" }, { "bbox": [ 469, 639, 502, 652 ], "score": 0.92, "content": "V _ { A } ( z _ { t } ^ { A } )", "type": "inline_equation" }, { "bbox": [ 502, 638, 507, 654 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 43 }, { "bbox": [ 107, 649, 506, 664 ], "spans": [ { "bbox": [ 107, 650, 141, 662 ], "score": 0.92, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 141, 649, 160, 664 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 161, 650, 195, 663 ], "score": 0.93, "content": "V _ { C } ( z _ { t } ^ { C } )", "type": "inline_equation" }, { "bbox": [ 195, 649, 506, 664 ], "score": 1.0, "content": "aligned. Therefore, CoDi can seamlessly achieve joint generation of modalities", "type": "text" } ], "index": 44 }, { "bbox": [ 107, 661, 505, 675 ], "spans": [ { "bbox": [ 107, 663, 115, 672 ], "score": 0.77, "content": "A", "type": "inline_equation" }, { "bbox": [ 116, 661, 133, 675 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 134, 663, 142, 672 ], "score": 0.84, "content": "C", "type": "inline_equation" }, { "bbox": [ 143, 661, 505, 675 ], "score": 1.0, "content": "without any additional training. Moreover, such design automatically effortlessly enables", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 672, 505, 685 ], "spans": [ { "bbox": [ 104, 672, 229, 685 ], "score": 1.0, "content": "joint generation of modalities", "type": "text" }, { "bbox": [ 230, 674, 237, 682 ], "score": 0.66, "content": "A", "type": "inline_equation" }, { "bbox": [ 237, 672, 241, 685 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 242, 674, 251, 682 ], "score": 0.71, "content": "B", "type": "inline_equation" }, { "bbox": [ 251, 672, 272, 685 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 272, 674, 281, 682 ], "score": 0.86, "content": "C", "type": "inline_equation" }, { "bbox": [ 281, 672, 426, 685 ], "score": 1.0, "content": "concurrently. Specifically, UNet of", "type": "text" }, { "bbox": [ 427, 673, 436, 682 ], "score": 0.76, "content": "A", "type": "inline_equation" }, { "bbox": [ 436, 672, 505, 685 ], "score": 1.0, "content": "can cross-attend", "type": "text" } ], "index": 46 }, { "bbox": [ 104, 682, 506, 698 ], "spans": [ { "bbox": [ 104, 682, 205, 698 ], "score": 1.0, "content": "with the interpolation of", "type": "text" }, { "bbox": [ 205, 684, 239, 695 ], "score": 0.93, "content": "V _ { B } \\big ( z _ { t } ^ { B } \\big )", "type": "inline_equation" }, { "bbox": [ 240, 682, 261, 698 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 261, 684, 294, 695 ], "score": 0.92, "content": "V _ { C } ( z _ { t } ^ { C } )", "type": "inline_equation" }, { "bbox": [ 294, 682, 506, 698 ], "score": 1.0, "content": ", although CoDi has not been trained with such task.", "type": "text" } ], "index": 47 } ], "index": 42.5, "bbox_fs": [ 103, 576, 507, 698 ] }, { "type": "text", "bbox": [ 106, 699, 503, 722 ], "lines": [ { "bbox": [ 105, 699, 505, 713 ], "spans": [ { "bbox": [ 105, 699, 505, 713 ], "score": 1.0, "content": "As shown in Fig. 2(b)(3), we follow similar designs to the \"Bridging Alignment\" in training joint", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "generation: (1) We first train the cross-attention weights in the image and text diffusers, as well", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 446, 505, 457 ], "spans": [ { "bbox": [ 106, 446, 231, 457 ], "score": 1.0, "content": "as their environment encoders", "type": "text", "cross_page": true }, { "bbox": [ 231, 446, 240, 456 ], "score": 0.63, "content": "V", "type": "inline_equation", "cross_page": true }, { "bbox": [ 241, 446, 505, 457 ], "score": 1.0, "content": ", on text-image paired data. (2) We freeze the weights of the text", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 456, 506, 469 ], "spans": [ { "bbox": [ 105, 456, 506, 469 ], "score": 1.0, "content": "diffuser and train the environment encoder and cross-attention weights of the audio diffuser on", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 106, 468, 506, 480 ], "spans": [ { "bbox": [ 106, 468, 506, 480 ], "score": 1.0, "content": "text-audio paired data. (3) Finally we freeze the audio diffuser and its environment encoder, and", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 106, 479, 505, 490 ], "spans": [ { "bbox": [ 106, 479, 505, 490 ], "score": 1.0, "content": "train the joint generation of the video modality on audio-video paired data. As demonstrated in", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 105, 489, 507, 502 ], "spans": [ { "bbox": [ 105, 489, 420, 502 ], "score": 1.0, "content": "Section 5.3, although only trained on three paired joint generation tasks (i.e, Text", "type": "text", "cross_page": true }, { "bbox": [ 420, 491, 426, 499 ], "score": 0.34, "content": "^ +", "type": "inline_equation", "cross_page": true }, { "bbox": [ 427, 489, 507, 502 ], "score": 1.0, "content": "Audio, Text+Image,", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 106, 501, 505, 513 ], "spans": [ { "bbox": [ 106, 501, 505, 513 ], "score": 1.0, "content": "and Video+Audio), CoDi is capable of generating assorted combinations of modalities simultaneously", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 105, 511, 409, 524 ], "spans": [ { "bbox": [ 105, 511, 409, 524 ], "score": 1.0, "content": "that are unseen in training, e.g., joint image-text-audio generation in Fig. 5.", "type": "text", "cross_page": true } ], "index": 16 } ], "index": 48.5, "bbox_fs": [ 105, 699, 505, 723 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 107, 97, 503, 219 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 504, 92 ], "group_id": 0, "lines": [ { "bbox": [ 106, 69, 505, 82 ], "spans": [ { "bbox": [ 106, 69, 505, 82 ], "score": 1.0, "content": "Table 1: Training tasks (CT stands for “contrastive learning” to align prompt encoders) and datasets", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 80, 502, 93 ], "spans": [ { "bbox": [ 106, 80, 502, 93 ], "score": 1.0, "content": "with corresponding statistics. * denotes the number of accessible examples in the original datasets.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 107, 97, 503, 219 ], "group_id": 0, "lines": [ { "bbox": [ 107, 97, 503, 219 ], "spans": [ { "bbox": [ 107, 97, 503, 219 ], "score": 0.769, "html": "
CategoriesTasksDatasets# of samplesDomain
Image + TextImage-→Text,Text-→Image Text-→Image+TextLaion400M [42]400MOpen
Audio + TextText→Audio,Audio-→Text, Text-→Audio+Text,Audio-Text CTAudioSet [16] AudioCaps [24] Freesound 500K BBC Sound Effect900K* 46K 2.5M 30KYouTube YouTube Public audio samples Authentic natural sound
AudiovisualImage→Audio,Image→Video+AudioAudioSet SoundNet [3]900K* 1.0M*YouTube Flickr, natural sound
VideoText-→Video,Image→Video, Video-Text CTWebvid10M[4] HD-Villa-100M [54]10.7M 100MShort videos YouTube
", "type": "table", "image_path": "a1227e1b3cb165e5b75d57d2176ab2a10da01b03a47cc44ccf5819cbf35e61b5.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 107, 97, 503, 137.66666666666666 ], "spans": [], "index": 2 }, { "bbox": [ 107, 137.66666666666666, 503, 178.33333333333331 ], "spans": [], "index": 3 }, { "bbox": [ 107, 178.33333333333331, 503, 218.99999999999997 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "image", "bbox": [ 107, 231, 503, 391 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 231, 503, 391 ], "group_id": 0, "lines": [ { "bbox": [ 107, 231, 503, 391 ], "spans": [ { "bbox": [ 107, 231, 503, 391 ], "score": 0.631, "type": "image", "image_path": "493f24169a89b5450efae4d85c3805ebc13a132fcf5e28a3a488cbd564530264.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 107, 231, 503, 284.3333333333333 ], "spans": [], "index": 5 }, { "bbox": [ 107, 284.3333333333333, 503, 337.66666666666663 ], "spans": [], "index": 6 }, { "bbox": [ 107, 337.66666666666663, 503, 390.99999999999994 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 107, 401, 505, 424 ], "group_id": 0, "lines": [ { "bbox": [ 105, 400, 506, 416 ], "spans": [ { "bbox": [ 105, 400, 414, 416 ], "score": 1.0, "content": "Figure 3: Single-to-single modality generation. Clockwise from top left: text", "type": "text" }, { "bbox": [ 414, 403, 424, 412 ], "score": 0.77, "content": "", "type": "inline_equation" }, { "bbox": [ 424, 400, 477, 416 ], "score": 1.0, "content": "image, image", "type": "text" }, { "bbox": [ 477, 403, 487, 412 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 488, 400, 506, 416 ], "score": 1.0, "content": "text,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 411, 229, 426 ], "spans": [ { "bbox": [ 105, 411, 131, 426 ], "score": 1.0, "content": "image", "type": "text" }, { "bbox": [ 132, 414, 142, 423 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 142, 411, 190, 426 ], "score": 1.0, "content": "video, audio", "type": "text" }, { "bbox": [ 191, 414, 201, 423 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 201, 411, 229, 426 ], "score": 1.0, "content": "image.", "type": "text" } ], "index": 9 } ], "index": 8.5 } ], "index": 7.25 }, { "type": "text", "bbox": [ 107, 445, 505, 523 ], "lines": [ { "bbox": [ 106, 446, 505, 457 ], "spans": [ { "bbox": [ 106, 446, 231, 457 ], "score": 1.0, "content": "as their environment encoders", "type": "text" }, { "bbox": [ 231, 446, 240, 456 ], "score": 0.63, "content": "V", "type": "inline_equation" }, { "bbox": [ 241, 446, 505, 457 ], "score": 1.0, "content": ", on text-image paired data. (2) We freeze the weights of the text", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 456, 506, 469 ], "spans": [ { "bbox": [ 105, 456, 506, 469 ], "score": 1.0, "content": "diffuser and train the environment encoder and cross-attention weights of the audio diffuser on", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 468, 506, 480 ], "spans": [ { "bbox": [ 106, 468, 506, 480 ], "score": 1.0, "content": "text-audio paired data. (3) Finally we freeze the audio diffuser and its environment encoder, and", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 479, 505, 490 ], "spans": [ { "bbox": [ 106, 479, 505, 490 ], "score": 1.0, "content": "train the joint generation of the video modality on audio-video paired data. As demonstrated in", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 489, 507, 502 ], "spans": [ { "bbox": [ 105, 489, 420, 502 ], "score": 1.0, "content": "Section 5.3, although only trained on three paired joint generation tasks (i.e, Text", "type": "text" }, { "bbox": [ 420, 491, 426, 499 ], "score": 0.34, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 427, 489, 507, 502 ], "score": 1.0, "content": "Audio, Text+Image,", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 501, 505, 513 ], "spans": [ { "bbox": [ 106, 501, 505, 513 ], "score": 1.0, "content": "and Video+Audio), CoDi is capable of generating assorted combinations of modalities simultaneously", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 511, 409, 524 ], "spans": [ { "bbox": [ 105, 511, 409, 524 ], "score": 1.0, "content": "that are unseen in training, e.g., joint image-text-audio generation in Fig. 5.", "type": "text" } ], "index": 16 } ], "index": 13 }, { "type": "title", "bbox": [ 107, 539, 191, 552 ], "lines": [ { "bbox": [ 104, 537, 193, 555 ], "spans": [ { "bbox": [ 104, 537, 193, 555 ], "score": 1.0, "content": "4 Experiments", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "title", "bbox": [ 108, 564, 271, 576 ], "lines": [ { "bbox": [ 106, 564, 272, 578 ], "spans": [ { "bbox": [ 106, 564, 272, 578 ], "score": 1.0, "content": "4.1 Training Objectives and Datasets", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 585, 505, 695 ], "lines": [ { "bbox": [ 106, 585, 505, 597 ], "spans": [ { "bbox": [ 106, 585, 505, 597 ], "score": 1.0, "content": "We list training tasks of CoDi in Table 1, including single modality synthesis, joint multimodal", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 596, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 506, 609 ], "score": 1.0, "content": "generation, and contrastive learning to align prompt encoders. Table 1 provides an overview of the", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 605, 506, 621 ], "spans": [ { "bbox": [ 105, 605, 506, 621 ], "score": 1.0, "content": "datasets, tasks, number of samples, and domain. Datasets are from the following domains: image", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 616, 507, 632 ], "spans": [ { "bbox": [ 105, 619, 113, 628 ], "score": 0.65, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 113, 616, 271, 632 ], "score": 1.0, "content": "text (e.g. image with caption), audio", "type": "text" }, { "bbox": [ 271, 619, 280, 628 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 280, 616, 450, 632 ], "score": 1.0, "content": "text (e.g. audio with description), audio", "type": "text" }, { "bbox": [ 451, 619, 459, 628 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 460, 616, 507, 632 ], "score": 1.0, "content": "video (e.g.", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 628, 506, 641 ], "spans": [ { "bbox": [ 105, 628, 228, 641 ], "score": 1.0, "content": "video with sound), and video", "type": "text" }, { "bbox": [ 228, 630, 237, 639 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 237, 628, 506, 641 ], "score": 1.0, "content": "text (e.g. video with description). As one may have noticed, the", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 640, 505, 653 ], "spans": [ { "bbox": [ 106, 640, 505, 653 ], "score": 1.0, "content": "language modality appears in most datasets and domains. This echos the idea of using text as the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 650, 505, 662 ], "spans": [ { "bbox": [ 106, 650, 505, 662 ], "score": 1.0, "content": "bridge modality to be able to extrapolate and generate new unseen combinations such as audio", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 662, 505, 674 ], "spans": [ { "bbox": [ 106, 662, 505, 674 ], "score": 1.0, "content": "and image bridged by text, as mentioned in Section 3.2 and Section 3.4. Due to space limit, more", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 672, 505, 685 ], "spans": [ { "bbox": [ 105, 672, 505, 685 ], "score": 1.0, "content": "details on training datasets and can be found in Appendix C, model architecture details in Appendix", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 684, 311, 696 ], "spans": [ { "bbox": [ 106, 684, 311, 696 ], "score": 1.0, "content": "Appendix A.1, and training details in Appendix B.", "type": "text" } ], "index": 28 } ], "index": 23.5 }, { "type": "text", "bbox": [ 106, 700, 504, 722 ], "lines": [ { "bbox": [ 106, 699, 504, 712 ], "spans": [ { "bbox": [ 106, 699, 135, 712 ], "score": 1.0, "content": "Image", "type": "text" }, { "bbox": [ 135, 701, 144, 710 ], "score": 0.77, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 145, 699, 504, 712 ], "score": 1.0, "content": "Text. We use a recently developed large-scale image caption dataset, Laion400M [42].", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 354, 723 ], "score": 1.0, "content": "This image-text paired data allows us to train with tasks text", "type": "text" }, { "bbox": [ 354, 712, 365, 720 ], "score": 0.76, "content": "", "type": "inline_equation" }, { "bbox": [ 365, 711, 420, 723 ], "score": 1.0, "content": "image, image", "type": "text" }, { "bbox": [ 421, 712, 431, 720 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 431, 711, 505, 723 ], "score": 1.0, "content": "text, and the joint", "type": "text" } ], "index": 30 } ], "index": 29.5 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 308, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 107, 97, 503, 219 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 504, 92 ], "group_id": 0, "lines": [ { "bbox": [ 106, 69, 505, 82 ], "spans": [ { "bbox": [ 106, 69, 505, 82 ], "score": 1.0, "content": "Table 1: Training tasks (CT stands for “contrastive learning” to align prompt encoders) and datasets", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 80, 502, 93 ], "spans": [ { "bbox": [ 106, 80, 502, 93 ], "score": 1.0, "content": "with corresponding statistics. * denotes the number of accessible examples in the original datasets.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 107, 97, 503, 219 ], "group_id": 0, "lines": [ { "bbox": [ 107, 97, 503, 219 ], "spans": [ { "bbox": [ 107, 97, 503, 219 ], "score": 0.769, "html": "
CategoriesTasksDatasets# of samplesDomain
Image + TextImage-→Text,Text-→Image Text-→Image+TextLaion400M [42]400MOpen
Audio + TextText→Audio,Audio-→Text, Text-→Audio+Text,Audio-Text CTAudioSet [16] AudioCaps [24] Freesound 500K BBC Sound Effect900K* 46K 2.5M 30KYouTube YouTube Public audio samples Authentic natural sound
AudiovisualImage→Audio,Image→Video+AudioAudioSet SoundNet [3]900K* 1.0M*YouTube Flickr, natural sound
VideoText-→Video,Image→Video, Video-Text CTWebvid10M[4] HD-Villa-100M [54]10.7M 100MShort videos YouTube
", "type": "table", "image_path": "a1227e1b3cb165e5b75d57d2176ab2a10da01b03a47cc44ccf5819cbf35e61b5.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 107, 97, 503, 137.66666666666666 ], "spans": [], "index": 2 }, { "bbox": [ 107, 137.66666666666666, 503, 178.33333333333331 ], "spans": [], "index": 3 }, { "bbox": [ 107, 178.33333333333331, 503, 218.99999999999997 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "image", "bbox": [ 107, 231, 503, 391 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 231, 503, 391 ], "group_id": 0, "lines": [ { "bbox": [ 107, 231, 503, 391 ], "spans": [ { "bbox": [ 107, 231, 503, 391 ], "score": 0.631, "type": "image", "image_path": "493f24169a89b5450efae4d85c3805ebc13a132fcf5e28a3a488cbd564530264.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 107, 231, 503, 284.3333333333333 ], "spans": [], "index": 5 }, { "bbox": [ 107, 284.3333333333333, 503, 337.66666666666663 ], "spans": [], "index": 6 }, { "bbox": [ 107, 337.66666666666663, 503, 390.99999999999994 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 107, 401, 505, 424 ], "group_id": 0, "lines": [ { "bbox": [ 105, 400, 506, 416 ], "spans": [ { "bbox": [ 105, 400, 414, 416 ], "score": 1.0, "content": "Figure 3: Single-to-single modality generation. Clockwise from top left: text", "type": "text" }, { "bbox": [ 414, 403, 424, 412 ], "score": 0.77, "content": "", "type": "inline_equation" }, { "bbox": [ 424, 400, 477, 416 ], "score": 1.0, "content": "image, image", "type": "text" }, { "bbox": [ 477, 403, 487, 412 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 488, 400, 506, 416 ], "score": 1.0, "content": "text,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 411, 229, 426 ], "spans": [ { "bbox": [ 105, 411, 131, 426 ], "score": 1.0, "content": "image", "type": "text" }, { "bbox": [ 132, 414, 142, 423 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 142, 411, 190, 426 ], "score": 1.0, "content": "video, audio", "type": "text" }, { "bbox": [ 191, 414, 201, 423 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 201, 411, 229, 426 ], "score": 1.0, "content": "image.", "type": "text" } ], "index": 9 } ], "index": 8.5 } ], "index": 7.25 }, { "type": "text", "bbox": [ 107, 445, 505, 523 ], "lines": [], "index": 13, "bbox_fs": [ 105, 446, 507, 524 ], "lines_deleted": true }, { "type": "title", "bbox": [ 107, 539, 191, 552 ], "lines": [ { "bbox": [ 104, 537, 193, 555 ], "spans": [ { "bbox": [ 104, 537, 193, 555 ], "score": 1.0, "content": "4 Experiments", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "title", "bbox": [ 108, 564, 271, 576 ], "lines": [ { "bbox": [ 106, 564, 272, 578 ], "spans": [ { "bbox": [ 106, 564, 272, 578 ], "score": 1.0, "content": "4.1 Training Objectives and Datasets", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 585, 505, 695 ], "lines": [ { "bbox": [ 106, 585, 505, 597 ], "spans": [ { "bbox": [ 106, 585, 505, 597 ], "score": 1.0, "content": "We list training tasks of CoDi in Table 1, including single modality synthesis, joint multimodal", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 596, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 506, 609 ], "score": 1.0, "content": "generation, and contrastive learning to align prompt encoders. Table 1 provides an overview of the", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 605, 506, 621 ], "spans": [ { "bbox": [ 105, 605, 506, 621 ], "score": 1.0, "content": "datasets, tasks, number of samples, and domain. Datasets are from the following domains: image", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 616, 507, 632 ], "spans": [ { "bbox": [ 105, 619, 113, 628 ], "score": 0.65, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 113, 616, 271, 632 ], "score": 1.0, "content": "text (e.g. image with caption), audio", "type": "text" }, { "bbox": [ 271, 619, 280, 628 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 280, 616, 450, 632 ], "score": 1.0, "content": "text (e.g. audio with description), audio", "type": "text" }, { "bbox": [ 451, 619, 459, 628 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 460, 616, 507, 632 ], "score": 1.0, "content": "video (e.g.", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 628, 506, 641 ], "spans": [ { "bbox": [ 105, 628, 228, 641 ], "score": 1.0, "content": "video with sound), and video", "type": "text" }, { "bbox": [ 228, 630, 237, 639 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 237, 628, 506, 641 ], "score": 1.0, "content": "text (e.g. video with description). As one may have noticed, the", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 640, 505, 653 ], "spans": [ { "bbox": [ 106, 640, 505, 653 ], "score": 1.0, "content": "language modality appears in most datasets and domains. This echos the idea of using text as the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 650, 505, 662 ], "spans": [ { "bbox": [ 106, 650, 505, 662 ], "score": 1.0, "content": "bridge modality to be able to extrapolate and generate new unseen combinations such as audio", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 662, 505, 674 ], "spans": [ { "bbox": [ 106, 662, 505, 674 ], "score": 1.0, "content": "and image bridged by text, as mentioned in Section 3.2 and Section 3.4. Due to space limit, more", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 672, 505, 685 ], "spans": [ { "bbox": [ 105, 672, 505, 685 ], "score": 1.0, "content": "details on training datasets and can be found in Appendix C, model architecture details in Appendix", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 684, 311, 696 ], "spans": [ { "bbox": [ 106, 684, 311, 696 ], "score": 1.0, "content": "Appendix A.1, and training details in Appendix B.", "type": "text" } ], "index": 28 } ], "index": 23.5, "bbox_fs": [ 105, 585, 507, 696 ] }, { "type": "text", "bbox": [ 106, 700, 504, 722 ], "lines": [ { "bbox": [ 106, 699, 504, 712 ], "spans": [ { "bbox": [ 106, 699, 135, 712 ], "score": 1.0, "content": "Image", "type": "text" }, { "bbox": [ 135, 701, 144, 710 ], "score": 0.77, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 145, 699, 504, 712 ], "score": 1.0, "content": "Text. We use a recently developed large-scale image caption dataset, Laion400M [42].", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 354, 723 ], "score": 1.0, "content": "This image-text paired data allows us to train with tasks text", "type": "text" }, { "bbox": [ 354, 712, 365, 720 ], "score": 0.76, "content": "", "type": "inline_equation" }, { "bbox": [ 365, 711, 420, 723 ], "score": 1.0, "content": "image, image", "type": "text" }, { "bbox": [ 421, 712, 431, 720 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 431, 711, 505, 723 ], "score": 1.0, "content": "text, and the joint", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 465, 506, 478 ], "spans": [ { "bbox": [ 105, 465, 445, 478 ], "score": 1.0, "content": "generation of image and text. For the joint generation task, we propose to train with text", "type": "text", "cross_page": true }, { "bbox": [ 446, 467, 455, 475 ], "score": 0.66, "content": "", "type": "inline_equation", "cross_page": true }, { "bbox": [ 456, 465, 506, 478 ], "score": 1.0, "content": "image+text,", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 105, 476, 506, 489 ], "spans": [ { "bbox": [ 105, 476, 506, 489 ], "score": 1.0, "content": "where the prompt text is the truncated image caption, and the output text is the original caption. Since", "type": "text", "cross_page": true } ], "index": 30 }, { "bbox": [ 106, 488, 505, 499 ], "spans": [ { "bbox": [ 106, 488, 505, 499 ], "score": 1.0, "content": "the condition information is incomplete, the text and image diffuser will need to learn to attend with", "type": "text", "cross_page": true } ], "index": 31 }, { "bbox": [ 105, 498, 299, 511 ], "spans": [ { "bbox": [ 105, 498, 299, 511 ], "score": 1.0, "content": "each other through the joint generation process.", "type": "text", "cross_page": true } ], "index": 32 } ], "index": 29.5, "bbox_fs": [ 106, 699, 505, 723 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 109, 109, 227, 204 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 228, 102 ], "group_id": 0, "lines": [ { "bbox": [ 107, 69, 228, 81 ], "spans": [ { "bbox": [ 107, 69, 228, 81 ], "score": 1.0, "content": "Table 2: COCO-caption [32]", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 79, 228, 92 ], "spans": [ { "bbox": [ 106, 79, 228, 92 ], "score": 1.0, "content": "FID scores for text-to-image", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 92, 155, 103 ], "spans": [ { "bbox": [ 106, 92, 155, 103 ], "score": 1.0, "content": "generation.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 109, 109, 227, 204 ], "group_id": 0, "lines": [ { "bbox": [ 109, 109, 227, 204 ], "spans": [ { "bbox": [ 109, 109, 227, 204 ], "score": 0.962, "html": "
MethodFID↓
CogView [10]27.10
GLIDE [36]12.24
Make-a-Scene [15]11.84
LDM [41]12.63
Stable Diffusion-1.411.21
Stable Diffusion-1.511.12
Versatile Diffusion [53]11.10
CoDi (Ours)11.26
", "type": "table", "image_path": "f0713a5e3868134fdb21d162b330910771c9fe92eb4e6a85f24933ece1e1781b.jpg" } ] } ], "index": 6.5, "virtual_lines": [ { "bbox": [ 109, 109, 227, 156.5 ], "spans": [], "index": 5 }, { "bbox": [ 109, 156.5, 227, 204.0 ], "spans": [], "index": 8 } ] } ], "index": 3.75 }, { "type": "table", "bbox": [ 239, 121, 373, 183 ], "blocks": [ { "type": "table_caption", "bbox": [ 237, 93, 504, 116 ], "group_id": 2, "lines": [ { "bbox": [ 237, 92, 504, 105 ], "spans": [ { "bbox": [ 237, 92, 504, 105 ], "score": 1.0, "content": "Table 3: MSR-VTT text-to-video Table 4: UCF-101 text-to-video", "type": "text" } ], "index": 3 }, { "bbox": [ 237, 104, 475, 117 ], "spans": [ { "bbox": [ 237, 104, 338, 117 ], "score": 1.0, "content": "generation performance.", "type": "text" }, { "bbox": [ 374, 105, 475, 117 ], "score": 1.0, "content": "generation performance.", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "table_body", "bbox": [ 239, 121, 373, 183 ], "group_id": 2, "lines": [ { "bbox": [ 239, 121, 373, 183 ], "spans": [ { "bbox": [ 239, 121, 373, 183 ], "score": 0.966, "html": "
MethodZero-ShotCLIPSIM ↑
GODIVA [50]No0.2402
NUWA [51]No0.2439
CogVideo [22]Yes0.2631
Make-A-Video [44]Yes0.3049
Video LDM[5]Yes0.2929
CoDi(Ours)Yes0.2890
", "type": "table", "image_path": "092e9968c4819972025aaf57bfda705e16f8eed1cfa742fa9dda77afc438a0d5.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 239, 121, 373, 152.0 ], "spans": [], "index": 6 }, { "bbox": [ 239, 152.0, 373, 183.0 ], "spans": [], "index": 9 } ] } ], "index": 5.5 }, { "type": "table", "bbox": [ 378, 122, 503, 182 ], "blocks": [ { "type": "table_body", "bbox": [ 378, 122, 503, 182 ], "group_id": 6, "lines": [ { "bbox": [ 378, 122, 503, 182 ], "spans": [ { "bbox": [ 378, 122, 503, 182 ], "score": 0.933, "html": "
MethodIS(1)FVD (↑)
Cog Video (Chinese)23.55751.34
CogVideo (English)25.27701.59
Make-A-Video33.00367.23
Video LDM33.45550.61
CoDi(Ours)32.88596.34
", "type": "table", "image_path": "89d31c9fff04f8286fa349bee00035aa85492d934c65c9e6e455e2a373e24edf.jpg" } ] } ], "index": 8.5, "virtual_lines": [ { "bbox": [ 378, 122, 503, 152.0 ], "spans": [], "index": 7 }, { "bbox": [ 378, 152.0, 503, 182.0 ], "spans": [], "index": 10 } ] } ], "index": 8.5 }, { "type": "table", "bbox": [ 126, 251, 484, 318 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 213, 505, 246 ], "group_id": 1, "lines": [ { "bbox": [ 105, 212, 505, 225 ], "spans": [ { "bbox": [ 105, 212, 505, 225 ], "score": 1.0, "content": "Table 5: The comparison between our audio diffuser and baseline TTA generation models. Evaluation", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 223, 507, 237 ], "spans": [ { "bbox": [ 105, 223, 507, 237 ], "score": 1.0, "content": "is conducted on AudioCaps test set. AS, AC, FSD, BBC, and SDN stand for AudioSet, AudioCaps,", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 235, 291, 246 ], "spans": [ { "bbox": [ 106, 235, 291, 246 ], "score": 1.0, "content": "Freesound, BBC Sound Effect, and Soundnet.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 126, 251, 484, 318 ], "group_id": 1, "lines": [ { "bbox": [ 126, 251, 484, 318 ], "spans": [ { "bbox": [ 126, 251, 484, 318 ], "score": 0.978, "html": "
ModelDatasetsFD↓IS个KL←FAD↓OVL ↑REL个
Ground truth-----83.6180.11
DiffSoundAS+AC47.684.012.527.7545.0043.83
AudioGenAS +AC+8others=-2.093.13-=
AudioLDM-L-FullAS+AC+FSD+BBC23.318.131.591.9665.9165.97
CoDi(Ours)AS+AC+FSD+BBC+SDN22.908.771.401.8066.8767.60
", "type": "table", "image_path": "6cdd7e474acf79309e07b53bb583c906924dd31228adbed791c23115f83c02ab.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 126, 251, 484, 273.3333333333333 ], "spans": [], "index": 14 }, { "bbox": [ 126, 273.3333333333333, 484, 295.66666666666663 ], "spans": [], "index": 15 }, { "bbox": [ 126, 295.66666666666663, 484, 317.99999999999994 ], "spans": [], "index": 16 } ] } ], "index": 13.5 }, { "type": "table", "bbox": [ 108, 356, 231, 439 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 325, 232, 348 ], "group_id": 3, "lines": [ { "bbox": [ 105, 324, 233, 338 ], "spans": [ { "bbox": [ 105, 324, 233, 338 ], "score": 1.0, "content": "Table 6: COCO image caption-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 337, 200, 349 ], "spans": [ { "bbox": [ 106, 337, 200, 349 ], "score": 1.0, "content": "ing scores comparison.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "table_body", "bbox": [ 108, 356, 231, 439 ], "group_id": 3, "lines": [ { "bbox": [ 108, 356, 231, 439 ], "spans": [ { "bbox": [ 108, 356, 231, 439 ], "score": 0.966, "html": "
ModelB@4METEORCIDEr
Autoregressive Model
Oscar [31]36.5830.4124.12
ClipCap [35]32.1527.1108.35
OFA [49]44.932.5154.9
BLIP2 [30]43.7-145.8
Diffusion Model
DDCap [59]35.028.2117.8
SCD-Net [34]39.429.2131.6
CoDi (Ours)40.231.0149.9
", "type": "table", "image_path": "3b08dd3a07ff80c69c148aa0491f160ff27946a1af6b4a57632f75d5ff7af62b.jpg" } ] } ], "index": 23.5, "virtual_lines": [ { "bbox": [ 108, 356, 231, 397.5 ], "spans": [], "index": 23 }, { "bbox": [ 108, 397.5, 231, 439.0 ], "spans": [], "index": 24 } ] } ], "index": 20.5 }, { "type": "table", "bbox": [ 239, 374, 366, 423 ], "blocks": [ { "type": "table_caption", "bbox": [ 237, 344, 367, 366 ], "group_id": 4, "lines": [ { "bbox": [ 236, 342, 368, 357 ], "spans": [ { "bbox": [ 236, 342, 368, 357 ], "score": 1.0, "content": "Table 7: AudioCaps audio cap-", "type": "text" } ], "index": 19 }, { "bbox": [ 237, 354, 347, 367 ], "spans": [ { "bbox": [ 237, 354, 347, 367 ], "score": 1.0, "content": "tioning scores comparison.", "type": "text" } ], "index": 21 } ], "index": 20.0 }, { "type": "table_body", "bbox": [ 239, 374, 366, 423 ], "group_id": 4, "lines": [ { "bbox": [ 239, 374, 366, 423 ], "spans": [ { "bbox": [ 239, 374, 366, 423 ], "score": 0.971, "html": "
ModelSPIDErCIDErSPICE
AudioCaps [24]0.3690.5930.144
BART-Finetune [17]0.4650.7530.176
VALOR[7]0.741
AL-MixGen [25]0.4660.7550.177
CoDi (Ours)0.4800.7890.182
", "type": "table", "image_path": "d48a4e4d621172d01f3a4f8bc34da6f60c2863153ae2635e211633ead5c366e1.jpg" } ] } ], "index": 26.0, "virtual_lines": [ { "bbox": [ 239, 374, 366, 398.5 ], "spans": [], "index": 25 }, { "bbox": [ 239, 398.5, 366, 423.0 ], "spans": [], "index": 27 } ] } ], "index": 23.0 }, { "type": "table", "bbox": [ 374, 373, 495, 423 ], "blocks": [ { "type": "table_caption", "bbox": [ 372, 343, 495, 365 ], "group_id": 5, "lines": [ { "bbox": [ 371, 341, 496, 356 ], "spans": [ { "bbox": [ 371, 341, 496, 356 ], "score": 1.0, "content": "Table 8: MSRVTT video cap-", "type": "text" } ], "index": 20 }, { "bbox": [ 372, 354, 482, 366 ], "spans": [ { "bbox": [ 372, 354, 482, 366 ], "score": 1.0, "content": "tioning scores comparison.", "type": "text" } ], "index": 22 } ], "index": 21.0 }, { "type": "table_body", "bbox": [ 374, 373, 495, 423 ], "group_id": 5, "lines": [ { "bbox": [ 374, 373, 495, 423 ], "spans": [ { "bbox": [ 374, 373, 495, 423 ], "score": 0.968, "html": "
ModelB@4METEORCIDEr
ORG-TRL[58]43.628.850.9
MV-GPT[43]48.938.760.0
GIT[48]54.833.175.9
mPLUG-2 [52]57.834.980.3
CoDi(Ours)52.132.574.4
", "type": "table", "image_path": "555831ba6bc0c59839fc917171599939764c30018cb4c379929213d7ca95eaf8.jpg" } ] } ], "index": 27.0, "virtual_lines": [ { "bbox": [ 374, 373, 495, 398.0 ], "spans": [], "index": 26 }, { "bbox": [ 374, 398.0, 495, 423.0 ], "spans": [], "index": 28 } ] } ], "index": 24.0 }, { "type": "text", "bbox": [ 107, 465, 505, 509 ], "lines": [ { "bbox": [ 105, 465, 506, 478 ], "spans": [ { "bbox": [ 105, 465, 445, 478 ], "score": 1.0, "content": "generation of image and text. For the joint generation task, we propose to train with text", "type": "text" }, { "bbox": [ 446, 467, 455, 475 ], "score": 0.66, "content": "", "type": "inline_equation" }, { "bbox": [ 456, 465, 506, 478 ], "score": 1.0, "content": "image+text,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 476, 506, 489 ], "spans": [ { "bbox": [ 105, 476, 506, 489 ], "score": 1.0, "content": "where the prompt text is the truncated image caption, and the output text is the original caption. Since", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 488, 505, 499 ], "spans": [ { "bbox": [ 106, 488, 505, 499 ], "score": 1.0, "content": "the condition information is incomplete, the text and image diffuser will need to learn to attend with", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 498, 299, 511 ], "spans": [ { "bbox": [ 105, 498, 299, 511 ], "score": 1.0, "content": "each other through the joint generation process.", "type": "text" } ], "index": 32 } ], "index": 30.5 }, { "type": "text", "bbox": [ 106, 514, 505, 591 ], "lines": [ { "bbox": [ 105, 513, 506, 527 ], "spans": [ { "bbox": [ 105, 513, 133, 527 ], "score": 1.0, "content": "Audio", "type": "text" }, { "bbox": [ 133, 515, 142, 524 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 142, 513, 506, 527 ], "score": 1.0, "content": "Text. We curated a new dataset, Freesound 500K, by crawling 500K audio samples together", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 526, 505, 537 ], "spans": [ { "bbox": [ 106, 526, 505, 537 ], "score": 1.0, "content": "with tags and descriptions from the Freesound website. We also use AudioSet [42] with 2 million", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 537, 506, 547 ], "spans": [ { "bbox": [ 106, 537, 506, 547 ], "score": 1.0, "content": "human-labeled 10-second sound clips from YouTube videos and AudioCaps [24] with 46K audio-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 547, 506, 560 ], "spans": [ { "bbox": [ 105, 547, 506, 560 ], "score": 1.0, "content": "text pairs derived from the AudioSet dataset. Audio samples are clipped into 10-second segments", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 558, 507, 571 ], "spans": [ { "bbox": [ 105, 558, 271, 571 ], "score": 1.0, "content": "for training purposes. The paired audio", "type": "text" }, { "bbox": [ 271, 559, 280, 568 ], "score": 0.79, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 280, 558, 412, 571 ], "score": 1.0, "content": "text data enables us to train text", "type": "text" }, { "bbox": [ 413, 559, 423, 568 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 423, 558, 476, 571 ], "score": 1.0, "content": "audio, audio", "type": "text" }, { "bbox": [ 477, 559, 486, 568 ], "score": 0.74, "content": "", "type": "inline_equation" }, { "bbox": [ 486, 558, 507, 571 ], "score": 1.0, "content": "text,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 568, 506, 581 ], "spans": [ { "bbox": [ 105, 568, 123, 581 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 570, 133, 578 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 134, 568, 158, 581 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 159, 570, 167, 579 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 168, 568, 456, 581 ], "score": 1.0, "content": "text generation, and audio-text contrastive learning. Similar to image", "type": "text" }, { "bbox": [ 456, 570, 465, 579 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 465, 568, 506, 581 ], "score": 1.0, "content": "text joint", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 580, 506, 592 ], "spans": [ { "bbox": [ 105, 580, 179, 592 ], "score": 1.0, "content": "generation, in text", "type": "text" }, { "bbox": [ 179, 581, 189, 590 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 189, 580, 212, 592 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 212, 581, 220, 590 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 221, 580, 506, 592 ], "score": 1.0, "content": "text, text prompt is the truncated text, and the output is the original text.", "type": "text" } ], "index": 39 } ], "index": 36 }, { "type": "text", "bbox": [ 107, 596, 505, 651 ], "lines": [ { "bbox": [ 106, 596, 505, 608 ], "spans": [ { "bbox": [ 106, 596, 505, 608 ], "score": 1.0, "content": "Video. We use the following diverse and high-quality video datasets to train video generation and", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 606, 506, 620 ], "spans": [ { "bbox": [ 105, 606, 506, 620 ], "score": 1.0, "content": "video prompt encoder. WebVid [4], a large-scale dataset of web videos together with descriptions;", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 618, 505, 630 ], "spans": [ { "bbox": [ 106, 618, 471, 630 ], "score": 1.0, "content": "HD-Villa-100M [54] with high resolution YouTube videos of at least 720P. We perform text", "type": "text" }, { "bbox": [ 471, 619, 481, 627 ], "score": 0.73, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 618, 505, 630 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 628, 505, 642 ], "spans": [ { "bbox": [ 106, 628, 470, 642 ], "score": 1.0, "content": "and video-text contrastive learning task with WebVid. We use HD-Villa-100M for image", "type": "text" }, { "bbox": [ 470, 631, 481, 639 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 628, 505, 642 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 639, 324, 654 ], "spans": [ { "bbox": [ 105, 639, 324, 654 ], "score": 1.0, "content": "generation where the middle frame is the input image.", "type": "text" } ], "index": 44 } ], "index": 42 }, { "type": "text", "bbox": [ 107, 656, 505, 722 ], "lines": [ { "bbox": [ 105, 654, 505, 669 ], "spans": [ { "bbox": [ 105, 654, 505, 669 ], "score": 1.0, "content": "Audiovisual. Web videos are a natural aligned audio-video data resource. However, many existing", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 667, 506, 678 ], "spans": [ { "bbox": [ 106, 667, 506, 678 ], "score": 1.0, "content": "datasets, e.g., ACAV100M [28], feature heavily on videos of human speech rather than natural sounds.", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 677, 506, 690 ], "spans": [ { "bbox": [ 105, 677, 506, 690 ], "score": 1.0, "content": "Therefore, we leverage sound-oriented datasets AudioSet and SoundNet [3] for joint audio-video", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 689, 506, 702 ], "spans": [ { "bbox": [ 105, 689, 199, 702 ], "score": 1.0, "content": "generation. For image", "type": "text" }, { "bbox": [ 200, 690, 210, 699 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 210, 689, 234, 702 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 235, 690, 244, 699 ], "score": 0.75, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 244, 689, 506, 702 ], "score": 1.0, "content": "video, we use the middle frame of the target video as the input", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 699, 506, 713 ], "spans": [ { "bbox": [ 105, 699, 506, 713 ], "score": 1.0, "content": "prompt image. We also use the middle frame as the prompt input to train the model to generate the", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 711, 214, 723 ], "spans": [ { "bbox": [ 106, 711, 177, 723 ], "score": 1.0, "content": "audio, i.e., image", "type": "text" }, { "bbox": [ 177, 712, 187, 721 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 187, 711, 214, 723 ], "score": 1.0, "content": "audio.", "type": "text" } ], "index": 50 } ], "index": 47.5 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 308, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 109, 109, 227, 204 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 228, 102 ], "group_id": 0, "lines": [ { "bbox": [ 107, 69, 228, 81 ], "spans": [ { "bbox": [ 107, 69, 228, 81 ], "score": 1.0, "content": "Table 2: COCO-caption [32]", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 79, 228, 92 ], "spans": [ { "bbox": [ 106, 79, 228, 92 ], "score": 1.0, "content": "FID scores for text-to-image", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 92, 155, 103 ], "spans": [ { "bbox": [ 106, 92, 155, 103 ], "score": 1.0, "content": "generation.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 109, 109, 227, 204 ], "group_id": 0, "lines": [ { "bbox": [ 109, 109, 227, 204 ], "spans": [ { "bbox": [ 109, 109, 227, 204 ], "score": 0.962, "html": "
MethodFID↓
CogView [10]27.10
GLIDE [36]12.24
Make-a-Scene [15]11.84
LDM [41]12.63
Stable Diffusion-1.411.21
Stable Diffusion-1.511.12
Versatile Diffusion [53]11.10
CoDi (Ours)11.26
", "type": "table", "image_path": "f0713a5e3868134fdb21d162b330910771c9fe92eb4e6a85f24933ece1e1781b.jpg" } ] } ], "index": 6.5, "virtual_lines": [ { "bbox": [ 109, 109, 227, 156.5 ], "spans": [], "index": 5 }, { "bbox": [ 109, 156.5, 227, 204.0 ], "spans": [], "index": 8 } ] } ], "index": 3.75 }, { "type": "table", "bbox": [ 239, 121, 373, 183 ], "blocks": [ { "type": "table_caption", "bbox": [ 237, 93, 504, 116 ], "group_id": 2, "lines": [ { "bbox": [ 237, 92, 504, 105 ], "spans": [ { "bbox": [ 237, 92, 504, 105 ], "score": 1.0, "content": "Table 3: MSR-VTT text-to-video Table 4: UCF-101 text-to-video", "type": "text" } ], "index": 3 }, { "bbox": [ 237, 104, 475, 117 ], "spans": [ { "bbox": [ 237, 104, 338, 117 ], "score": 1.0, "content": "generation performance.", "type": "text" }, { "bbox": [ 374, 105, 475, 117 ], "score": 1.0, "content": "generation performance.", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "table_body", "bbox": [ 239, 121, 373, 183 ], "group_id": 2, "lines": [ { "bbox": [ 239, 121, 373, 183 ], "spans": [ { "bbox": [ 239, 121, 373, 183 ], "score": 0.966, "html": "
MethodZero-ShotCLIPSIM ↑
GODIVA [50]No0.2402
NUWA [51]No0.2439
CogVideo [22]Yes0.2631
Make-A-Video [44]Yes0.3049
Video LDM[5]Yes0.2929
CoDi(Ours)Yes0.2890
", "type": "table", "image_path": "092e9968c4819972025aaf57bfda705e16f8eed1cfa742fa9dda77afc438a0d5.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 239, 121, 373, 152.0 ], "spans": [], "index": 6 }, { "bbox": [ 239, 152.0, 373, 183.0 ], "spans": [], "index": 9 } ] } ], "index": 5.5 }, { "type": "table", "bbox": [ 378, 122, 503, 182 ], "blocks": [ { "type": "table_body", "bbox": [ 378, 122, 503, 182 ], "group_id": 6, "lines": [ { "bbox": [ 378, 122, 503, 182 ], "spans": [ { "bbox": [ 378, 122, 503, 182 ], "score": 0.933, "html": "
MethodIS(1)FVD (↑)
Cog Video (Chinese)23.55751.34
CogVideo (English)25.27701.59
Make-A-Video33.00367.23
Video LDM33.45550.61
CoDi(Ours)32.88596.34
", "type": "table", "image_path": "89d31c9fff04f8286fa349bee00035aa85492d934c65c9e6e455e2a373e24edf.jpg" } ] } ], "index": 8.5, "virtual_lines": [ { "bbox": [ 378, 122, 503, 152.0 ], "spans": [], "index": 7 }, { "bbox": [ 378, 152.0, 503, 182.0 ], "spans": [], "index": 10 } ] } ], "index": 8.5 }, { "type": "table", "bbox": [ 126, 251, 484, 318 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 213, 505, 246 ], "group_id": 1, "lines": [ { "bbox": [ 105, 212, 505, 225 ], "spans": [ { "bbox": [ 105, 212, 505, 225 ], "score": 1.0, "content": "Table 5: The comparison between our audio diffuser and baseline TTA generation models. Evaluation", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 223, 507, 237 ], "spans": [ { "bbox": [ 105, 223, 507, 237 ], "score": 1.0, "content": "is conducted on AudioCaps test set. AS, AC, FSD, BBC, and SDN stand for AudioSet, AudioCaps,", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 235, 291, 246 ], "spans": [ { "bbox": [ 106, 235, 291, 246 ], "score": 1.0, "content": "Freesound, BBC Sound Effect, and Soundnet.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 126, 251, 484, 318 ], "group_id": 1, "lines": [ { "bbox": [ 126, 251, 484, 318 ], "spans": [ { "bbox": [ 126, 251, 484, 318 ], "score": 0.978, "html": "
ModelDatasetsFD↓IS个KL←FAD↓OVL ↑REL个
Ground truth-----83.6180.11
DiffSoundAS+AC47.684.012.527.7545.0043.83
AudioGenAS +AC+8others=-2.093.13-=
AudioLDM-L-FullAS+AC+FSD+BBC23.318.131.591.9665.9165.97
CoDi(Ours)AS+AC+FSD+BBC+SDN22.908.771.401.8066.8767.60
", "type": "table", "image_path": "6cdd7e474acf79309e07b53bb583c906924dd31228adbed791c23115f83c02ab.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 126, 251, 484, 273.3333333333333 ], "spans": [], "index": 14 }, { "bbox": [ 126, 273.3333333333333, 484, 295.66666666666663 ], "spans": [], "index": 15 }, { "bbox": [ 126, 295.66666666666663, 484, 317.99999999999994 ], "spans": [], "index": 16 } ] } ], "index": 13.5 }, { "type": "table", "bbox": [ 108, 356, 231, 439 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 325, 232, 348 ], "group_id": 3, "lines": [ { "bbox": [ 105, 324, 233, 338 ], "spans": [ { "bbox": [ 105, 324, 233, 338 ], "score": 1.0, "content": "Table 6: COCO image caption-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 337, 200, 349 ], "spans": [ { "bbox": [ 106, 337, 200, 349 ], "score": 1.0, "content": "ing scores comparison.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "table_body", "bbox": [ 108, 356, 231, 439 ], "group_id": 3, "lines": [ { "bbox": [ 108, 356, 231, 439 ], "spans": [ { "bbox": [ 108, 356, 231, 439 ], "score": 0.966, "html": "
ModelB@4METEORCIDEr
Autoregressive Model
Oscar [31]36.5830.4124.12
ClipCap [35]32.1527.1108.35
OFA [49]44.932.5154.9
BLIP2 [30]43.7-145.8
Diffusion Model
DDCap [59]35.028.2117.8
SCD-Net [34]39.429.2131.6
CoDi (Ours)40.231.0149.9
", "type": "table", "image_path": "3b08dd3a07ff80c69c148aa0491f160ff27946a1af6b4a57632f75d5ff7af62b.jpg" } ] } ], "index": 23.5, "virtual_lines": [ { "bbox": [ 108, 356, 231, 397.5 ], "spans": [], "index": 23 }, { "bbox": [ 108, 397.5, 231, 439.0 ], "spans": [], "index": 24 } ] } ], "index": 20.5 }, { "type": "table", "bbox": [ 239, 374, 366, 423 ], "blocks": [ { "type": "table_caption", "bbox": [ 237, 344, 367, 366 ], "group_id": 4, "lines": [ { "bbox": [ 236, 342, 368, 357 ], "spans": [ { "bbox": [ 236, 342, 368, 357 ], "score": 1.0, "content": "Table 7: AudioCaps audio cap-", "type": "text" } ], "index": 19 }, { "bbox": [ 237, 354, 347, 367 ], "spans": [ { "bbox": [ 237, 354, 347, 367 ], "score": 1.0, "content": "tioning scores comparison.", "type": "text" } ], "index": 21 } ], "index": 20.0 }, { "type": "table_body", "bbox": [ 239, 374, 366, 423 ], "group_id": 4, "lines": [ { "bbox": [ 239, 374, 366, 423 ], "spans": [ { "bbox": [ 239, 374, 366, 423 ], "score": 0.971, "html": "
ModelSPIDErCIDErSPICE
AudioCaps [24]0.3690.5930.144
BART-Finetune [17]0.4650.7530.176
VALOR[7]0.741
AL-MixGen [25]0.4660.7550.177
CoDi (Ours)0.4800.7890.182
", "type": "table", "image_path": "d48a4e4d621172d01f3a4f8bc34da6f60c2863153ae2635e211633ead5c366e1.jpg" } ] } ], "index": 26.0, "virtual_lines": [ { "bbox": [ 239, 374, 366, 398.5 ], "spans": [], "index": 25 }, { "bbox": [ 239, 398.5, 366, 423.0 ], "spans": [], "index": 27 } ] } ], "index": 23.0 }, { "type": "table", "bbox": [ 374, 373, 495, 423 ], "blocks": [ { "type": "table_caption", "bbox": [ 372, 343, 495, 365 ], "group_id": 5, "lines": [ { "bbox": [ 371, 341, 496, 356 ], "spans": [ { "bbox": [ 371, 341, 496, 356 ], "score": 1.0, "content": "Table 8: MSRVTT video cap-", "type": "text" } ], "index": 20 }, { "bbox": [ 372, 354, 482, 366 ], "spans": [ { "bbox": [ 372, 354, 482, 366 ], "score": 1.0, "content": "tioning scores comparison.", "type": "text" } ], "index": 22 } ], "index": 21.0 }, { "type": "table_body", "bbox": [ 374, 373, 495, 423 ], "group_id": 5, "lines": [ { "bbox": [ 374, 373, 495, 423 ], "spans": [ { "bbox": [ 374, 373, 495, 423 ], "score": 0.968, "html": "
ModelB@4METEORCIDEr
ORG-TRL[58]43.628.850.9
MV-GPT[43]48.938.760.0
GIT[48]54.833.175.9
mPLUG-2 [52]57.834.980.3
CoDi(Ours)52.132.574.4
", "type": "table", "image_path": "555831ba6bc0c59839fc917171599939764c30018cb4c379929213d7ca95eaf8.jpg" } ] } ], "index": 27.0, "virtual_lines": [ { "bbox": [ 374, 373, 495, 398.0 ], "spans": [], "index": 26 }, { "bbox": [ 374, 398.0, 495, 423.0 ], "spans": [], "index": 28 } ] } ], "index": 24.0 }, { "type": "text", "bbox": [ 107, 465, 505, 509 ], "lines": [], "index": 30.5, "bbox_fs": [ 105, 465, 506, 511 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 514, 505, 591 ], "lines": [ { "bbox": [ 105, 513, 506, 527 ], "spans": [ { "bbox": [ 105, 513, 133, 527 ], "score": 1.0, "content": "Audio", "type": "text" }, { "bbox": [ 133, 515, 142, 524 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 142, 513, 506, 527 ], "score": 1.0, "content": "Text. We curated a new dataset, Freesound 500K, by crawling 500K audio samples together", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 526, 505, 537 ], "spans": [ { "bbox": [ 106, 526, 505, 537 ], "score": 1.0, "content": "with tags and descriptions from the Freesound website. We also use AudioSet [42] with 2 million", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 537, 506, 547 ], "spans": [ { "bbox": [ 106, 537, 506, 547 ], "score": 1.0, "content": "human-labeled 10-second sound clips from YouTube videos and AudioCaps [24] with 46K audio-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 547, 506, 560 ], "spans": [ { "bbox": [ 105, 547, 506, 560 ], "score": 1.0, "content": "text pairs derived from the AudioSet dataset. Audio samples are clipped into 10-second segments", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 558, 507, 571 ], "spans": [ { "bbox": [ 105, 558, 271, 571 ], "score": 1.0, "content": "for training purposes. The paired audio", "type": "text" }, { "bbox": [ 271, 559, 280, 568 ], "score": 0.79, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 280, 558, 412, 571 ], "score": 1.0, "content": "text data enables us to train text", "type": "text" }, { "bbox": [ 413, 559, 423, 568 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 423, 558, 476, 571 ], "score": 1.0, "content": "audio, audio", "type": "text" }, { "bbox": [ 477, 559, 486, 568 ], "score": 0.74, "content": "", "type": "inline_equation" }, { "bbox": [ 486, 558, 507, 571 ], "score": 1.0, "content": "text,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 568, 506, 581 ], "spans": [ { "bbox": [ 105, 568, 123, 581 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 570, 133, 578 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 134, 568, 158, 581 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 159, 570, 167, 579 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 168, 568, 456, 581 ], "score": 1.0, "content": "text generation, and audio-text contrastive learning. Similar to image", "type": "text" }, { "bbox": [ 456, 570, 465, 579 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 465, 568, 506, 581 ], "score": 1.0, "content": "text joint", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 580, 506, 592 ], "spans": [ { "bbox": [ 105, 580, 179, 592 ], "score": 1.0, "content": "generation, in text", "type": "text" }, { "bbox": [ 179, 581, 189, 590 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 189, 580, 212, 592 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 212, 581, 220, 590 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 221, 580, 506, 592 ], "score": 1.0, "content": "text, text prompt is the truncated text, and the output is the original text.", "type": "text" } ], "index": 39 } ], "index": 36, "bbox_fs": [ 105, 513, 507, 592 ] }, { "type": "text", "bbox": [ 107, 596, 505, 651 ], "lines": [ { "bbox": [ 106, 596, 505, 608 ], "spans": [ { "bbox": [ 106, 596, 505, 608 ], "score": 1.0, "content": "Video. We use the following diverse and high-quality video datasets to train video generation and", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 606, 506, 620 ], "spans": [ { "bbox": [ 105, 606, 506, 620 ], "score": 1.0, "content": "video prompt encoder. WebVid [4], a large-scale dataset of web videos together with descriptions;", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 618, 505, 630 ], "spans": [ { "bbox": [ 106, 618, 471, 630 ], "score": 1.0, "content": "HD-Villa-100M [54] with high resolution YouTube videos of at least 720P. We perform text", "type": "text" }, { "bbox": [ 471, 619, 481, 627 ], "score": 0.73, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 618, 505, 630 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 628, 505, 642 ], "spans": [ { "bbox": [ 106, 628, 470, 642 ], "score": 1.0, "content": "and video-text contrastive learning task with WebVid. We use HD-Villa-100M for image", "type": "text" }, { "bbox": [ 470, 631, 481, 639 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 628, 505, 642 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 639, 324, 654 ], "spans": [ { "bbox": [ 105, 639, 324, 654 ], "score": 1.0, "content": "generation where the middle frame is the input image.", "type": "text" } ], "index": 44 } ], "index": 42, "bbox_fs": [ 105, 596, 506, 654 ] }, { "type": "text", "bbox": [ 107, 656, 505, 722 ], "lines": [ { "bbox": [ 105, 654, 505, 669 ], "spans": [ { "bbox": [ 105, 654, 505, 669 ], "score": 1.0, "content": "Audiovisual. Web videos are a natural aligned audio-video data resource. However, many existing", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 667, 506, 678 ], "spans": [ { "bbox": [ 106, 667, 506, 678 ], "score": 1.0, "content": "datasets, e.g., ACAV100M [28], feature heavily on videos of human speech rather than natural sounds.", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 677, 506, 690 ], "spans": [ { "bbox": [ 105, 677, 506, 690 ], "score": 1.0, "content": "Therefore, we leverage sound-oriented datasets AudioSet and SoundNet [3] for joint audio-video", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 689, 506, 702 ], "spans": [ { "bbox": [ 105, 689, 199, 702 ], "score": 1.0, "content": "generation. For image", "type": "text" }, { "bbox": [ 200, 690, 210, 699 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 210, 689, 234, 702 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 235, 690, 244, 699 ], "score": 0.75, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 244, 689, 506, 702 ], "score": 1.0, "content": "video, we use the middle frame of the target video as the input", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 699, 506, 713 ], "spans": [ { "bbox": [ 105, 699, 506, 713 ], "score": 1.0, "content": "prompt image. We also use the middle frame as the prompt input to train the model to generate the", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 711, 214, 723 ], "spans": [ { "bbox": [ 106, 711, 177, 723 ], "score": 1.0, "content": "audio, i.e., image", "type": "text" }, { "bbox": [ 177, 712, 187, 721 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 187, 711, 214, 723 ], "score": 1.0, "content": "audio.", "type": "text" } ], "index": 50 } ], "index": 47.5, "bbox_fs": [ 105, 654, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 71, 504, 371 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 71, 504, 371 ], "group_id": 0, "lines": [ { "bbox": [ 107, 71, 504, 371 ], "spans": [ { "bbox": [ 107, 71, 504, 371 ], "score": 0.966, "type": "image", "image_path": "dc63eb6c668a08077a3c76c77a471ccad92d8b97c381a63765856efb970b42b0.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 71, 504, 171.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 171.0, 504, 271.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 271.0, 504, 371.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 104, 379, 505, 402 ], "group_id": 0, "lines": [ { "bbox": [ 105, 378, 507, 393 ], "spans": [ { "bbox": [ 105, 378, 467, 393 ], "score": 1.0, "content": "Figure 4: Generation with multiple input modality conditions. Top to bottom: text+audio", "type": "text" }, { "bbox": [ 467, 381, 477, 389 ], "score": 0.62, "content": "", "type": "inline_equation" }, { "bbox": [ 478, 378, 507, 393 ], "score": 1.0, "content": "image,", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 391, 266, 402 ], "spans": [ { "bbox": [ 106, 391, 150, 402 ], "score": 1.0, "content": "text+audio", "type": "text" }, { "bbox": [ 150, 392, 160, 400 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 160, 391, 236, 402 ], "score": 1.0, "content": "video, video+audio", "type": "text" }, { "bbox": [ 237, 392, 247, 400 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 247, 391, 266, 402 ], "score": 1.0, "content": "text.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "title", "bbox": [ 107, 421, 222, 435 ], "lines": [ { "bbox": [ 104, 420, 223, 437 ], "spans": [ { "bbox": [ 104, 420, 223, 437 ], "score": 1.0, "content": "5 Evaluation Results", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 108, 447, 505, 481 ], "lines": [ { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "In this section, we will evaluate the model generation quality in different settings including single", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 458, 505, 471 ], "spans": [ { "bbox": [ 105, 458, 505, 471 ], "score": 1.0, "content": "modality generation, multi-condition generation, and multi-output joint generation. We provide both", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 470, 506, 481 ], "spans": [ { "bbox": [ 106, 470, 506, 481 ], "score": 1.0, "content": "quantitative benchmarking on evaluation datasets as well as qualitative visualization demonstrations.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "title", "bbox": [ 108, 495, 282, 507 ], "lines": [ { "bbox": [ 105, 495, 283, 509 ], "spans": [ { "bbox": [ 105, 495, 283, 509 ], "score": 1.0, "content": "5.1 Single Modality Generation Results", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 106, 516, 505, 604 ], "lines": [ { "bbox": [ 106, 516, 506, 528 ], "spans": [ { "bbox": [ 106, 516, 506, 528 ], "score": 1.0, "content": "We first show example demo in Fig. 3, where we present various single to single modality generation.", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 528, 505, 540 ], "spans": [ { "bbox": [ 106, 528, 505, 540 ], "score": 1.0, "content": "Then, we evaluate the synthesis quality of the unimodal generation on text, image, video, and", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 538, 506, 550 ], "spans": [ { "bbox": [ 105, 538, 506, 550 ], "score": 1.0, "content": "audio. CoDi achieves SOTA on audio captions and audio generation, as shown in Table 7 and", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 550, 505, 560 ], "spans": [ { "bbox": [ 106, 550, 505, 560 ], "score": 1.0, "content": "Table 5. Notably for the first time in the field, CoDi, a diffusion-base model, exhibits comparable", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 560, 505, 572 ], "spans": [ { "bbox": [ 105, 560, 505, 572 ], "score": 1.0, "content": "performance on image captioning with autoregressive transformer-based SOTA (Table 6). CoDi is", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 571, 505, 583 ], "spans": [ { "bbox": [ 106, 571, 505, 583 ], "score": 1.0, "content": "the first diffusion-model based for video captioning Table 8. On image and video generation, CoDi", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 582, 506, 595 ], "spans": [ { "bbox": [ 105, 582, 506, 595 ], "score": 1.0, "content": "performs competitively with state-of-the-art (Tables 2 to 4). This gives us strong starting points for", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 592, 506, 604 ], "spans": [ { "bbox": [ 105, 592, 506, 604 ], "score": 1.0, "content": "multi-condition and multi-output generation that will be presented next in Section 5.2 and Section 5.3.", "type": "text" } ], "index": 17 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 609, 504, 631 ], "lines": [ { "bbox": [ 106, 608, 506, 621 ], "spans": [ { "bbox": [ 106, 608, 506, 621 ], "score": 1.0, "content": "We demonstrate in Section 3.2 that CoDi is capable of integrating representation from different", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 621, 506, 632 ], "spans": [ { "bbox": [ 106, 621, 506, 632 ], "score": 1.0, "content": "modalities in the generation. Thus, we first show multi-condition generation demo as shown in Fig. 4.", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 645, 284, 658 ], "lines": [ { "bbox": [ 105, 645, 285, 658 ], "spans": [ { "bbox": [ 105, 645, 285, 658 ], "score": 1.0, "content": "5.2 Multi-Condition Generation Results", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 667, 504, 722 ], "lines": [ { "bbox": [ 106, 667, 505, 679 ], "spans": [ { "bbox": [ 106, 667, 505, 679 ], "score": 1.0, "content": "For quantitative evaluation, we focus on multiple inputs to image synthesis output since the evaluation", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 678, 505, 690 ], "spans": [ { "bbox": [ 105, 678, 505, 690 ], "score": 1.0, "content": "metric for this case (FID) does not require specific modality inputs like text. We test with several", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 689, 505, 701 ], "spans": [ { "bbox": [ 106, 689, 240, 701 ], "score": 1.0, "content": "input combinations including text", "type": "text" }, { "bbox": [ 240, 690, 249, 699 ], "score": 0.71, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 249, 689, 294, 701 ], "score": 1.0, "content": "image, text", "type": "text" }, { "bbox": [ 294, 690, 303, 699 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 303, 689, 355, 701 ], "score": 1.0, "content": "audio, image", "type": "text" }, { "bbox": [ 355, 690, 363, 699 ], "score": 0.7, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 364, 689, 407, 701 ], "score": 1.0, "content": "audio, text", "type": "text" }, { "bbox": [ 407, 690, 415, 699 ], "score": 0.74, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 416, 689, 505, 701 ], "score": 1.0, "content": "video, as well as three", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 149, 712 ], "score": 1.0, "content": "inputs text", "type": "text" }, { "bbox": [ 149, 701, 158, 710 ], "score": 0.74, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 158, 699, 181, 712 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 182, 701, 190, 710 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 190, 699, 506, 712 ], "score": 1.0, "content": "image. We test on the validation set of AudioCaps [24] since all four modalities", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "are present in this dataset. The prompt image input is the middle frame of the video. As shown in", "type": "text" } ], "index": 25 } ], "index": 23 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 308, 750 ], "lines": [ { "bbox": [ 301, 740, 310, 752 ], "spans": [ { "bbox": [ 301, 740, 310, 752 ], "score": 1.0, "content": "", "type": "text", "height": 12, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 107, 71, 504, 371 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 71, 504, 371 ], "group_id": 0, "lines": [ { "bbox": [ 107, 71, 504, 371 ], "spans": [ { "bbox": [ 107, 71, 504, 371 ], "score": 0.966, "type": "image", "image_path": "dc63eb6c668a08077a3c76c77a471ccad92d8b97c381a63765856efb970b42b0.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 71, 504, 171.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 171.0, 504, 271.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 271.0, 504, 371.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 104, 379, 505, 402 ], "group_id": 0, "lines": [ { "bbox": [ 105, 378, 507, 393 ], "spans": [ { "bbox": [ 105, 378, 467, 393 ], "score": 1.0, "content": "Figure 4: Generation with multiple input modality conditions. Top to bottom: text+audio", "type": "text" }, { "bbox": [ 467, 381, 477, 389 ], "score": 0.62, "content": "", "type": "inline_equation" }, { "bbox": [ 478, 378, 507, 393 ], "score": 1.0, "content": "image,", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 391, 266, 402 ], "spans": [ { "bbox": [ 106, 391, 150, 402 ], "score": 1.0, "content": "text+audio", "type": "text" }, { "bbox": [ 150, 392, 160, 400 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 160, 391, 236, 402 ], "score": 1.0, "content": "video, video+audio", "type": "text" }, { "bbox": [ 237, 392, 247, 400 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 247, 391, 266, 402 ], "score": 1.0, "content": "text.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "title", "bbox": [ 107, 421, 222, 435 ], "lines": [ { "bbox": [ 104, 420, 223, 437 ], "spans": [ { "bbox": [ 104, 420, 223, 437 ], "score": 1.0, "content": "5 Evaluation Results", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 108, 447, 505, 481 ], "lines": [ { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "In this section, we will evaluate the model generation quality in different settings including single", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 458, 505, 471 ], "spans": [ { "bbox": [ 105, 458, 505, 471 ], "score": 1.0, "content": "modality generation, multi-condition generation, and multi-output joint generation. We provide both", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 470, 506, 481 ], "spans": [ { "bbox": [ 106, 470, 506, 481 ], "score": 1.0, "content": "quantitative benchmarking on evaluation datasets as well as qualitative visualization demonstrations.", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 105, 447, 506, 481 ] }, { "type": "title", "bbox": [ 108, 495, 282, 507 ], "lines": [ { "bbox": [ 105, 495, 283, 509 ], "spans": [ { "bbox": [ 105, 495, 283, 509 ], "score": 1.0, "content": "5.1 Single Modality Generation Results", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 106, 516, 505, 604 ], "lines": [ { "bbox": [ 106, 516, 506, 528 ], "spans": [ { "bbox": [ 106, 516, 506, 528 ], "score": 1.0, "content": "We first show example demo in Fig. 3, where we present various single to single modality generation.", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 528, 505, 540 ], "spans": [ { "bbox": [ 106, 528, 505, 540 ], "score": 1.0, "content": "Then, we evaluate the synthesis quality of the unimodal generation on text, image, video, and", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 538, 506, 550 ], "spans": [ { "bbox": [ 105, 538, 506, 550 ], "score": 1.0, "content": "audio. CoDi achieves SOTA on audio captions and audio generation, as shown in Table 7 and", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 550, 505, 560 ], "spans": [ { "bbox": [ 106, 550, 505, 560 ], "score": 1.0, "content": "Table 5. Notably for the first time in the field, CoDi, a diffusion-base model, exhibits comparable", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 560, 505, 572 ], "spans": [ { "bbox": [ 105, 560, 505, 572 ], "score": 1.0, "content": "performance on image captioning with autoregressive transformer-based SOTA (Table 6). CoDi is", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 571, 505, 583 ], "spans": [ { "bbox": [ 106, 571, 505, 583 ], "score": 1.0, "content": "the first diffusion-model based for video captioning Table 8. On image and video generation, CoDi", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 582, 506, 595 ], "spans": [ { "bbox": [ 105, 582, 506, 595 ], "score": 1.0, "content": "performs competitively with state-of-the-art (Tables 2 to 4). This gives us strong starting points for", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 592, 506, 604 ], "spans": [ { "bbox": [ 105, 592, 506, 604 ], "score": 1.0, "content": "multi-condition and multi-output generation that will be presented next in Section 5.2 and Section 5.3.", "type": "text" } ], "index": 17 } ], "index": 13.5, "bbox_fs": [ 105, 516, 506, 604 ] }, { "type": "text", "bbox": [ 106, 609, 504, 631 ], "lines": [ { "bbox": [ 106, 608, 506, 621 ], "spans": [ { "bbox": [ 106, 608, 506, 621 ], "score": 1.0, "content": "We demonstrate in Section 3.2 that CoDi is capable of integrating representation from different", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 621, 506, 632 ], "spans": [ { "bbox": [ 106, 621, 506, 632 ], "score": 1.0, "content": "modalities in the generation. Thus, we first show multi-condition generation demo as shown in Fig. 4.", "type": "text" } ], "index": 19 } ], "index": 18.5, "bbox_fs": [ 106, 608, 506, 632 ] }, { "type": "title", "bbox": [ 108, 645, 284, 658 ], "lines": [ { "bbox": [ 105, 645, 285, 658 ], "spans": [ { "bbox": [ 105, 645, 285, 658 ], "score": 1.0, "content": "5.2 Multi-Condition Generation Results", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 667, 504, 722 ], "lines": [ { "bbox": [ 106, 667, 505, 679 ], "spans": [ { "bbox": [ 106, 667, 505, 679 ], "score": 1.0, "content": "For quantitative evaluation, we focus on multiple inputs to image synthesis output since the evaluation", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 678, 505, 690 ], "spans": [ { "bbox": [ 105, 678, 505, 690 ], "score": 1.0, "content": "metric for this case (FID) does not require specific modality inputs like text. We test with several", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 689, 505, 701 ], "spans": [ { "bbox": [ 106, 689, 240, 701 ], "score": 1.0, "content": "input combinations including text", "type": "text" }, { "bbox": [ 240, 690, 249, 699 ], "score": 0.71, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 249, 689, 294, 701 ], "score": 1.0, "content": "image, text", "type": "text" }, { "bbox": [ 294, 690, 303, 699 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 303, 689, 355, 701 ], "score": 1.0, "content": "audio, image", "type": "text" }, { "bbox": [ 355, 690, 363, 699 ], "score": 0.7, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 364, 689, 407, 701 ], "score": 1.0, "content": "audio, text", "type": "text" }, { "bbox": [ 407, 690, 415, 699 ], "score": 0.74, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 416, 689, 505, 701 ], "score": 1.0, "content": "video, as well as three", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 149, 712 ], "score": 1.0, "content": "inputs text", "type": "text" }, { "bbox": [ 149, 701, 158, 710 ], "score": 0.74, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 158, 699, 181, 712 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 182, 701, 190, 710 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 190, 699, 506, 712 ], "score": 1.0, "content": "image. We test on the validation set of AudioCaps [24] since all four modalities", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "are present in this dataset. The prompt image input is the middle frame of the video. As shown in", "type": "text" } ], "index": 25 } ], "index": 23, "bbox_fs": [ 105, 667, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 145, 109, 270, 180 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 307, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 68, 307, 82 ], "spans": [ { "bbox": [ 105, 68, 307, 82 ], "score": 1.0, "content": "Table 9: CoDi is capable of generating high quality", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 80, 308, 92 ], "spans": [ { "bbox": [ 106, 80, 308, 92 ], "score": 1.0, "content": "output (image in this case) from various combina-", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 92, 217, 103 ], "spans": [ { "bbox": [ 106, 92, 217, 103 ], "score": 1.0, "content": "tions of prompt modalities.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 145, 109, 270, 180 ], "group_id": 0, "lines": [ { "bbox": [ 145, 109, 270, 180 ], "spans": [ { "bbox": [ 145, 109, 270, 180 ], "score": 0.963, "html": "
InputsFID↓
Single-modality Prompt
Text14.2
Audio14.3
Dual-modality Prompt
Text+Audio14.9
", "type": "table", "image_path": "e4534d7ebcfba76448956e498aa36d45e0b4f106882850134d22e24e5043a3c9.jpg" } ] } ], "index": 5.5, "virtual_lines": [ { "bbox": [ 145, 109, 270, 144.5 ], "spans": [], "index": 5 }, { "bbox": [ 145, 144.5, 270, 180.0 ], "spans": [], "index": 6 } ] } ], "index": 3.75 }, { "type": "table", "bbox": [ 344, 97, 493, 180 ], "blocks": [ { "type": "table_caption", "bbox": [ 318, 70, 519, 92 ], "group_id": 1, "lines": [ { "bbox": [ 317, 69, 520, 83 ], "spans": [ { "bbox": [ 317, 69, 520, 83 ], "score": 1.0, "content": "Table 10: MSR-VTT text-to-video generation per-", "type": "text" } ], "index": 1 }, { "bbox": [ 317, 82, 361, 93 ], "spans": [ { "bbox": [ 317, 82, 361, 93 ], "score": 1.0, "content": "formance.", "type": "text" } ], "index": 3 } ], "index": 2.0 }, { "type": "table_body", "bbox": [ 344, 97, 493, 180 ], "group_id": 1, "lines": [ { "bbox": [ 344, 97, 493, 180 ], "spans": [ { "bbox": [ 344, 97, 493, 180 ], "score": 0.967, "html": "
InputsCLIPSIM个
Single-modality Prompt
Text0.2890
Dual-modality Prompt
Text+Audio0.2912
Text+Image0.2891
Text+Audio+Image0.2923
", "type": "table", "image_path": "5dac823453343eca2eec8a73d3324d6d6bf7776e7f743d995f77fa1083127e85.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 344, 97, 493, 138.5 ], "spans": [], "index": 7 }, { "bbox": [ 344, 138.5, 493, 180.0 ], "spans": [], "index": 8 } ] } ], "index": 4.75 }, { "type": "image", "bbox": [ 109, 201, 501, 473 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 201, 501, 473 ], "group_id": 0, "lines": [ { "bbox": [ 109, 201, 501, 473 ], "spans": [ { "bbox": [ 109, 201, 501, 473 ], "score": 0.961, "type": "image", "image_path": "570523fa62e48124b6e7ee0927d72eb815a36ea17ca483b39355fc5b79c4dd8d.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 109, 201, 501, 291.6666666666667 ], "spans": [], "index": 9 }, { "bbox": [ 109, 291.6666666666667, 501, 382.33333333333337 ], "spans": [], "index": 10 }, { "bbox": [ 109, 382.33333333333337, 501, 473.00000000000006 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 481, 504, 504 ], "group_id": 0, "lines": [ { "bbox": [ 105, 480, 506, 495 ], "spans": [ { "bbox": [ 105, 480, 506, 495 ], "score": 1.0, "content": "Figure 5: Joint generation of multiple output modalities by CoDi. From top to bottom:", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 493, 425, 505 ], "spans": [ { "bbox": [ 106, 493, 122, 505 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 494, 132, 502 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 133, 493, 201, 505 ], "score": 1.0, "content": "video+audio, tex", "type": "text" }, { "bbox": [ 201, 494, 212, 502 ], "score": 0.77, "content": "", "type": "inline_equation" }, { "bbox": [ 212, 493, 361, 505 ], "score": 1.0, "content": "image+text+audio, text+audio+image", "type": "text" }, { "bbox": [ 362, 494, 372, 502 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 372, 493, 425, 505 ], "score": 1.0, "content": "video+audio.", "type": "text" } ], "index": 13 } ], "index": 12.5 } ], "index": 11.25 }, { "type": "text", "bbox": [ 106, 525, 505, 614 ], "lines": [ { "bbox": [ 104, 524, 506, 539 ], "spans": [ { "bbox": [ 104, 524, 506, 539 ], "score": 1.0, "content": "Table 9, CoDi achieves high image generation quality given assorted groups of input modalities. We", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 537, 506, 549 ], "spans": [ { "bbox": [ 105, 537, 435, 549 ], "score": 1.0, "content": "also test with several input combinations with video as output including text, text", "type": "text" }, { "bbox": [ 435, 538, 443, 547 ], "score": 0.75, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 443, 537, 497, 549 ], "score": 1.0, "content": "audio, image", "type": "text" }, { "bbox": [ 498, 538, 506, 547 ], "score": 0.73, "content": "^ +", "type": "inline_equation" } ], "index": 15 }, { "bbox": [ 105, 548, 506, 560 ], "spans": [ { "bbox": [ 105, 548, 194, 560 ], "score": 1.0, "content": "image, as well as text", "type": "text" }, { "bbox": [ 195, 549, 203, 558 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 203, 548, 227, 560 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 227, 549, 236, 558 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 236, 548, 506, 560 ], "score": 1.0, "content": "image. We also test on MSRVTT [24] since all four modalities are", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 559, 505, 570 ], "spans": [ { "bbox": [ 105, 559, 505, 570 ], "score": 1.0, "content": "present in this dataset. Similarly, the prompt image input is the middle frame of the video. As shown", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 569, 505, 582 ], "spans": [ { "bbox": [ 105, 569, 505, 582 ], "score": 1.0, "content": "in Table 10, CoDi achieves high video and ground truth text similarity given assorted groups of input", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 580, 505, 592 ], "spans": [ { "bbox": [ 106, 580, 461, 592 ], "score": 1.0, "content": "modalities. Again our model does not need to train on multi-condition generation like text", "type": "text" }, { "bbox": [ 462, 582, 470, 591 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 470, 580, 505, 592 ], "score": 1.0, "content": "audio or", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 591, 505, 604 ], "spans": [ { "bbox": [ 106, 591, 123, 604 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 592, 132, 601 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 132, 591, 505, 604 ], "score": 1.0, "content": "image. Through bridging alignment and composable multimodal conditioning as proposed in", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 601, 480, 614 ], "spans": [ { "bbox": [ 105, 601, 480, 614 ], "score": 1.0, "content": "Section 3.2, our model trained on single condition can zero-shot infer on multiple conditions.", "type": "text" } ], "index": 21 } ], "index": 17.5 }, { "type": "title", "bbox": [ 107, 627, 297, 640 ], "lines": [ { "bbox": [ 106, 627, 298, 641 ], "spans": [ { "bbox": [ 106, 627, 298, 641 ], "score": 1.0, "content": "5.3 Multi-Output Joint Generation Results", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 648, 505, 703 ], "lines": [ { "bbox": [ 106, 648, 505, 659 ], "spans": [ { "bbox": [ 106, 648, 505, 659 ], "score": 1.0, "content": "For joint multimodal generation, we first demonstrate high-quality multimodal output joint generation", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 658, 505, 672 ], "spans": [ { "bbox": [ 105, 658, 505, 672 ], "score": 1.0, "content": "demo as shown in Fig. 5. For quantitative evaluation, there is no existing evaluation metric since we", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 668, 505, 684 ], "spans": [ { "bbox": [ 105, 668, 505, 684 ], "score": 1.0, "content": "are the first model that can simultaneously generate across all 4 modalities. Therefore, we propose", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "the following metric SIM that quantifies the coherence and consistency between the two generated", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 690, 296, 705 ], "spans": [ { "bbox": [ 105, 690, 296, 705 ], "score": 1.0, "content": "modalities by cosine similarity of embeddings:", "type": "text" } ], "index": 27 } ], "index": 25 }, { "type": "interline_equation", "bbox": [ 232, 709, 379, 723 ], "lines": [ { "bbox": [ 232, 709, 379, 723 ], "spans": [ { "bbox": [ 232, 709, 379, 723 ], "score": 0.91, "content": "\\operatorname { S I M } ( A , B ) = \\cos { ( C _ { A } ( A ) , C _ { B } ( B ) ) }", "type": "interline_equation", "image_path": "442991e83372788db1166c87e1f118b81b1449278a900571ed564c3772d81cd4.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 232, 709, 379, 723 ], "spans": [], "index": 28 } ] } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 145, 109, 270, 180 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 307, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 68, 307, 82 ], "spans": [ { "bbox": [ 105, 68, 307, 82 ], "score": 1.0, "content": "Table 9: CoDi is capable of generating high quality", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 80, 308, 92 ], "spans": [ { "bbox": [ 106, 80, 308, 92 ], "score": 1.0, "content": "output (image in this case) from various combina-", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 92, 217, 103 ], "spans": [ { "bbox": [ 106, 92, 217, 103 ], "score": 1.0, "content": "tions of prompt modalities.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 145, 109, 270, 180 ], "group_id": 0, "lines": [ { "bbox": [ 145, 109, 270, 180 ], "spans": [ { "bbox": [ 145, 109, 270, 180 ], "score": 0.963, "html": "
InputsFID↓
Single-modality Prompt
Text14.2
Audio14.3
Dual-modality Prompt
Text+Audio14.9
", "type": "table", "image_path": "e4534d7ebcfba76448956e498aa36d45e0b4f106882850134d22e24e5043a3c9.jpg" } ] } ], "index": 5.5, "virtual_lines": [ { "bbox": [ 145, 109, 270, 144.5 ], "spans": [], "index": 5 }, { "bbox": [ 145, 144.5, 270, 180.0 ], "spans": [], "index": 6 } ] } ], "index": 3.75 }, { "type": "table", "bbox": [ 344, 97, 493, 180 ], "blocks": [ { "type": "table_caption", "bbox": [ 318, 70, 519, 92 ], "group_id": 1, "lines": [ { "bbox": [ 317, 69, 520, 83 ], "spans": [ { "bbox": [ 317, 69, 520, 83 ], "score": 1.0, "content": "Table 10: MSR-VTT text-to-video generation per-", "type": "text" } ], "index": 1 }, { "bbox": [ 317, 82, 361, 93 ], "spans": [ { "bbox": [ 317, 82, 361, 93 ], "score": 1.0, "content": "formance.", "type": "text" } ], "index": 3 } ], "index": 2.0 }, { "type": "table_body", "bbox": [ 344, 97, 493, 180 ], "group_id": 1, "lines": [ { "bbox": [ 344, 97, 493, 180 ], "spans": [ { "bbox": [ 344, 97, 493, 180 ], "score": 0.967, "html": "
InputsCLIPSIM个
Single-modality Prompt
Text0.2890
Dual-modality Prompt
Text+Audio0.2912
Text+Image0.2891
Text+Audio+Image0.2923
", "type": "table", "image_path": "5dac823453343eca2eec8a73d3324d6d6bf7776e7f743d995f77fa1083127e85.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 344, 97, 493, 138.5 ], "spans": [], "index": 7 }, { "bbox": [ 344, 138.5, 493, 180.0 ], "spans": [], "index": 8 } ] } ], "index": 4.75 }, { "type": "image", "bbox": [ 109, 201, 501, 473 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 201, 501, 473 ], "group_id": 0, "lines": [ { "bbox": [ 109, 201, 501, 473 ], "spans": [ { "bbox": [ 109, 201, 501, 473 ], "score": 0.961, "type": "image", "image_path": "570523fa62e48124b6e7ee0927d72eb815a36ea17ca483b39355fc5b79c4dd8d.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 109, 201, 501, 291.6666666666667 ], "spans": [], "index": 9 }, { "bbox": [ 109, 291.6666666666667, 501, 382.33333333333337 ], "spans": [], "index": 10 }, { "bbox": [ 109, 382.33333333333337, 501, 473.00000000000006 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 481, 504, 504 ], "group_id": 0, "lines": [ { "bbox": [ 105, 480, 506, 495 ], "spans": [ { "bbox": [ 105, 480, 506, 495 ], "score": 1.0, "content": "Figure 5: Joint generation of multiple output modalities by CoDi. From top to bottom:", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 493, 425, 505 ], "spans": [ { "bbox": [ 106, 493, 122, 505 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 494, 132, 502 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 133, 493, 201, 505 ], "score": 1.0, "content": "video+audio, tex", "type": "text" }, { "bbox": [ 201, 494, 212, 502 ], "score": 0.77, "content": "", "type": "inline_equation" }, { "bbox": [ 212, 493, 361, 505 ], "score": 1.0, "content": "image+text+audio, text+audio+image", "type": "text" }, { "bbox": [ 362, 494, 372, 502 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 372, 493, 425, 505 ], "score": 1.0, "content": "video+audio.", "type": "text" } ], "index": 13 } ], "index": 12.5 } ], "index": 11.25 }, { "type": "text", "bbox": [ 106, 525, 505, 614 ], "lines": [ { "bbox": [ 104, 524, 506, 539 ], "spans": [ { "bbox": [ 104, 524, 506, 539 ], "score": 1.0, "content": "Table 9, CoDi achieves high image generation quality given assorted groups of input modalities. We", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 537, 506, 549 ], "spans": [ { "bbox": [ 105, 537, 435, 549 ], "score": 1.0, "content": "also test with several input combinations with video as output including text, text", "type": "text" }, { "bbox": [ 435, 538, 443, 547 ], "score": 0.75, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 443, 537, 497, 549 ], "score": 1.0, "content": "audio, image", "type": "text" }, { "bbox": [ 498, 538, 506, 547 ], "score": 0.73, "content": "^ +", "type": "inline_equation" } ], "index": 15 }, { "bbox": [ 105, 548, 506, 560 ], "spans": [ { "bbox": [ 105, 548, 194, 560 ], "score": 1.0, "content": "image, as well as text", "type": "text" }, { "bbox": [ 195, 549, 203, 558 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 203, 548, 227, 560 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 227, 549, 236, 558 ], "score": 0.73, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 236, 548, 506, 560 ], "score": 1.0, "content": "image. We also test on MSRVTT [24] since all four modalities are", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 559, 505, 570 ], "spans": [ { "bbox": [ 105, 559, 505, 570 ], "score": 1.0, "content": "present in this dataset. Similarly, the prompt image input is the middle frame of the video. As shown", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 569, 505, 582 ], "spans": [ { "bbox": [ 105, 569, 505, 582 ], "score": 1.0, "content": "in Table 10, CoDi achieves high video and ground truth text similarity given assorted groups of input", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 580, 505, 592 ], "spans": [ { "bbox": [ 106, 580, 461, 592 ], "score": 1.0, "content": "modalities. Again our model does not need to train on multi-condition generation like text", "type": "text" }, { "bbox": [ 462, 582, 470, 591 ], "score": 0.76, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 470, 580, 505, 592 ], "score": 1.0, "content": "audio or", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 591, 505, 604 ], "spans": [ { "bbox": [ 106, 591, 123, 604 ], "score": 1.0, "content": "text", "type": "text" }, { "bbox": [ 123, 592, 132, 601 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 132, 591, 505, 604 ], "score": 1.0, "content": "image. Through bridging alignment and composable multimodal conditioning as proposed in", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 601, 480, 614 ], "spans": [ { "bbox": [ 105, 601, 480, 614 ], "score": 1.0, "content": "Section 3.2, our model trained on single condition can zero-shot infer on multiple conditions.", "type": "text" } ], "index": 21 } ], "index": 17.5, "bbox_fs": [ 104, 524, 506, 614 ] }, { "type": "title", "bbox": [ 107, 627, 297, 640 ], "lines": [ { "bbox": [ 106, 627, 298, 641 ], "spans": [ { "bbox": [ 106, 627, 298, 641 ], "score": 1.0, "content": "5.3 Multi-Output Joint Generation Results", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 648, 505, 703 ], "lines": [ { "bbox": [ 106, 648, 505, 659 ], "spans": [ { "bbox": [ 106, 648, 505, 659 ], "score": 1.0, "content": "For joint multimodal generation, we first demonstrate high-quality multimodal output joint generation", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 658, 505, 672 ], "spans": [ { "bbox": [ 105, 658, 505, 672 ], "score": 1.0, "content": "demo as shown in Fig. 5. For quantitative evaluation, there is no existing evaluation metric since we", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 668, 505, 684 ], "spans": [ { "bbox": [ 105, 668, 505, 684 ], "score": 1.0, "content": "are the first model that can simultaneously generate across all 4 modalities. Therefore, we propose", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "the following metric SIM that quantifies the coherence and consistency between the two generated", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 690, 296, 705 ], "spans": [ { "bbox": [ 105, 690, 296, 705 ], "score": 1.0, "content": "modalities by cosine similarity of embeddings:", "type": "text" } ], "index": 27 } ], "index": 25, "bbox_fs": [ 105, 648, 505, 705 ] }, { "type": "interline_equation", "bbox": [ 232, 709, 379, 723 ], "lines": [ { "bbox": [ 232, 709, 379, 723 ], "spans": [ { "bbox": [ 232, 709, 379, 723 ], "score": 0.91, "content": "\\operatorname { S I M } ( A , B ) = \\cos { ( C _ { A } ( A ) , C _ { B } ( B ) ) }", "type": "interline_equation", "image_path": "442991e83372788db1166c87e1f118b81b1449278a900571ed564c3772d81cd4.jpg" } ] } ], "index": 28, "virtual_lines": [ { "bbox": [ 232, 709, 379, 723 ], "spans": [], "index": 28 } ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 109, 108, 500, 242 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 505, 103 ], "group_id": 0, "lines": [ { "bbox": [ 105, 69, 505, 82 ], "spans": [ { "bbox": [ 105, 69, 447, 82 ], "score": 1.0, "content": "Table 11: Similarity scores between generated modalities. The number on the left of", "type": "text" }, { "bbox": [ 448, 70, 461, 80 ], "score": 0.36, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 461, 69, 505, 82 ], "score": 1.0, "content": "represents", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 79, 506, 93 ], "spans": [ { "bbox": [ 105, 79, 506, 93 ], "score": 1.0, "content": "the similarity score of independent generation, and the right it represents the case of joint generation.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 91, 363, 104 ], "spans": [ { "bbox": [ 105, 91, 363, 104 ], "score": 1.0, "content": "Jointly generated outputs consistently show stronger coherence.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 109, 108, 500, 242 ], "group_id": 0, "lines": [ { "bbox": [ 109, 108, 500, 242 ], "spans": [ { "bbox": [ 109, 108, 500, 242 ], "score": 0.983, "html": "
InputsSIM-ITSIM-ATSIM-VTSIM-VA
Two Joint Outputs
Audio → Image+Text0.251 / 0.260
Image→Audio+Text■0.244 / 0.256
Text →Video+Audio0.240 / 0.255
Audio →Video+Text0.256 / 0.261
Three Joint Outputs
Text-→ Video+Image+Audio 0.256/0.270 0.240/0.2570.240 / 0.257
Multi-Inputs-Outputs
Text+Image -→ Video+Audio0.247 / 0.259
", "type": "table", "image_path": "7c6e18d97f95273627bd5cdb16fd3a210522268a86a54bd9275114265361cbb0.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 109, 108, 500, 152.66666666666666 ], "spans": [], "index": 3 }, { "bbox": [ 109, 152.66666666666666, 500, 197.33333333333331 ], "spans": [], "index": 4 }, { "bbox": [ 109, 197.33333333333331, 500, 241.99999999999997 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 263, 505, 308 ], "lines": [ { "bbox": [ 105, 264, 505, 277 ], "spans": [ { "bbox": [ 105, 264, 133, 277 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 264, 142, 274 ], "score": 0.58, "content": "A", "type": "inline_equation" }, { "bbox": [ 142, 264, 146, 277 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 146, 264, 155, 274 ], "score": 0.62, "content": "B", "type": "inline_equation" }, { "bbox": [ 156, 264, 293, 277 ], "score": 1.0, "content": "are the generated modalities, and", "type": "text" }, { "bbox": [ 293, 264, 308, 275 ], "score": 0.9, "content": "C _ { A }", "type": "inline_equation" }, { "bbox": [ 308, 264, 326, 277 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 327, 264, 342, 275 ], "score": 0.9, "content": "C _ { B }", "type": "inline_equation" }, { "bbox": [ 342, 264, 478, 277 ], "score": 1.0, "content": "are aligned encoders that project", "type": "text" }, { "bbox": [ 478, 264, 487, 274 ], "score": 0.7, "content": "A", "type": "inline_equation" }, { "bbox": [ 487, 264, 505, 277 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 275, 506, 288 ], "spans": [ { "bbox": [ 107, 276, 116, 285 ], "score": 0.79, "content": "B", "type": "inline_equation" }, { "bbox": [ 116, 275, 506, 288 ], "score": 1.0, "content": "to the same space. We use the prompt encoder as described in Section 3.2. This metric aims to", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "compute the cosine similarity of the embedding of two modalities using contrastive learned prompt", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 297, 493, 309 ], "spans": [ { "bbox": [ 106, 297, 493, 309 ], "score": 1.0, "content": "encoders. Thus, the higher the metric, the more aligned and similar the generated modalities are.", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 313, 505, 390 ], "lines": [ { "bbox": [ 105, 313, 505, 326 ], "spans": [ { "bbox": [ 105, 313, 441, 326 ], "score": 1.0, "content": "To demonstrate the effectiveness of joint generation, assume the prompt modality is", "type": "text" }, { "bbox": [ 442, 314, 450, 323 ], "score": 0.8, "content": "P", "type": "inline_equation" }, { "bbox": [ 451, 313, 505, 326 ], "score": 1.0, "content": ", we compare", "type": "text" } ], "index": 10 }, { "bbox": [ 107, 323, 505, 337 ], "spans": [ { "bbox": [ 107, 324, 154, 336 ], "score": 0.91, "content": "\\mathrm { S I M } ( A , B )", "type": "inline_equation" }, { "bbox": [ 154, 323, 168, 337 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 168, 324, 177, 334 ], "score": 0.71, "content": "A", "type": "inline_equation" }, { "bbox": [ 178, 323, 197, 337 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 198, 324, 207, 334 ], "score": 0.77, "content": "B", "type": "inline_equation" }, { "bbox": [ 208, 323, 369, 337 ], "score": 1.0, "content": "generated separately vs. jointly, i.e.,", "type": "text" }, { "bbox": [ 369, 324, 453, 335 ], "score": 0.75, "content": "\\{ P \\ { \\overset { - } { \\to } } \\ A , \\ P \\ { \\overset { - } { \\to } } \\ B \\}", "type": "inline_equation" }, { "bbox": [ 454, 323, 477, 337 ], "score": 1.0, "content": "vs.", "type": "text" }, { "bbox": [ 477, 324, 505, 335 ], "score": 0.81, "content": "\\{ P ", "type": "inline_equation" } ], "index": 11 }, { "bbox": [ 107, 334, 506, 348 ], "spans": [ { "bbox": [ 107, 336, 140, 347 ], "score": 0.9, "content": "A + B \\}", "type": "inline_equation" }, { "bbox": [ 140, 334, 506, 348 ], "score": 1.0, "content": ". The benchmark is the validation set of AudioCaps [24]. We test on the following settings,", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 347, 505, 358 ], "spans": [ { "bbox": [ 106, 347, 131, 358 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 131, 348, 142, 356 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 143, 347, 219, 358 ], "score": 1.0, "content": "image+text, image", "type": "text" }, { "bbox": [ 219, 348, 231, 356 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 231, 347, 311, 358 ], "score": 1.0, "content": "audio+text, and text", "type": "text" }, { "bbox": [ 312, 348, 321, 356 ], "score": 0.75, "content": "", "type": "inline_equation" }, { "bbox": [ 321, 347, 403, 358 ], "score": 1.0, "content": "video+audio, image", "type": "text" }, { "bbox": [ 403, 347, 414, 356 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 415, 347, 493, 358 ], "score": 1.0, "content": "video+audio. audio", "type": "text" }, { "bbox": [ 493, 347, 505, 356 ], "score": 0.8, "content": "", "type": "inline_equation" } ], "index": 13 }, { "bbox": [ 105, 357, 505, 369 ], "spans": [ { "bbox": [ 105, 357, 177, 369 ], "score": 1.0, "content": "video+text, audio", "type": "text" }, { "bbox": [ 177, 358, 189, 367 ], "score": 0.84, "content": "", "type": "inline_equation" }, { "bbox": [ 189, 357, 284, 369 ], "score": 1.0, "content": "text+video+image, text", "type": "text" }, { "bbox": [ 284, 358, 296, 367 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 357, 505, 369 ], "score": 1.0, "content": "video+image+audio, where the image prompt is the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 368, 505, 380 ], "spans": [ { "bbox": [ 105, 368, 505, 380 ], "score": 1.0, "content": "middle frame of the video clip. As shown in Table 11, joint generation (similarity shown on the right", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 379, 445, 391 ], "spans": [ { "bbox": [ 105, 379, 136, 391 ], "score": 1.0, "content": "side of", "type": "text" }, { "bbox": [ 136, 379, 148, 389 ], "score": 0.6, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 149, 379, 426, 391 ], "score": 1.0, "content": ") consistently outperforms independent generation (on the left side of", "type": "text" }, { "bbox": [ 427, 379, 438, 389 ], "score": 0.49, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 439, 379, 445, 391 ], "score": 1.0, "content": ").", "type": "text" } ], "index": 16 } ], "index": 13 }, { "type": "title", "bbox": [ 107, 407, 183, 420 ], "lines": [ { "bbox": [ 104, 405, 185, 423 ], "spans": [ { "bbox": [ 104, 405, 185, 423 ], "score": 1.0, "content": "6 Conclusion", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 433, 505, 521 ], "lines": [ { "bbox": [ 105, 433, 506, 446 ], "spans": [ { "bbox": [ 105, 433, 506, 446 ], "score": 1.0, "content": "In this paper, we present Composable Diffusion (CoDi), a groundbreaking model in multimodal", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 444, 507, 458 ], "spans": [ { "bbox": [ 104, 444, 507, 458 ], "score": 1.0, "content": "generation that is capable of processing and simultaneously generating modalities across text, image,", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 456, 506, 468 ], "spans": [ { "bbox": [ 106, 456, 506, 468 ], "score": 1.0, "content": "video, and audio. Our approach enables the synergistic generation of high-quality and coherent", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 466, 506, 479 ], "spans": [ { "bbox": [ 105, 466, 506, 479 ], "score": 1.0, "content": "outputs spanning various modalities, from assorted combinations of input modalities. Through", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 477, 505, 490 ], "spans": [ { "bbox": [ 105, 477, 505, 490 ], "score": 1.0, "content": "extensive experiments, we demonstrate CoDi’s remarkable capabilities in flexibly generating single", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 488, 505, 500 ], "spans": [ { "bbox": [ 106, 488, 505, 500 ], "score": 1.0, "content": "or multiple modalities from a wide range of inputs. Our work marks a significant step towards", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 499, 506, 512 ], "spans": [ { "bbox": [ 105, 499, 506, 512 ], "score": 1.0, "content": "more engaging and holistic human-computer interactions, establishing a solid foundation for future", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 510, 305, 522 ], "spans": [ { "bbox": [ 106, 510, 305, 522 ], "score": 1.0, "content": "investigations in generative artificial intelligence.", "type": "text" } ], "index": 25 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 526, 389, 538 ], "lines": [ { "bbox": [ 105, 525, 390, 540 ], "spans": [ { "bbox": [ 105, 525, 390, 540 ], "score": 1.0, "content": "Limitations & Broader Impacts. See Appendix D for the discussion.", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "title", "bbox": [ 108, 555, 202, 568 ], "lines": [ { "bbox": [ 106, 554, 204, 571 ], "spans": [ { "bbox": [ 106, 554, 204, 571 ], "score": 1.0, "content": "Acknowledgement", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 108, 581, 505, 614 ], "lines": [ { "bbox": [ 105, 579, 505, 594 ], "spans": [ { "bbox": [ 105, 579, 505, 594 ], "score": 1.0, "content": "We would like to thank Bei Liu for HD-VILA-100M data support. We also thank Shi Dong, Mahmoud", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 591, 505, 604 ], "spans": [ { "bbox": [ 106, 591, 505, 604 ], "score": 1.0, "content": "Khademi, Junheng Hao, Yuwei Fang, Yichong Xu and Azure Cognitive Services Research team", "type": "text" } ], "index": 29 }, { "bbox": [ 107, 603, 221, 614 ], "spans": [ { "bbox": [ 107, 603, 221, 614 ], "score": 1.0, "content": "members for their feedback.", "type": "text" } ], "index": 30 } ], "index": 29 }, { "type": "title", "bbox": [ 107, 631, 163, 643 ], "lines": [ { "bbox": [ 106, 630, 165, 645 ], "spans": [ { "bbox": [ 106, 630, 165, 645 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 109, 649, 506, 722 ], "lines": [ { "bbox": [ 109, 648, 507, 663 ], "spans": [ { "bbox": [ 109, 648, 507, 663 ], "score": 1.0, "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc,", "type": "text" } ], "index": 32 }, { "bbox": [ 125, 659, 506, 672 ], "spans": [ { "bbox": [ 125, 659, 506, 672 ], "score": 1.0, "content": "Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for", "type": "text" } ], "index": 33 }, { "bbox": [ 126, 670, 483, 682 ], "spans": [ { "bbox": [ 126, 670, 483, 682 ], "score": 1.0, "content": "few-shot learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022. 3", "type": "text" } ], "index": 34 }, { "bbox": [ 110, 691, 506, 704 ], "spans": [ { "bbox": [ 110, 691, 506, 704 ], "score": 1.0, "content": "[2] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift:", "type": "text" } ], "index": 35 }, { "bbox": [ 125, 702, 506, 713 ], "spans": [ { "bbox": [ 125, 702, 506, 713 ], "score": 1.0, "content": "Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477,", "type": "text" } ], "index": 36 }, { "bbox": [ 126, 711, 185, 722 ], "spans": [ { "bbox": [ 126, 711, 185, 722 ], "score": 1.0, "content": "2023. 5, 15, 16", "type": "text" } ], "index": 37 } ], "index": 34.5 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 755 ], "spans": [ { "bbox": [ 299, 740, 313, 755 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 109, 108, 500, 242 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 69, 505, 103 ], "group_id": 0, "lines": [ { "bbox": [ 105, 69, 505, 82 ], "spans": [ { "bbox": [ 105, 69, 447, 82 ], "score": 1.0, "content": "Table 11: Similarity scores between generated modalities. The number on the left of", "type": "text" }, { "bbox": [ 448, 70, 461, 80 ], "score": 0.36, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 461, 69, 505, 82 ], "score": 1.0, "content": "represents", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 79, 506, 93 ], "spans": [ { "bbox": [ 105, 79, 506, 93 ], "score": 1.0, "content": "the similarity score of independent generation, and the right it represents the case of joint generation.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 91, 363, 104 ], "spans": [ { "bbox": [ 105, 91, 363, 104 ], "score": 1.0, "content": "Jointly generated outputs consistently show stronger coherence.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 109, 108, 500, 242 ], "group_id": 0, "lines": [ { "bbox": [ 109, 108, 500, 242 ], "spans": [ { "bbox": [ 109, 108, 500, 242 ], "score": 0.983, "html": "
InputsSIM-ITSIM-ATSIM-VTSIM-VA
Two Joint Outputs
Audio → Image+Text0.251 / 0.260
Image→Audio+Text■0.244 / 0.256
Text →Video+Audio0.240 / 0.255
Audio →Video+Text0.256 / 0.261
Three Joint Outputs
Text-→ Video+Image+Audio 0.256/0.270 0.240/0.2570.240 / 0.257
Multi-Inputs-Outputs
Text+Image -→ Video+Audio0.247 / 0.259
", "type": "table", "image_path": "7c6e18d97f95273627bd5cdb16fd3a210522268a86a54bd9275114265361cbb0.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 109, 108, 500, 152.66666666666666 ], "spans": [], "index": 3 }, { "bbox": [ 109, 152.66666666666666, 500, 197.33333333333331 ], "spans": [], "index": 4 }, { "bbox": [ 109, 197.33333333333331, 500, 241.99999999999997 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 263, 505, 308 ], "lines": [ { "bbox": [ 105, 264, 505, 277 ], "spans": [ { "bbox": [ 105, 264, 133, 277 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 264, 142, 274 ], "score": 0.58, "content": "A", "type": "inline_equation" }, { "bbox": [ 142, 264, 146, 277 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 146, 264, 155, 274 ], "score": 0.62, "content": "B", "type": "inline_equation" }, { "bbox": [ 156, 264, 293, 277 ], "score": 1.0, "content": "are the generated modalities, and", "type": "text" }, { "bbox": [ 293, 264, 308, 275 ], "score": 0.9, "content": "C _ { A }", "type": "inline_equation" }, { "bbox": [ 308, 264, 326, 277 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 327, 264, 342, 275 ], "score": 0.9, "content": "C _ { B }", "type": "inline_equation" }, { "bbox": [ 342, 264, 478, 277 ], "score": 1.0, "content": "are aligned encoders that project", "type": "text" }, { "bbox": [ 478, 264, 487, 274 ], "score": 0.7, "content": "A", "type": "inline_equation" }, { "bbox": [ 487, 264, 505, 277 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 275, 506, 288 ], "spans": [ { "bbox": [ 107, 276, 116, 285 ], "score": 0.79, "content": "B", "type": "inline_equation" }, { "bbox": [ 116, 275, 506, 288 ], "score": 1.0, "content": "to the same space. We use the prompt encoder as described in Section 3.2. This metric aims to", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "compute the cosine similarity of the embedding of two modalities using contrastive learned prompt", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 297, 493, 309 ], "spans": [ { "bbox": [ 106, 297, 493, 309 ], "score": 1.0, "content": "encoders. Thus, the higher the metric, the more aligned and similar the generated modalities are.", "type": "text" } ], "index": 9 } ], "index": 7.5, "bbox_fs": [ 105, 264, 506, 309 ] }, { "type": "text", "bbox": [ 107, 313, 505, 390 ], "lines": [ { "bbox": [ 105, 313, 505, 326 ], "spans": [ { "bbox": [ 105, 313, 441, 326 ], "score": 1.0, "content": "To demonstrate the effectiveness of joint generation, assume the prompt modality is", "type": "text" }, { "bbox": [ 442, 314, 450, 323 ], "score": 0.8, "content": "P", "type": "inline_equation" }, { "bbox": [ 451, 313, 505, 326 ], "score": 1.0, "content": ", we compare", "type": "text" } ], "index": 10 }, { "bbox": [ 107, 323, 505, 337 ], "spans": [ { "bbox": [ 107, 324, 154, 336 ], "score": 0.91, "content": "\\mathrm { S I M } ( A , B )", "type": "inline_equation" }, { "bbox": [ 154, 323, 168, 337 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 168, 324, 177, 334 ], "score": 0.71, "content": "A", "type": "inline_equation" }, { "bbox": [ 178, 323, 197, 337 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 198, 324, 207, 334 ], "score": 0.77, "content": "B", "type": "inline_equation" }, { "bbox": [ 208, 323, 369, 337 ], "score": 1.0, "content": "generated separately vs. jointly, i.e.,", "type": "text" }, { "bbox": [ 369, 324, 453, 335 ], "score": 0.75, "content": "\\{ P \\ { \\overset { - } { \\to } } \\ A , \\ P \\ { \\overset { - } { \\to } } \\ B \\}", "type": "inline_equation" }, { "bbox": [ 454, 323, 477, 337 ], "score": 1.0, "content": "vs.", "type": "text" }, { "bbox": [ 477, 324, 505, 335 ], "score": 0.81, "content": "\\{ P ", "type": "inline_equation" } ], "index": 11 }, { "bbox": [ 107, 334, 506, 348 ], "spans": [ { "bbox": [ 107, 336, 140, 347 ], "score": 0.9, "content": "A + B \\}", "type": "inline_equation" }, { "bbox": [ 140, 334, 506, 348 ], "score": 1.0, "content": ". The benchmark is the validation set of AudioCaps [24]. We test on the following settings,", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 347, 505, 358 ], "spans": [ { "bbox": [ 106, 347, 131, 358 ], "score": 1.0, "content": "audio", "type": "text" }, { "bbox": [ 131, 348, 142, 356 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 143, 347, 219, 358 ], "score": 1.0, "content": "image+text, image", "type": "text" }, { "bbox": [ 219, 348, 231, 356 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 231, 347, 311, 358 ], "score": 1.0, "content": "audio+text, and text", "type": "text" }, { "bbox": [ 312, 348, 321, 356 ], "score": 0.75, "content": "", "type": "inline_equation" }, { "bbox": [ 321, 347, 403, 358 ], "score": 1.0, "content": "video+audio, image", "type": "text" }, { "bbox": [ 403, 347, 414, 356 ], "score": 0.79, "content": "", "type": "inline_equation" }, { "bbox": [ 415, 347, 493, 358 ], "score": 1.0, "content": "video+audio. audio", "type": "text" }, { "bbox": [ 493, 347, 505, 356 ], "score": 0.8, "content": "", "type": "inline_equation" } ], "index": 13 }, { "bbox": [ 105, 357, 505, 369 ], "spans": [ { "bbox": [ 105, 357, 177, 369 ], "score": 1.0, "content": "video+text, audio", "type": "text" }, { "bbox": [ 177, 358, 189, 367 ], "score": 0.84, "content": "", "type": "inline_equation" }, { "bbox": [ 189, 357, 284, 369 ], "score": 1.0, "content": "text+video+image, text", "type": "text" }, { "bbox": [ 284, 358, 296, 367 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 357, 505, 369 ], "score": 1.0, "content": "video+image+audio, where the image prompt is the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 368, 505, 380 ], "spans": [ { "bbox": [ 105, 368, 505, 380 ], "score": 1.0, "content": "middle frame of the video clip. As shown in Table 11, joint generation (similarity shown on the right", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 379, 445, 391 ], "spans": [ { "bbox": [ 105, 379, 136, 391 ], "score": 1.0, "content": "side of", "type": "text" }, { "bbox": [ 136, 379, 148, 389 ], "score": 0.6, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 149, 379, 426, 391 ], "score": 1.0, "content": ") consistently outperforms independent generation (on the left side of", "type": "text" }, { "bbox": [ 427, 379, 438, 389 ], "score": 0.49, "content": "\" / \"", "type": "inline_equation" }, { "bbox": [ 439, 379, 445, 391 ], "score": 1.0, "content": ").", "type": "text" } ], "index": 16 } ], "index": 13, "bbox_fs": [ 105, 313, 506, 391 ] }, { "type": "title", "bbox": [ 107, 407, 183, 420 ], "lines": [ { "bbox": [ 104, 405, 185, 423 ], "spans": [ { "bbox": [ 104, 405, 185, 423 ], "score": 1.0, "content": "6 Conclusion", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 433, 505, 521 ], "lines": [ { "bbox": [ 105, 433, 506, 446 ], "spans": [ { "bbox": [ 105, 433, 506, 446 ], "score": 1.0, "content": "In this paper, we present Composable Diffusion (CoDi), a groundbreaking model in multimodal", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 444, 507, 458 ], "spans": [ { "bbox": [ 104, 444, 507, 458 ], "score": 1.0, "content": "generation that is capable of processing and simultaneously generating modalities across text, image,", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 456, 506, 468 ], "spans": [ { "bbox": [ 106, 456, 506, 468 ], "score": 1.0, "content": "video, and audio. Our approach enables the synergistic generation of high-quality and coherent", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 466, 506, 479 ], "spans": [ { "bbox": [ 105, 466, 506, 479 ], "score": 1.0, "content": "outputs spanning various modalities, from assorted combinations of input modalities. Through", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 477, 505, 490 ], "spans": [ { "bbox": [ 105, 477, 505, 490 ], "score": 1.0, "content": "extensive experiments, we demonstrate CoDi’s remarkable capabilities in flexibly generating single", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 488, 505, 500 ], "spans": [ { "bbox": [ 106, 488, 505, 500 ], "score": 1.0, "content": "or multiple modalities from a wide range of inputs. Our work marks a significant step towards", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 499, 506, 512 ], "spans": [ { "bbox": [ 105, 499, 506, 512 ], "score": 1.0, "content": "more engaging and holistic human-computer interactions, establishing a solid foundation for future", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 510, 305, 522 ], "spans": [ { "bbox": [ 106, 510, 305, 522 ], "score": 1.0, "content": "investigations in generative artificial intelligence.", "type": "text" } ], "index": 25 } ], "index": 21.5, "bbox_fs": [ 104, 433, 507, 522 ] }, { "type": "text", "bbox": [ 107, 526, 389, 538 ], "lines": [ { "bbox": [ 105, 525, 390, 540 ], "spans": [ { "bbox": [ 105, 525, 390, 540 ], "score": 1.0, "content": "Limitations & Broader Impacts. See Appendix D for the discussion.", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 105, 525, 390, 540 ] }, { "type": "title", "bbox": [ 108, 555, 202, 568 ], "lines": [ { "bbox": [ 106, 554, 204, 571 ], "spans": [ { "bbox": [ 106, 554, 204, 571 ], "score": 1.0, "content": "Acknowledgement", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 108, 581, 505, 614 ], "lines": [ { "bbox": [ 105, 579, 505, 594 ], "spans": [ { "bbox": [ 105, 579, 505, 594 ], "score": 1.0, "content": "We would like to thank Bei Liu for HD-VILA-100M data support. We also thank Shi Dong, Mahmoud", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 591, 505, 604 ], "spans": [ { "bbox": [ 106, 591, 505, 604 ], "score": 1.0, "content": "Khademi, Junheng Hao, Yuwei Fang, Yichong Xu and Azure Cognitive Services Research team", "type": "text" } ], "index": 29 }, { "bbox": [ 107, 603, 221, 614 ], "spans": [ { "bbox": [ 107, 603, 221, 614 ], "score": 1.0, "content": "members for their feedback.", "type": "text" } ], "index": 30 } ], "index": 29, "bbox_fs": [ 105, 579, 505, 614 ] }, { "type": "title", "bbox": [ 107, 631, 163, 643 ], "lines": [ { "bbox": [ 106, 630, 165, 645 ], "spans": [ { "bbox": [ 106, 630, 165, 645 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "list", "bbox": [ 109, 649, 506, 722 ], "lines": [ { "bbox": [ 109, 648, 507, 663 ], "spans": [ { "bbox": [ 109, 648, 507, 663 ], "score": 1.0, "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc,", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 125, 659, 506, 672 ], "spans": [ { "bbox": [ 125, 659, 506, 672 ], "score": 1.0, "content": "Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for", "type": "text" } ], "index": 33 }, { "bbox": [ 126, 670, 483, 682 ], "spans": [ { "bbox": [ 126, 670, 483, 682 ], "score": 1.0, "content": "few-shot learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022. 3", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 110, 691, 506, 704 ], "spans": [ { "bbox": [ 110, 691, 506, 704 ], "score": 1.0, "content": "[2] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift:", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 125, 702, 506, 713 ], "spans": [ { "bbox": [ 125, 702, 506, 713 ], "score": 1.0, "content": "Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477,", "type": "text" } ], "index": 36 }, { "bbox": [ 126, 711, 185, 722 ], "spans": [ { "bbox": [ 126, 711, 185, 722 ], "score": 1.0, "content": "2023. 5, 15, 16", "type": "text" } ], "index": 37, "is_list_end_line": true } ], "index": 34.5, "bbox_fs": [ 109, 648, 507, 722 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 109, 73, 504, 94 ], "lines": [ { "bbox": [ 109, 72, 505, 86 ], "spans": [ { "bbox": [ 109, 72, 505, 86 ], "score": 1.0, "content": "[3] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from", "type": "text" } ], "index": 0 }, { "bbox": [ 125, 82, 430, 95 ], "spans": [ { "bbox": [ 125, 82, 430, 95 ], "score": 1.0, "content": "unlabeled video. Advances in neural information processing systems, 29, 2016. 6, 7", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 110, 102, 502, 133 ], "lines": [ { "bbox": [ 110, 101, 505, 115 ], "spans": [ { "bbox": [ 110, 101, 505, 115 ], "score": 1.0, "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 505, 123 ], "spans": [ { "bbox": [ 125, 111, 505, 123 ], "score": 1.0, "content": "encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer", "type": "text" } ], "index": 3 }, { "bbox": [ 126, 122, 274, 133 ], "spans": [ { "bbox": [ 126, 122, 274, 133 ], "score": 1.0, "content": "Vision, pages 1728–1738, 2021. 6, 7, 16", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "text", "bbox": [ 109, 140, 503, 171 ], "lines": [ { "bbox": [ 109, 140, 505, 153 ], "spans": [ { "bbox": [ 109, 140, 505, 153 ], "score": 1.0, "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 151, 505, 162 ], "spans": [ { "bbox": [ 126, 151, 505, 162 ], "score": 1.0, "content": "Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. arXiv", "type": "text" } ], "index": 6 }, { "bbox": [ 124, 161, 259, 171 ], "spans": [ { "bbox": [ 124, 161, 259, 171 ], "score": 1.0, "content": "preprint arXiv:2304.08818, 2023. 7", "type": "text" } ], "index": 7 } ], "index": 6 }, { "type": "text", "bbox": [ 109, 179, 504, 210 ], "lines": [ { "bbox": [ 110, 179, 506, 192 ], "spans": [ { "bbox": [ 110, 179, 506, 192 ], "score": 1.0, "content": "[6] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", "type": "text" } ], "index": 8 }, { "bbox": [ 125, 189, 505, 201 ], "spans": [ { "bbox": [ 125, 189, 505, 201 ], "score": 1.0, "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early", "type": "text" } ], "index": 9 }, { "bbox": [ 125, 199, 370, 211 ], "spans": [ { "bbox": [ 125, 199, 370, 211 ], "score": 1.0, "content": "experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 2", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "text", "bbox": [ 109, 218, 504, 249 ], "lines": [ { "bbox": [ 109, 217, 506, 230 ], "spans": [ { "bbox": [ 109, 217, 506, 230 ], "score": 1.0, "content": "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor:", "type": "text" } ], "index": 11 }, { "bbox": [ 126, 228, 506, 239 ], "spans": [ { "bbox": [ 126, 228, 506, 239 ], "score": 1.0, "content": "Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345,", "type": "text" } ], "index": 12 }, { "bbox": [ 125, 237, 158, 249 ], "spans": [ { "bbox": [ 125, 237, 158, 249 ], "score": 1.0, "content": "2023. 7", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 109, 256, 503, 278 ], "lines": [ { "bbox": [ 110, 257, 504, 268 ], "spans": [ { "bbox": [ 110, 257, 504, 268 ], "score": 1.0, "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation.", "type": "text" } ], "index": 14 }, { "bbox": [ 125, 267, 439, 278 ], "spans": [ { "bbox": [ 125, 267, 439, 278 ], "score": 1.0, "content": "In International Conference on Machine Learning, pages 1931–1942. PMLR, 2021. 3", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 109, 285, 504, 307 ], "lines": [ { "bbox": [ 109, 285, 505, 297 ], "spans": [ { "bbox": [ 109, 285, 505, 297 ], "score": 1.0, "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirec-", "type": "text" } ], "index": 16 }, { "bbox": [ 125, 295, 455, 307 ], "spans": [ { "bbox": [ 125, 295, 455, 307 ], "score": 1.0, "content": "tional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "text", "bbox": [ 106, 313, 504, 345 ], "lines": [ { "bbox": [ 105, 314, 506, 326 ], "spans": [ { "bbox": [ 105, 314, 506, 326 ], "score": 1.0, "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou", "type": "text" } ], "index": 18 }, { "bbox": [ 126, 324, 505, 337 ], "spans": [ { "bbox": [ 126, 324, 505, 337 ], "score": 1.0, "content": "Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. arXiv", "type": "text" } ], "index": 19 }, { "bbox": [ 124, 335, 258, 346 ], "spans": [ { "bbox": [ 124, 335, 258, 346 ], "score": 1.0, "content": "preprint arXiv:2105.13290, 2021. 7", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 106, 353, 503, 394 ], "lines": [ { "bbox": [ 106, 353, 505, 365 ], "spans": [ { "bbox": [ 106, 353, 505, 365 ], "score": 1.0, "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 363, 505, 375 ], "spans": [ { "bbox": [ 126, 363, 505, 375 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 372, 505, 385 ], "spans": [ { "bbox": [ 126, 372, 505, 385 ], "score": 1.0, "content": "16x16 words: Transformers for image recognition at scale. In Proceedings of the IEEE/CVF Conference", "type": "text" } ], "index": 23 }, { "bbox": [ 126, 383, 399, 395 ], "spans": [ { "bbox": [ 126, 383, 399, 395 ], "score": 1.0, "content": "on Computer Vision and Pattern Recognition, pages 10687–10696, 2021. 3", "type": "text" } ], "index": 24 } ], "index": 22.5 }, { "type": "text", "bbox": [ 106, 401, 504, 423 ], "lines": [ { "bbox": [ 106, 401, 505, 413 ], "spans": [ { "bbox": [ 106, 401, 505, 413 ], "score": 1.0, "content": "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap: Learning audio", "type": "text" } ], "index": 25 }, { "bbox": [ 125, 411, 443, 423 ], "spans": [ { "bbox": [ 125, 411, 443, 423 ], "score": 1.0, "content": "concepts from natural language supervision. arXiv preprint arXiv:2206.04769, 2022. 3", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 106, 430, 505, 461 ], "lines": [ { "bbox": [ 105, 430, 507, 442 ], "spans": [ { "bbox": [ 105, 430, 507, 442 ], "score": 1.0, "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis.", "type": "text" } ], "index": 27 }, { "bbox": [ 126, 441, 506, 451 ], "spans": [ { "bbox": [ 126, 441, 506, 451 ], "score": 1.0, "content": "Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011,", "type": "text" } ], "index": 28 }, { "bbox": [ 125, 450, 167, 462 ], "spans": [ { "bbox": [ 125, 450, 167, 462 ], "score": 1.0, "content": "2023. 2, 5", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 105, 469, 505, 500 ], "lines": [ { "bbox": [ 105, 468, 506, 481 ], "spans": [ { "bbox": [ 105, 468, 506, 481 ], "score": 1.0, "content": "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis.", "type": "text" } ], "index": 30 }, { "bbox": [ 126, 479, 506, 491 ], "spans": [ { "bbox": [ 126, 479, 506, 491 ], "score": 1.0, "content": "Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011,", "type": "text" } ], "index": 31 }, { "bbox": [ 124, 489, 158, 500 ], "spans": [ { "bbox": [ 124, 489, 158, 500 ], "score": 1.0, "content": "2023. 3", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 105, 507, 504, 539 ], "lines": [ { "bbox": [ 106, 507, 507, 520 ], "spans": [ { "bbox": [ 106, 507, 507, 520 ], "score": 1.0, "content": "[15] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene:", "type": "text" } ], "index": 33 }, { "bbox": [ 126, 518, 506, 529 ], "spans": [ { "bbox": [ 126, 518, 506, 529 ], "score": 1.0, "content": "Scene-based text-to-image generation with human priors. In Computer Vision–ECCV 2022: 17th European", "type": "text" } ], "index": 34 }, { "bbox": [ 126, 528, 506, 540 ], "spans": [ { "bbox": [ 126, 528, 506, 540 ], "score": 1.0, "content": "Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 89–106. Springer, 2022. 7", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "text", "bbox": [ 106, 546, 507, 587 ], "lines": [ { "bbox": [ 106, 545, 506, 558 ], "spans": [ { "bbox": [ 106, 545, 506, 558 ], "score": 1.0, "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore,", "type": "text" } ], "index": 36 }, { "bbox": [ 125, 556, 506, 568 ], "spans": [ { "bbox": [ 125, 556, 506, 568 ], "score": 1.0, "content": "Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In", "type": "text" } ], "index": 37 }, { "bbox": [ 125, 565, 507, 579 ], "spans": [ { "bbox": [ 125, 565, 507, 579 ], "score": 1.0, "content": "2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780.", "type": "text" } ], "index": 38 }, { "bbox": [ 126, 576, 182, 587 ], "spans": [ { "bbox": [ 126, 576, 182, 587 ], "score": 1.0, "content": "IEEE, 2017. 6", "type": "text" } ], "index": 39 } ], "index": 37.5 }, { "type": "text", "bbox": [ 105, 595, 504, 616 ], "lines": [ { "bbox": [ 106, 595, 505, 606 ], "spans": [ { "bbox": [ 106, 595, 505, 606 ], "score": 1.0, "content": "[17] Félix Gontier, Romain Serizel, and Christophe Cerisara. Automated audio captioning by fine-tuning bart", "type": "text" } ], "index": 40 }, { "bbox": [ 126, 605, 505, 616 ], "spans": [ { "bbox": [ 126, 605, 505, 616 ], "score": 1.0, "content": "with audioset tags. In Detection and Classification of Acoustic Scenes and Events-DCASE 2021, 2021. 7", "type": "text" } ], "index": 41 } ], "index": 40.5 }, { "type": "text", "bbox": [ 106, 623, 505, 654 ], "lines": [ { "bbox": [ 106, 624, 506, 636 ], "spans": [ { "bbox": [ 106, 624, 506, 636 ], "score": 1.0, "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", "type": "text" } ], "index": 42 }, { "bbox": [ 125, 634, 506, 646 ], "spans": [ { "bbox": [ 125, 634, 506, 646 ], "score": 1.0, "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.", "type": "text" } ], "index": 43 }, { "bbox": [ 125, 644, 137, 654 ], "spans": [ { "bbox": [ 125, 644, 137, 654 ], "score": 1.0, "content": "15", "type": "text" } ], "index": 44 } ], "index": 43 }, { "type": "text", "bbox": [ 106, 662, 504, 694 ], "lines": [ { "bbox": [ 105, 662, 505, 675 ], "spans": [ { "bbox": [ 105, 662, 505, 675 ], "score": 1.0, "content": "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P", "type": "text" } ], "index": 45 }, { "bbox": [ 126, 672, 505, 684 ], "spans": [ { "bbox": [ 126, 672, 505, 684 ], "score": 1.0, "content": "Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video", "type": "text" } ], "index": 46 }, { "bbox": [ 125, 682, 414, 694 ], "spans": [ { "bbox": [ 125, 682, 414, 694 ], "score": 1.0, "content": "generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3", "type": "text" } ], "index": 47 } ], "index": 46 }, { "type": "text", "bbox": [ 104, 701, 505, 722 ], "lines": [ { "bbox": [ 106, 701, 505, 713 ], "spans": [ { "bbox": [ 106, 701, 505, 713 ], "score": 1.0, "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural", "type": "text" } ], "index": 48 }, { "bbox": [ 126, 711, 343, 722 ], "spans": [ { "bbox": [ 126, 711, 343, 722 ], "score": 1.0, "content": "Information Processing Systems, 33:6840–6851, 2020. 3, 4", "type": "text" } ], "index": 49 } ], "index": 48.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 310, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 755 ], "spans": [ { "bbox": [ 299, 740, 312, 755 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 109, 73, 504, 94 ], "lines": [ { "bbox": [ 109, 72, 505, 86 ], "spans": [ { "bbox": [ 109, 72, 505, 86 ], "score": 1.0, "content": "[3] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from", "type": "text" } ], "index": 0 }, { "bbox": [ 125, 82, 430, 95 ], "spans": [ { "bbox": [ 125, 82, 430, 95 ], "score": 1.0, "content": "unlabeled video. Advances in neural information processing systems, 29, 2016. 6, 7", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 109, 72, 505, 95 ] }, { "type": "text", "bbox": [ 110, 102, 502, 133 ], "lines": [ { "bbox": [ 110, 101, 505, 115 ], "spans": [ { "bbox": [ 110, 101, 505, 115 ], "score": 1.0, "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 505, 123 ], "spans": [ { "bbox": [ 125, 111, 505, 123 ], "score": 1.0, "content": "encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer", "type": "text" } ], "index": 3 }, { "bbox": [ 126, 122, 274, 133 ], "spans": [ { "bbox": [ 126, 122, 274, 133 ], "score": 1.0, "content": "Vision, pages 1728–1738, 2021. 6, 7, 16", "type": "text" } ], "index": 4 } ], "index": 3, "bbox_fs": [ 110, 101, 505, 133 ] }, { "type": "text", "bbox": [ 109, 140, 503, 171 ], "lines": [ { "bbox": [ 109, 140, 505, 153 ], "spans": [ { "bbox": [ 109, 140, 505, 153 ], "score": 1.0, "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 151, 505, 162 ], "spans": [ { "bbox": [ 126, 151, 505, 162 ], "score": 1.0, "content": "Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. arXiv", "type": "text" } ], "index": 6 }, { "bbox": [ 124, 161, 259, 171 ], "spans": [ { "bbox": [ 124, 161, 259, 171 ], "score": 1.0, "content": "preprint arXiv:2304.08818, 2023. 7", "type": "text" } ], "index": 7 } ], "index": 6, "bbox_fs": [ 109, 140, 505, 171 ] }, { "type": "text", "bbox": [ 109, 179, 504, 210 ], "lines": [ { "bbox": [ 110, 179, 506, 192 ], "spans": [ { "bbox": [ 110, 179, 506, 192 ], "score": 1.0, "content": "[6] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", "type": "text" } ], "index": 8 }, { "bbox": [ 125, 189, 505, 201 ], "spans": [ { "bbox": [ 125, 189, 505, 201 ], "score": 1.0, "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early", "type": "text" } ], "index": 9 }, { "bbox": [ 125, 199, 370, 211 ], "spans": [ { "bbox": [ 125, 199, 370, 211 ], "score": 1.0, "content": "experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 2", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 110, 179, 506, 211 ] }, { "type": "text", "bbox": [ 109, 218, 504, 249 ], "lines": [ { "bbox": [ 109, 217, 506, 230 ], "spans": [ { "bbox": [ 109, 217, 506, 230 ], "score": 1.0, "content": "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor:", "type": "text" } ], "index": 11 }, { "bbox": [ 126, 228, 506, 239 ], "spans": [ { "bbox": [ 126, 228, 506, 239 ], "score": 1.0, "content": "Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345,", "type": "text" } ], "index": 12 }, { "bbox": [ 125, 237, 158, 249 ], "spans": [ { "bbox": [ 125, 237, 158, 249 ], "score": 1.0, "content": "2023. 7", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 109, 217, 506, 249 ] }, { "type": "text", "bbox": [ 109, 256, 503, 278 ], "lines": [ { "bbox": [ 110, 257, 504, 268 ], "spans": [ { "bbox": [ 110, 257, 504, 268 ], "score": 1.0, "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation.", "type": "text" } ], "index": 14 }, { "bbox": [ 125, 267, 439, 278 ], "spans": [ { "bbox": [ 125, 267, 439, 278 ], "score": 1.0, "content": "In International Conference on Machine Learning, pages 1931–1942. PMLR, 2021. 3", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 110, 257, 504, 278 ] }, { "type": "text", "bbox": [ 109, 285, 504, 307 ], "lines": [ { "bbox": [ 109, 285, 505, 297 ], "spans": [ { "bbox": [ 109, 285, 505, 297 ], "score": 1.0, "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirec-", "type": "text" } ], "index": 16 }, { "bbox": [ 125, 295, 455, 307 ], "spans": [ { "bbox": [ 125, 295, 455, 307 ], "score": 1.0, "content": "tional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 109, 285, 505, 307 ] }, { "type": "text", "bbox": [ 106, 313, 504, 345 ], "lines": [ { "bbox": [ 105, 314, 506, 326 ], "spans": [ { "bbox": [ 105, 314, 506, 326 ], "score": 1.0, "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou", "type": "text" } ], "index": 18 }, { "bbox": [ 126, 324, 505, 337 ], "spans": [ { "bbox": [ 126, 324, 505, 337 ], "score": 1.0, "content": "Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. arXiv", "type": "text" } ], "index": 19 }, { "bbox": [ 124, 335, 258, 346 ], "spans": [ { "bbox": [ 124, 335, 258, 346 ], "score": 1.0, "content": "preprint arXiv:2105.13290, 2021. 7", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 105, 314, 506, 346 ] }, { "type": "text", "bbox": [ 106, 353, 503, 394 ], "lines": [ { "bbox": [ 106, 353, 505, 365 ], "spans": [ { "bbox": [ 106, 353, 505, 365 ], "score": 1.0, "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 363, 505, 375 ], "spans": [ { "bbox": [ 126, 363, 505, 375 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 372, 505, 385 ], "spans": [ { "bbox": [ 126, 372, 505, 385 ], "score": 1.0, "content": "16x16 words: Transformers for image recognition at scale. In Proceedings of the IEEE/CVF Conference", "type": "text" } ], "index": 23 }, { "bbox": [ 126, 383, 399, 395 ], "spans": [ { "bbox": [ 126, 383, 399, 395 ], "score": 1.0, "content": "on Computer Vision and Pattern Recognition, pages 10687–10696, 2021. 3", "type": "text" } ], "index": 24 } ], "index": 22.5, "bbox_fs": [ 106, 353, 505, 395 ] }, { "type": "text", "bbox": [ 106, 401, 504, 423 ], "lines": [ { "bbox": [ 106, 401, 505, 413 ], "spans": [ { "bbox": [ 106, 401, 505, 413 ], "score": 1.0, "content": "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap: Learning audio", "type": "text" } ], "index": 25 }, { "bbox": [ 125, 411, 443, 423 ], "spans": [ { "bbox": [ 125, 411, 443, 423 ], "score": 1.0, "content": "concepts from natural language supervision. arXiv preprint arXiv:2206.04769, 2022. 3", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 106, 401, 505, 423 ] }, { "type": "text", "bbox": [ 106, 430, 505, 461 ], "lines": [ { "bbox": [ 105, 430, 507, 442 ], "spans": [ { "bbox": [ 105, 430, 507, 442 ], "score": 1.0, "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis.", "type": "text" } ], "index": 27 }, { "bbox": [ 126, 441, 506, 451 ], "spans": [ { "bbox": [ 126, 441, 506, 451 ], "score": 1.0, "content": "Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011,", "type": "text" } ], "index": 28 }, { "bbox": [ 125, 450, 167, 462 ], "spans": [ { "bbox": [ 125, 450, 167, 462 ], "score": 1.0, "content": "2023. 2, 5", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 105, 430, 507, 462 ] }, { "type": "text", "bbox": [ 105, 469, 505, 500 ], "lines": [ { "bbox": [ 105, 468, 506, 481 ], "spans": [ { "bbox": [ 105, 468, 506, 481 ], "score": 1.0, "content": "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis.", "type": "text" } ], "index": 30 }, { "bbox": [ 126, 479, 506, 491 ], "spans": [ { "bbox": [ 126, 479, 506, 491 ], "score": 1.0, "content": "Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011,", "type": "text" } ], "index": 31 }, { "bbox": [ 124, 489, 158, 500 ], "spans": [ { "bbox": [ 124, 489, 158, 500 ], "score": 1.0, "content": "2023. 3", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 105, 468, 506, 500 ] }, { "type": "text", "bbox": [ 105, 507, 504, 539 ], "lines": [ { "bbox": [ 106, 507, 507, 520 ], "spans": [ { "bbox": [ 106, 507, 507, 520 ], "score": 1.0, "content": "[15] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene:", "type": "text" } ], "index": 33 }, { "bbox": [ 126, 518, 506, 529 ], "spans": [ { "bbox": [ 126, 518, 506, 529 ], "score": 1.0, "content": "Scene-based text-to-image generation with human priors. In Computer Vision–ECCV 2022: 17th European", "type": "text" } ], "index": 34 }, { "bbox": [ 126, 528, 506, 540 ], "spans": [ { "bbox": [ 126, 528, 506, 540 ], "score": 1.0, "content": "Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 89–106. Springer, 2022. 7", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 545, 506, 558 ], "spans": [ { "bbox": [ 106, 545, 506, 558 ], "score": 1.0, "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore,", "type": "text" } ], "index": 36 }, { "bbox": [ 125, 556, 506, 568 ], "spans": [ { "bbox": [ 125, 556, 506, 568 ], "score": 1.0, "content": "Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In", "type": "text" } ], "index": 37 }, { "bbox": [ 125, 565, 507, 579 ], "spans": [ { "bbox": [ 125, 565, 507, 579 ], "score": 1.0, "content": "2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780.", "type": "text" } ], "index": 38 }, { "bbox": [ 126, 576, 182, 587 ], "spans": [ { "bbox": [ 126, 576, 182, 587 ], "score": 1.0, "content": "IEEE, 2017. 6", "type": "text" } ], "index": 39 } ], "index": 34, "bbox_fs": [ 106, 507, 507, 540 ] }, { "type": "text", "bbox": [ 106, 546, 507, 587 ], "lines": [], "index": 37.5, "bbox_fs": [ 106, 545, 507, 587 ], "lines_deleted": true }, { "type": "text", "bbox": [ 105, 595, 504, 616 ], "lines": [ { "bbox": [ 106, 595, 505, 606 ], "spans": [ { "bbox": [ 106, 595, 505, 606 ], "score": 1.0, "content": "[17] Félix Gontier, Romain Serizel, and Christophe Cerisara. Automated audio captioning by fine-tuning bart", "type": "text" } ], "index": 40 }, { "bbox": [ 126, 605, 505, 616 ], "spans": [ { "bbox": [ 126, 605, 505, 616 ], "score": 1.0, "content": "with audioset tags. In Detection and Classification of Acoustic Scenes and Events-DCASE 2021, 2021. 7", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 624, 506, 636 ], "spans": [ { "bbox": [ 106, 624, 506, 636 ], "score": 1.0, "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", "type": "text" } ], "index": 42 }, { "bbox": [ 125, 634, 506, 646 ], "spans": [ { "bbox": [ 125, 634, 506, 646 ], "score": 1.0, "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.", "type": "text" } ], "index": 43 }, { "bbox": [ 125, 644, 137, 654 ], "spans": [ { "bbox": [ 125, 644, 137, 654 ], "score": 1.0, "content": "15", "type": "text" } ], "index": 44 } ], "index": 40.5, "bbox_fs": [ 106, 595, 505, 616 ] }, { "type": "text", "bbox": [ 106, 623, 505, 654 ], "lines": [], "index": 43, "bbox_fs": [ 106, 624, 506, 654 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 662, 504, 694 ], "lines": [ { "bbox": [ 105, 662, 505, 675 ], "spans": [ { "bbox": [ 105, 662, 505, 675 ], "score": 1.0, "content": "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P", "type": "text" } ], "index": 45 }, { "bbox": [ 126, 672, 505, 684 ], "spans": [ { "bbox": [ 126, 672, 505, 684 ], "score": 1.0, "content": "Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video", "type": "text" } ], "index": 46 }, { "bbox": [ 125, 682, 414, 694 ], "spans": [ { "bbox": [ 125, 682, 414, 694 ], "score": 1.0, "content": "generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3", "type": "text" } ], "index": 47 } ], "index": 46, "bbox_fs": [ 105, 662, 505, 694 ] }, { "type": "text", "bbox": [ 104, 701, 505, 722 ], "lines": [ { "bbox": [ 106, 701, 505, 713 ], "spans": [ { "bbox": [ 106, 701, 505, 713 ], "score": 1.0, "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural", "type": "text" } ], "index": 48 }, { "bbox": [ 126, 711, 343, 722 ], "spans": [ { "bbox": [ 126, 711, 343, 722 ], "score": 1.0, "content": "Information Processing Systems, 33:6840–6851, 2020. 3, 4", "type": "text" } ], "index": 49 } ], "index": 48.5, "bbox_fs": [ 106, 701, 505, 722 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 73, 505, 94 ], "lines": [ { "bbox": [ 106, 73, 507, 85 ], "spans": [ { "bbox": [ 106, 73, 507, 85 ], "score": 1.0, "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet.", "type": "text" } ], "index": 0 }, { "bbox": [ 126, 83, 383, 95 ], "spans": [ { "bbox": [ 126, 83, 383, 95 ], "score": 1.0, "content": "Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3, 15", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 106, 101, 503, 123 ], "lines": [ { "bbox": [ 106, 101, 505, 114 ], "spans": [ { "bbox": [ 106, 101, 505, 114 ], "score": 1.0, "content": "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 442, 123 ], "spans": [ { "bbox": [ 125, 111, 442, 123 ], "score": 1.0, "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 7", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 129, 503, 160 ], "lines": [ { "bbox": [ 105, 129, 505, 142 ], "spans": [ { "bbox": [ 105, 129, 505, 142 ], "score": 1.0, "content": "[23] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu,", "type": "text" } ], "index": 4 }, { "bbox": [ 125, 140, 505, 151 ], "spans": [ { "bbox": [ 125, 140, 505, 151 ], "score": 1.0, "content": "Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion", "type": "text" } ], "index": 5 }, { "bbox": [ 125, 149, 322, 161 ], "spans": [ { "bbox": [ 125, 149, 322, 161 ], "score": 1.0, "content": "models. arXiv preprint arXiv:2301.12661, 2023. 2, 3", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "text", "bbox": [ 106, 167, 505, 208 ], "lines": [ { "bbox": [ 105, 167, 505, 180 ], "spans": [ { "bbox": [ 105, 167, 505, 180 ], "score": 1.0, "content": "[24] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 178, 505, 189 ], "spans": [ { "bbox": [ 126, 178, 505, 189 ], "score": 1.0, "content": "for audios in the wild. In Proceedings of the 2019 Conference of the North American Chapter of the", "type": "text" } ], "index": 8 }, { "bbox": [ 126, 187, 505, 199 ], "spans": [ { "bbox": [ 126, 187, 505, 199 ], "score": 1.0, "content": "Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short", "type": "text" } ], "index": 9 }, { "bbox": [ 126, 198, 288, 209 ], "spans": [ { "bbox": [ 126, 198, 288, 209 ], "score": 1.0, "content": "Papers), pages 119–132, 2019. 6, 7, 8, 9, 10", "type": "text" } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 105, 216, 504, 246 ], "lines": [ { "bbox": [ 105, 215, 505, 228 ], "spans": [ { "bbox": [ 105, 215, 505, 228 ], "score": 1.0, "content": "[25] Eungbeom Kim, Jinhee Kim, Yoori Oh, Kyungsu Kim, Minju Park, Jaeheon Sim, Jinwoo Lee, and Kyogu", "type": "text" } ], "index": 11 }, { "bbox": [ 124, 225, 505, 237 ], "spans": [ { "bbox": [ 124, 225, 505, 237 ], "score": 1.0, "content": "Lee. Improving audio-language learning with mixgen and multi-level test-time augmentation. arXiv", "type": "text" } ], "index": 12 }, { "bbox": [ 125, 236, 259, 247 ], "spans": [ { "bbox": [ 125, 236, 259, 247 ], "score": 1.0, "content": "preprint arXiv:2210.17143, 2022. 7", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 253, 504, 274 ], "lines": [ { "bbox": [ 106, 252, 505, 266 ], "spans": [ { "bbox": [ 106, 252, 505, 266 ], "score": 1.0, "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", "type": "text" } ], "index": 14 }, { "bbox": [ 125, 264, 227, 275 ], "spans": [ { "bbox": [ 125, 264, 227, 275 ], "score": 1.0, "content": "arXiv:1412.6980, 2014. 15", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 107, 282, 505, 312 ], "lines": [ { "bbox": [ 105, 281, 506, 294 ], "spans": [ { "bbox": [ 105, 281, 506, 294 ], "score": 1.0, "content": "[27] Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. Hifi-gan: Generative adversarial networks for efficient", "type": "text" } ], "index": 16 }, { "bbox": [ 126, 291, 506, 303 ], "spans": [ { "bbox": [ 126, 291, 506, 303 ], "score": 1.0, "content": "and high fidelity speech synthesis. Advances in Neural Information Processing Systems, 33:17022–17033,", "type": "text" } ], "index": 17 }, { "bbox": [ 124, 300, 159, 314 ], "spans": [ { "bbox": [ 124, 300, 159, 314 ], "score": 1.0, "content": "2020. 5", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 320, 504, 351 ], "lines": [ { "bbox": [ 105, 319, 506, 334 ], "spans": [ { "bbox": [ 105, 319, 506, 334 ], "score": 1.0, "content": "[28] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song.", "type": "text" } ], "index": 19 }, { "bbox": [ 125, 330, 506, 342 ], "spans": [ { "bbox": [ 125, 330, 146, 342 ], "score": 1.0, "content": "Acav", "type": "text" }, { "bbox": [ 146, 330, 167, 340 ], "score": 0.27, "content": "1 0 0 \\mathrm { m }", "type": "inline_equation" }, { "bbox": [ 167, 330, 506, 342 ], "score": 1.0, "content": ": Automatic curation of large-scale datasets for audio-visual video representation learning. In", "type": "text" } ], "index": 20 }, { "bbox": [ 126, 340, 506, 352 ], "spans": [ { "bbox": [ 126, 340, 506, 352 ], "score": 1.0, "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10274–10284, 2021. 7", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 358, 505, 390 ], "lines": [ { "bbox": [ 106, 357, 506, 370 ], "spans": [ { "bbox": [ 106, 357, 506, 370 ], "score": 1.0, "content": "[29] Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. Optimus:", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 368, 506, 381 ], "spans": [ { "bbox": [ 126, 368, 506, 381 ], "score": 1.0, "content": "Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on", "type": "text" } ], "index": 23 }, { "bbox": [ 126, 378, 455, 390 ], "spans": [ { "bbox": [ 126, 378, 455, 390 ], "score": 1.0, "content": "Empirical Methods in Natural Language Processing (EMNLP), pages 4678–4699, 2020. 5", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "text", "bbox": [ 105, 396, 505, 417 ], "lines": [ { "bbox": [ 105, 394, 506, 410 ], "spans": [ { "bbox": [ 105, 394, 506, 410 ], "score": 1.0, "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training", "type": "text" } ], "index": 25 }, { "bbox": [ 125, 406, 483, 418 ], "spans": [ { "bbox": [ 125, 406, 483, 418 ], "score": 1.0, "content": "with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 7", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 105, 424, 506, 465 ], "lines": [ { "bbox": [ 106, 423, 505, 438 ], "spans": [ { "bbox": [ 106, 423, 505, 438 ], "score": 1.0, "content": "[31] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong", "type": "text" } ], "index": 27 }, { "bbox": [ 125, 434, 505, 446 ], "spans": [ { "bbox": [ 125, 434, 505, 446 ], "score": 1.0, "content": "Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In", "type": "text" } ], "index": 28 }, { "bbox": [ 126, 443, 506, 457 ], "spans": [ { "bbox": [ 126, 443, 506, 457 ], "score": 1.0, "content": "Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings,", "type": "text" } ], "index": 29 }, { "bbox": [ 125, 454, 300, 465 ], "spans": [ { "bbox": [ 125, 454, 300, 465 ], "score": 1.0, "content": "Part XXX 16, pages 121–137. Springer, 2020. 7", "type": "text" } ], "index": 30 } ], "index": 28.5 }, { "type": "text", "bbox": [ 106, 472, 504, 513 ], "lines": [ { "bbox": [ 106, 472, 506, 484 ], "spans": [ { "bbox": [ 106, 472, 506, 484 ], "score": 1.0, "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,", "type": "text" } ], "index": 31 }, { "bbox": [ 126, 483, 506, 494 ], "spans": [ { "bbox": [ 126, 483, 506, 494 ], "score": 1.0, "content": "and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014:", "type": "text" } ], "index": 32 }, { "bbox": [ 124, 491, 506, 505 ], "spans": [ { "bbox": [ 124, 491, 506, 505 ], "score": 1.0, "content": "13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages", "type": "text" } ], "index": 33 }, { "bbox": [ 125, 502, 227, 514 ], "spans": [ { "bbox": [ 125, 502, 227, 514 ], "score": 1.0, "content": "740–755. Springer, 2014. 7", "type": "text" } ], "index": 34 } ], "index": 32.5 }, { "type": "text", "bbox": [ 107, 520, 504, 551 ], "lines": [ { "bbox": [ 105, 520, 506, 533 ], "spans": [ { "bbox": [ 105, 520, 506, 533 ], "score": 1.0, "content": "[33] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and", "type": "text" } ], "index": 35 }, { "bbox": [ 125, 530, 506, 542 ], "spans": [ { "bbox": [ 125, 530, 506, 542 ], "score": 1.0, "content": "Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint", "type": "text" } ], "index": 36 }, { "bbox": [ 126, 541, 245, 552 ], "spans": [ { "bbox": [ 126, 541, 245, 552 ], "score": 1.0, "content": "arXiv:2301.12503, 2023. 2, 3, 5", "type": "text" } ], "index": 37 } ], "index": 36 }, { "type": "text", "bbox": [ 105, 558, 504, 580 ], "lines": [ { "bbox": [ 106, 559, 505, 570 ], "spans": [ { "bbox": [ 106, 559, 505, 570 ], "score": 1.0, "content": "[34] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-", "type": "text" } ], "index": 38 }, { "bbox": [ 126, 568, 473, 581 ], "spans": [ { "bbox": [ 126, 568, 473, 581 ], "score": 1.0, "content": "conditional diffusion networks for image captioning. arXiv preprint arXiv:2212.03099, 2022. 7", "type": "text" } ], "index": 39 } ], "index": 38.5 }, { "type": "text", "bbox": [ 106, 587, 504, 608 ], "lines": [ { "bbox": [ 106, 586, 505, 600 ], "spans": [ { "bbox": [ 106, 586, 505, 600 ], "score": 1.0, "content": "[35] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint", "type": "text" } ], "index": 40 }, { "bbox": [ 125, 596, 236, 608 ], "spans": [ { "bbox": [ 125, 596, 236, 608 ], "score": 1.0, "content": "arXiv:2111.09734, 2021. 3, 7", "type": "text" } ], "index": 41 } ], "index": 40.5 }, { "type": "text", "bbox": [ 106, 615, 504, 646 ], "lines": [ { "bbox": [ 106, 615, 505, 627 ], "spans": [ { "bbox": [ 106, 615, 505, 627 ], "score": 1.0, "content": "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya", "type": "text" } ], "index": 42 }, { "bbox": [ 125, 624, 505, 637 ], "spans": [ { "bbox": [ 125, 624, 505, 637 ], "score": 1.0, "content": "Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided", "type": "text" } ], "index": 43 }, { "bbox": [ 125, 634, 346, 646 ], "spans": [ { "bbox": [ 125, 634, 346, 646 ], "score": 1.0, "content": "diffusion models. arXiv preprint arXiv:2112.10741, 2021. 7", "type": "text" } ], "index": 44 } ], "index": 43 }, { "type": "text", "bbox": [ 105, 653, 504, 674 ], "lines": [ { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "[37] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Automatic prompt", "type": "text" } ], "index": 45 }, { "bbox": [ 125, 663, 481, 675 ], "spans": [ { "bbox": [ 125, 663, 481, 675 ], "score": 1.0, "content": "optimization with\" gradient descent\" and beam search. arXiv preprint arXiv:2305.03495, 2023. 2", "type": "text" } ], "index": 46 } ], "index": 45.5 }, { "type": "text", "bbox": [ 107, 681, 505, 722 ], "lines": [ { "bbox": [ 106, 681, 505, 693 ], "spans": [ { "bbox": [ 106, 681, 505, 693 ], "score": 1.0, "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", "type": "text" } ], "index": 47 }, { "bbox": [ 126, 691, 505, 703 ], "spans": [ { "bbox": [ 126, 691, 505, 703 ], "score": 1.0, "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from", "type": "text" } ], "index": 48 }, { "bbox": [ 126, 701, 506, 713 ], "spans": [ { "bbox": [ 126, 701, 506, 713 ], "score": 1.0, "content": "natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR,", "type": "text" } ], "index": 49 }, { "bbox": [ 125, 711, 167, 723 ], "spans": [ { "bbox": [ 125, 711, 167, 723 ], "score": 1.0, "content": "2021. 3, 4", "type": "text" } ], "index": 50 } ], "index": 48.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 73, 505, 94 ], "lines": [ { "bbox": [ 106, 73, 507, 85 ], "spans": [ { "bbox": [ 106, 73, 507, 85 ], "score": 1.0, "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet.", "type": "text" } ], "index": 0 }, { "bbox": [ 126, 83, 383, 95 ], "spans": [ { "bbox": [ 126, 83, 383, 95 ], "score": 1.0, "content": "Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3, 15", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 106, 73, 507, 95 ] }, { "type": "text", "bbox": [ 106, 101, 503, 123 ], "lines": [ { "bbox": [ 106, 101, 505, 114 ], "spans": [ { "bbox": [ 106, 101, 505, 114 ], "score": 1.0, "content": "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 442, 123 ], "spans": [ { "bbox": [ 125, 111, 442, 123 ], "score": 1.0, "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 7", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 106, 101, 505, 123 ] }, { "type": "text", "bbox": [ 107, 129, 503, 160 ], "lines": [ { "bbox": [ 105, 129, 505, 142 ], "spans": [ { "bbox": [ 105, 129, 505, 142 ], "score": 1.0, "content": "[23] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu,", "type": "text" } ], "index": 4 }, { "bbox": [ 125, 140, 505, 151 ], "spans": [ { "bbox": [ 125, 140, 505, 151 ], "score": 1.0, "content": "Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion", "type": "text" } ], "index": 5 }, { "bbox": [ 125, 149, 322, 161 ], "spans": [ { "bbox": [ 125, 149, 322, 161 ], "score": 1.0, "content": "models. arXiv preprint arXiv:2301.12661, 2023. 2, 3", "type": "text" } ], "index": 6 } ], "index": 5, "bbox_fs": [ 105, 129, 505, 161 ] }, { "type": "text", "bbox": [ 106, 167, 505, 208 ], "lines": [ { "bbox": [ 105, 167, 505, 180 ], "spans": [ { "bbox": [ 105, 167, 505, 180 ], "score": 1.0, "content": "[24] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 178, 505, 189 ], "spans": [ { "bbox": [ 126, 178, 505, 189 ], "score": 1.0, "content": "for audios in the wild. In Proceedings of the 2019 Conference of the North American Chapter of the", "type": "text" } ], "index": 8 }, { "bbox": [ 126, 187, 505, 199 ], "spans": [ { "bbox": [ 126, 187, 505, 199 ], "score": 1.0, "content": "Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short", "type": "text" } ], "index": 9 }, { "bbox": [ 126, 198, 288, 209 ], "spans": [ { "bbox": [ 126, 198, 288, 209 ], "score": 1.0, "content": "Papers), pages 119–132, 2019. 6, 7, 8, 9, 10", "type": "text" } ], "index": 10 } ], "index": 8.5, "bbox_fs": [ 105, 167, 505, 209 ] }, { "type": "text", "bbox": [ 105, 216, 504, 246 ], "lines": [ { "bbox": [ 105, 215, 505, 228 ], "spans": [ { "bbox": [ 105, 215, 505, 228 ], "score": 1.0, "content": "[25] Eungbeom Kim, Jinhee Kim, Yoori Oh, Kyungsu Kim, Minju Park, Jaeheon Sim, Jinwoo Lee, and Kyogu", "type": "text" } ], "index": 11 }, { "bbox": [ 124, 225, 505, 237 ], "spans": [ { "bbox": [ 124, 225, 505, 237 ], "score": 1.0, "content": "Lee. Improving audio-language learning with mixgen and multi-level test-time augmentation. arXiv", "type": "text" } ], "index": 12 }, { "bbox": [ 125, 236, 259, 247 ], "spans": [ { "bbox": [ 125, 236, 259, 247 ], "score": 1.0, "content": "preprint arXiv:2210.17143, 2022. 7", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 105, 215, 505, 247 ] }, { "type": "text", "bbox": [ 106, 253, 504, 274 ], "lines": [ { "bbox": [ 106, 252, 505, 266 ], "spans": [ { "bbox": [ 106, 252, 505, 266 ], "score": 1.0, "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", "type": "text" } ], "index": 14 }, { "bbox": [ 125, 264, 227, 275 ], "spans": [ { "bbox": [ 125, 264, 227, 275 ], "score": 1.0, "content": "arXiv:1412.6980, 2014. 15", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 106, 252, 505, 275 ] }, { "type": "text", "bbox": [ 107, 282, 505, 312 ], "lines": [ { "bbox": [ 105, 281, 506, 294 ], "spans": [ { "bbox": [ 105, 281, 506, 294 ], "score": 1.0, "content": "[27] Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. Hifi-gan: Generative adversarial networks for efficient", "type": "text" } ], "index": 16 }, { "bbox": [ 126, 291, 506, 303 ], "spans": [ { "bbox": [ 126, 291, 506, 303 ], "score": 1.0, "content": "and high fidelity speech synthesis. Advances in Neural Information Processing Systems, 33:17022–17033,", "type": "text" } ], "index": 17 }, { "bbox": [ 124, 300, 159, 314 ], "spans": [ { "bbox": [ 124, 300, 159, 314 ], "score": 1.0, "content": "2020. 5", "type": "text" } ], "index": 18 } ], "index": 17, "bbox_fs": [ 105, 281, 506, 314 ] }, { "type": "text", "bbox": [ 106, 320, 504, 351 ], "lines": [ { "bbox": [ 105, 319, 506, 334 ], "spans": [ { "bbox": [ 105, 319, 506, 334 ], "score": 1.0, "content": "[28] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song.", "type": "text" } ], "index": 19 }, { "bbox": [ 125, 330, 506, 342 ], "spans": [ { "bbox": [ 125, 330, 146, 342 ], "score": 1.0, "content": "Acav", "type": "text" }, { "bbox": [ 146, 330, 167, 340 ], "score": 0.27, "content": "1 0 0 \\mathrm { m }", "type": "inline_equation" }, { "bbox": [ 167, 330, 506, 342 ], "score": 1.0, "content": ": Automatic curation of large-scale datasets for audio-visual video representation learning. In", "type": "text" } ], "index": 20 }, { "bbox": [ 126, 340, 506, 352 ], "spans": [ { "bbox": [ 126, 340, 506, 352 ], "score": 1.0, "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10274–10284, 2021. 7", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 357, 506, 370 ], "spans": [ { "bbox": [ 106, 357, 506, 370 ], "score": 1.0, "content": "[29] Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. Optimus:", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 368, 506, 381 ], "spans": [ { "bbox": [ 126, 368, 506, 381 ], "score": 1.0, "content": "Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on", "type": "text" } ], "index": 23 }, { "bbox": [ 126, 378, 455, 390 ], "spans": [ { "bbox": [ 126, 378, 455, 390 ], "score": 1.0, "content": "Empirical Methods in Natural Language Processing (EMNLP), pages 4678–4699, 2020. 5", "type": "text" } ], "index": 24 } ], "index": 20, "bbox_fs": [ 105, 319, 506, 352 ] }, { "type": "text", "bbox": [ 106, 358, 505, 390 ], "lines": [], "index": 23, "bbox_fs": [ 106, 357, 506, 390 ], "lines_deleted": true }, { "type": "text", "bbox": [ 105, 396, 505, 417 ], "lines": [ { "bbox": [ 105, 394, 506, 410 ], "spans": [ { "bbox": [ 105, 394, 506, 410 ], "score": 1.0, "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training", "type": "text" } ], "index": 25 }, { "bbox": [ 125, 406, 483, 418 ], "spans": [ { "bbox": [ 125, 406, 483, 418 ], "score": 1.0, "content": "with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 7", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 105, 394, 506, 418 ] }, { "type": "text", "bbox": [ 105, 424, 506, 465 ], "lines": [ { "bbox": [ 106, 423, 505, 438 ], "spans": [ { "bbox": [ 106, 423, 505, 438 ], "score": 1.0, "content": "[31] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong", "type": "text" } ], "index": 27 }, { "bbox": [ 125, 434, 505, 446 ], "spans": [ { "bbox": [ 125, 434, 505, 446 ], "score": 1.0, "content": "Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In", "type": "text" } ], "index": 28 }, { "bbox": [ 126, 443, 506, 457 ], "spans": [ { "bbox": [ 126, 443, 506, 457 ], "score": 1.0, "content": "Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings,", "type": "text" } ], "index": 29 }, { "bbox": [ 125, 454, 300, 465 ], "spans": [ { "bbox": [ 125, 454, 300, 465 ], "score": 1.0, "content": "Part XXX 16, pages 121–137. Springer, 2020. 7", "type": "text" } ], "index": 30 } ], "index": 28.5, "bbox_fs": [ 106, 423, 506, 465 ] }, { "type": "text", "bbox": [ 106, 472, 504, 513 ], "lines": [ { "bbox": [ 106, 472, 506, 484 ], "spans": [ { "bbox": [ 106, 472, 506, 484 ], "score": 1.0, "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,", "type": "text" } ], "index": 31 }, { "bbox": [ 126, 483, 506, 494 ], "spans": [ { "bbox": [ 126, 483, 506, 494 ], "score": 1.0, "content": "and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014:", "type": "text" } ], "index": 32 }, { "bbox": [ 124, 491, 506, 505 ], "spans": [ { "bbox": [ 124, 491, 506, 505 ], "score": 1.0, "content": "13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages", "type": "text" } ], "index": 33 }, { "bbox": [ 125, 502, 227, 514 ], "spans": [ { "bbox": [ 125, 502, 227, 514 ], "score": 1.0, "content": "740–755. Springer, 2014. 7", "type": "text" } ], "index": 34 } ], "index": 32.5, "bbox_fs": [ 106, 472, 506, 514 ] }, { "type": "text", "bbox": [ 107, 520, 504, 551 ], "lines": [ { "bbox": [ 105, 520, 506, 533 ], "spans": [ { "bbox": [ 105, 520, 506, 533 ], "score": 1.0, "content": "[33] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and", "type": "text" } ], "index": 35 }, { "bbox": [ 125, 530, 506, 542 ], "spans": [ { "bbox": [ 125, 530, 506, 542 ], "score": 1.0, "content": "Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint", "type": "text" } ], "index": 36 }, { "bbox": [ 126, 541, 245, 552 ], "spans": [ { "bbox": [ 126, 541, 245, 552 ], "score": 1.0, "content": "arXiv:2301.12503, 2023. 2, 3, 5", "type": "text" } ], "index": 37 } ], "index": 36, "bbox_fs": [ 105, 520, 506, 552 ] }, { "type": "text", "bbox": [ 105, 558, 504, 580 ], "lines": [ { "bbox": [ 106, 559, 505, 570 ], "spans": [ { "bbox": [ 106, 559, 505, 570 ], "score": 1.0, "content": "[34] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-", "type": "text" } ], "index": 38 }, { "bbox": [ 126, 568, 473, 581 ], "spans": [ { "bbox": [ 126, 568, 473, 581 ], "score": 1.0, "content": "conditional diffusion networks for image captioning. arXiv preprint arXiv:2212.03099, 2022. 7", "type": "text" } ], "index": 39 } ], "index": 38.5, "bbox_fs": [ 106, 559, 505, 581 ] }, { "type": "text", "bbox": [ 106, 587, 504, 608 ], "lines": [ { "bbox": [ 106, 586, 505, 600 ], "spans": [ { "bbox": [ 106, 586, 505, 600 ], "score": 1.0, "content": "[35] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint", "type": "text" } ], "index": 40 }, { "bbox": [ 125, 596, 236, 608 ], "spans": [ { "bbox": [ 125, 596, 236, 608 ], "score": 1.0, "content": "arXiv:2111.09734, 2021. 3, 7", "type": "text" } ], "index": 41 } ], "index": 40.5, "bbox_fs": [ 106, 586, 505, 608 ] }, { "type": "text", "bbox": [ 106, 615, 504, 646 ], "lines": [ { "bbox": [ 106, 615, 505, 627 ], "spans": [ { "bbox": [ 106, 615, 505, 627 ], "score": 1.0, "content": "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya", "type": "text" } ], "index": 42 }, { "bbox": [ 125, 624, 505, 637 ], "spans": [ { "bbox": [ 125, 624, 505, 637 ], "score": 1.0, "content": "Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided", "type": "text" } ], "index": 43 }, { "bbox": [ 125, 634, 346, 646 ], "spans": [ { "bbox": [ 125, 634, 346, 646 ], "score": 1.0, "content": "diffusion models. arXiv preprint arXiv:2112.10741, 2021. 7", "type": "text" } ], "index": 44 } ], "index": 43, "bbox_fs": [ 106, 615, 505, 646 ] }, { "type": "text", "bbox": [ 105, 653, 504, 674 ], "lines": [ { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "[37] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Automatic prompt", "type": "text" } ], "index": 45 }, { "bbox": [ 125, 663, 481, 675 ], "spans": [ { "bbox": [ 125, 663, 481, 675 ], "score": 1.0, "content": "optimization with\" gradient descent\" and beam search. arXiv preprint arXiv:2305.03495, 2023. 2", "type": "text" } ], "index": 46 } ], "index": 45.5, "bbox_fs": [ 105, 652, 506, 675 ] }, { "type": "text", "bbox": [ 107, 681, 505, 722 ], "lines": [ { "bbox": [ 106, 681, 505, 693 ], "spans": [ { "bbox": [ 106, 681, 505, 693 ], "score": 1.0, "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", "type": "text" } ], "index": 47 }, { "bbox": [ 126, 691, 505, 703 ], "spans": [ { "bbox": [ 126, 691, 505, 703 ], "score": 1.0, "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from", "type": "text" } ], "index": 48 }, { "bbox": [ 126, 701, 506, 713 ], "spans": [ { "bbox": [ 126, 701, 506, 713 ], "score": 1.0, "content": "natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR,", "type": "text" } ], "index": 49 }, { "bbox": [ 125, 711, 167, 723 ], "spans": [ { "bbox": [ 125, 711, 167, 723 ], "score": 1.0, "content": "2021. 3, 4", "type": "text" } ], "index": 50 } ], "index": 48.5, "bbox_fs": [ 106, 681, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 104, 73, 504, 94 ], "lines": [ { "bbox": [ 105, 72, 506, 87 ], "spans": [ { "bbox": [ 105, 72, 506, 87 ], "score": 1.0, "content": "[39] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 0 }, { "bbox": [ 125, 83, 395, 94 ], "spans": [ { "bbox": [ 125, 83, 395, 94 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 5", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 106, 101, 504, 122 ], "lines": [ { "bbox": [ 105, 101, 505, 113 ], "spans": [ { "bbox": [ 105, 101, 505, 113 ], "score": 1.0, "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 408, 123 ], "spans": [ { "bbox": [ 125, 111, 408, 123 ], "score": 1.0, "content": "image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 129, 503, 160 ], "lines": [ { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution", "type": "text" } ], "index": 4 }, { "bbox": [ 124, 138, 506, 152 ], "spans": [ { "bbox": [ 124, 138, 506, 152 ], "score": 1.0, "content": "image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 149, 385, 161 ], "spans": [ { "bbox": [ 126, 149, 385, 161 ], "score": 1.0, "content": "Vision and Pattern Recognition, pages 10684–10695, 2022. 2, 3, 4, 5, 7", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "text", "bbox": [ 106, 167, 505, 218 ], "lines": [ { "bbox": [ 106, 168, 505, 178 ], "spans": [ { "bbox": [ 106, 168, 505, 178 ], "score": 1.0, "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 177, 505, 189 ], "spans": [ { "bbox": [ 126, 177, 505, 189 ], "score": 1.0, "content": "Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R", "type": "text" } ], "index": 8 }, { "bbox": [ 126, 187, 505, 198 ], "spans": [ { "bbox": [ 126, 187, 505, 198 ], "score": 1.0, "content": "Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An", "type": "text" } ], "index": 9 }, { "bbox": [ 126, 197, 506, 210 ], "spans": [ { "bbox": [ 126, 197, 506, 210 ], "score": 1.0, "content": "open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on", "type": "text" } ], "index": 10 }, { "bbox": [ 125, 207, 430, 218 ], "spans": [ { "bbox": [ 125, 207, 430, 218 ], "score": 1.0, "content": "Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 6, 7", "type": "text" } ], "index": 11 } ], "index": 9 }, { "type": "text", "bbox": [ 105, 225, 504, 256 ], "lines": [ { "bbox": [ 105, 223, 506, 238 ], "spans": [ { "bbox": [ 105, 223, 506, 238 ], "score": 1.0, "content": "[43] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining", "type": "text" } ], "index": 12 }, { "bbox": [ 126, 235, 505, 247 ], "spans": [ { "bbox": [ 126, 235, 505, 247 ], "score": 1.0, "content": "for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 13 }, { "bbox": [ 125, 245, 309, 256 ], "spans": [ { "bbox": [ 125, 245, 309, 256 ], "score": 1.0, "content": "Pattern Recognition, pages 17959–17968, 2022. 7", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 263, 504, 294 ], "lines": [ { "bbox": [ 105, 262, 506, 276 ], "spans": [ { "bbox": [ 105, 262, 506, 276 ], "score": 1.0, "content": "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang,", "type": "text" } ], "index": 15 }, { "bbox": [ 126, 273, 505, 284 ], "spans": [ { "bbox": [ 126, 273, 505, 284 ], "score": 1.0, "content": "Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv", "type": "text" } ], "index": 16 }, { "bbox": [ 124, 283, 277, 295 ], "spans": [ { "bbox": [ 124, 283, 277, 295 ], "score": 1.0, "content": "preprint arXiv:2209.14792, 2022. 2, 3, 7", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "text", "bbox": [ 107, 301, 503, 332 ], "lines": [ { "bbox": [ 105, 300, 506, 313 ], "spans": [ { "bbox": [ 105, 300, 506, 313 ], "score": 1.0, "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", "type": "text" } ], "index": 18 }, { "bbox": [ 125, 311, 505, 324 ], "spans": [ { "bbox": [ 125, 311, 505, 324 ], "score": 1.0, "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages", "type": "text" } ], "index": 19 }, { "bbox": [ 126, 321, 241, 333 ], "spans": [ { "bbox": [ 126, 321, 241, 333 ], "score": 1.0, "content": "2256–2265. PMLR, 2015. 2, 4", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 106, 339, 503, 370 ], "lines": [ { "bbox": [ 105, 339, 505, 351 ], "spans": [ { "bbox": [ 105, 339, 505, 351 ], "score": 1.0, "content": "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole.", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 349, 505, 361 ], "spans": [ { "bbox": [ 126, 349, 505, 361 ], "score": 1.0, "content": "Score-based generative modeling through stochastic differential equations. In International Conference on", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 359, 255, 371 ], "spans": [ { "bbox": [ 126, 359, 255, 371 ], "score": 1.0, "content": "Learning Representations, 2021. 3", "type": "text" } ], "index": 23 } ], "index": 22 }, { "type": "text", "bbox": [ 105, 376, 504, 408 ], "lines": [ { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "[47] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer.", "type": "text" } ], "index": 24 }, { "bbox": [ 126, 387, 505, 399 ], "spans": [ { "bbox": [ 126, 387, 505, 399 ], "score": 1.0, "content": "In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural", "type": "text" } ], "index": 25 }, { "bbox": [ 126, 397, 277, 409 ], "spans": [ { "bbox": [ 126, 397, 277, 409 ], "score": 1.0, "content": "Information Processing Systems, 2022. 3", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "text", "bbox": [ 106, 415, 504, 446 ], "lines": [ { "bbox": [ 105, 415, 506, 427 ], "spans": [ { "bbox": [ 105, 415, 506, 427 ], "score": 1.0, "content": "[48] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu,", "type": "text" } ], "index": 27 }, { "bbox": [ 125, 425, 505, 437 ], "spans": [ { "bbox": [ 125, 425, 505, 437 ], "score": 1.0, "content": "and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint", "type": "text" } ], "index": 28 }, { "bbox": [ 126, 435, 227, 446 ], "spans": [ { "bbox": [ 126, 435, 227, 446 ], "score": 1.0, "content": "arXiv:2205.14100, 2022. 7", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 105, 452, 504, 484 ], "lines": [ { "bbox": [ 105, 452, 506, 466 ], "spans": [ { "bbox": [ 105, 452, 506, 466 ], "score": 1.0, "content": "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren", "type": "text" } ], "index": 30 }, { "bbox": [ 126, 463, 506, 475 ], "spans": [ { "bbox": [ 126, 463, 506, 475 ], "score": 1.0, "content": "Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-to-", "type": "text" } ], "index": 31 }, { "bbox": [ 126, 474, 392, 484 ], "spans": [ { "bbox": [ 126, 474, 392, 484 ], "score": 1.0, "content": "sequence learning framework. arXiv preprint arXiv:2202.03052, 2022. 7", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 105, 491, 505, 522 ], "lines": [ { "bbox": [ 105, 490, 507, 504 ], "spans": [ { "bbox": [ 105, 490, 507, 504 ], "score": 1.0, "content": "[50] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan.", "type": "text" } ], "index": 33 }, { "bbox": [ 125, 501, 506, 513 ], "spans": [ { "bbox": [ 125, 501, 506, 513 ], "score": 1.0, "content": "Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806,", "type": "text" } ], "index": 34 }, { "bbox": [ 124, 510, 158, 523 ], "spans": [ { "bbox": [ 124, 510, 158, 523 ], "score": 1.0, "content": "2021. 7", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "text", "bbox": [ 106, 529, 507, 569 ], "lines": [ { "bbox": [ 106, 529, 505, 541 ], "spans": [ { "bbox": [ 106, 529, 505, 541 ], "score": 1.0, "content": "[51] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual", "type": "text" } ], "index": 36 }, { "bbox": [ 125, 538, 506, 552 ], "spans": [ { "bbox": [ 125, 538, 506, 552 ], "score": 1.0, "content": "synthesis pre-training for neural visual world creation. In Computer Vision–ECCV 2022: 17th European", "type": "text" } ], "index": 37 }, { "bbox": [ 126, 549, 506, 561 ], "spans": [ { "bbox": [ 126, 549, 506, 561 ], "score": 1.0, "content": "Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI, pages 720–736. Springer, 2022.", "type": "text" } ], "index": 38 }, { "bbox": [ 128, 562, 131, 567 ], "spans": [ { "bbox": [ 128, 562, 131, 567 ], "score": 1.0, "content": "7", "type": "text" } ], "index": 39 } ], "index": 37.5 }, { "type": "text", "bbox": [ 105, 577, 504, 608 ], "lines": [ { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian,", "type": "text" } ], "index": 40 }, { "bbox": [ 125, 587, 506, 599 ], "spans": [ { "bbox": [ 125, 587, 506, 599 ], "score": 1.0, "content": "Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video.", "type": "text" } ], "index": 41 }, { "bbox": [ 126, 597, 281, 609 ], "spans": [ { "bbox": [ 126, 597, 281, 609 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.00402, 2023. 7", "type": "text" } ], "index": 42 } ], "index": 41 }, { "type": "text", "bbox": [ 105, 615, 505, 636 ], "lines": [ { "bbox": [ 106, 615, 506, 626 ], "spans": [ { "bbox": [ 106, 615, 506, 626 ], "score": 1.0, "content": "[53] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text,", "type": "text" } ], "index": 43 }, { "bbox": [ 126, 625, 467, 636 ], "spans": [ { "bbox": [ 126, 625, 467, 636 ], "score": 1.0, "content": "images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 5, 7", "type": "text" } ], "index": 44 } ], "index": 43.5 }, { "type": "text", "bbox": [ 106, 643, 505, 684 ], "lines": [ { "bbox": [ 106, 642, 506, 656 ], "spans": [ { "bbox": [ 106, 642, 506, 656 ], "score": 1.0, "content": "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining", "type": "text" } ], "index": 45 }, { "bbox": [ 125, 653, 505, 665 ], "spans": [ { "bbox": [ 125, 653, 505, 665 ], "score": 1.0, "content": "Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In", "type": "text" } ], "index": 46 }, { "bbox": [ 125, 662, 506, 675 ], "spans": [ { "bbox": [ 125, 662, 506, 675 ], "score": 1.0, "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036–5045,", "type": "text" } ], "index": 47 }, { "bbox": [ 126, 673, 180, 684 ], "spans": [ { "bbox": [ 126, 673, 180, 684 ], "score": 1.0, "content": "2022. 6, 7, 16", "type": "text" } ], "index": 48 } ], "index": 46.5 }, { "type": "text", "bbox": [ 105, 691, 504, 722 ], "lines": [ { "bbox": [ 105, 691, 506, 703 ], "spans": [ { "bbox": [ 105, 691, 506, 703 ], "score": 1.0, "content": "[55] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian,", "type": "text" } ], "index": 49 }, { "bbox": [ 126, 701, 506, 713 ], "spans": [ { "bbox": [ 126, 701, 506, 713 ], "score": 1.0, "content": "Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework.", "type": "text" } ], "index": 50 }, { "bbox": [ 126, 711, 290, 723 ], "spans": [ { "bbox": [ 126, 711, 290, 723 ], "score": 1.0, "content": "arXiv preprint arXiv:2205.01818, 2022. 2, 3", "type": "text" } ], "index": 51 } ], "index": 50 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 104, 73, 504, 94 ], "lines": [ { "bbox": [ 105, 72, 506, 87 ], "spans": [ { "bbox": [ 105, 72, 506, 87 ], "score": 1.0, "content": "[39] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 0 }, { "bbox": [ 125, 83, 395, 94 ], "spans": [ { "bbox": [ 125, 83, 395, 94 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 5", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 72, 506, 94 ] }, { "type": "text", "bbox": [ 106, 101, 504, 122 ], "lines": [ { "bbox": [ 105, 101, 505, 113 ], "spans": [ { "bbox": [ 105, 101, 505, 113 ], "score": 1.0, "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional", "type": "text" } ], "index": 2 }, { "bbox": [ 125, 111, 408, 123 ], "spans": [ { "bbox": [ 125, 111, 408, 123 ], "score": 1.0, "content": "image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 105, 101, 505, 123 ] }, { "type": "text", "bbox": [ 106, 129, 503, 160 ], "lines": [ { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution", "type": "text" } ], "index": 4 }, { "bbox": [ 124, 138, 506, 152 ], "spans": [ { "bbox": [ 124, 138, 506, 152 ], "score": 1.0, "content": "image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 149, 385, 161 ], "spans": [ { "bbox": [ 126, 149, 385, 161 ], "score": 1.0, "content": "Vision and Pattern Recognition, pages 10684–10695, 2022. 2, 3, 4, 5, 7", "type": "text" } ], "index": 6 } ], "index": 5, "bbox_fs": [ 105, 128, 506, 161 ] }, { "type": "text", "bbox": [ 106, 167, 505, 218 ], "lines": [ { "bbox": [ 106, 168, 505, 178 ], "spans": [ { "bbox": [ 106, 168, 505, 178 ], "score": 1.0, "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 177, 505, 189 ], "spans": [ { "bbox": [ 126, 177, 505, 189 ], "score": 1.0, "content": "Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R", "type": "text" } ], "index": 8 }, { "bbox": [ 126, 187, 505, 198 ], "spans": [ { "bbox": [ 126, 187, 505, 198 ], "score": 1.0, "content": "Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An", "type": "text" } ], "index": 9 }, { "bbox": [ 126, 197, 506, 210 ], "spans": [ { "bbox": [ 126, 197, 506, 210 ], "score": 1.0, "content": "open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on", "type": "text" } ], "index": 10 }, { "bbox": [ 125, 207, 430, 218 ], "spans": [ { "bbox": [ 125, 207, 430, 218 ], "score": 1.0, "content": "Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 6, 7", "type": "text" } ], "index": 11 } ], "index": 9, "bbox_fs": [ 106, 168, 506, 218 ] }, { "type": "text", "bbox": [ 105, 225, 504, 256 ], "lines": [ { "bbox": [ 105, 223, 506, 238 ], "spans": [ { "bbox": [ 105, 223, 506, 238 ], "score": 1.0, "content": "[43] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining", "type": "text" } ], "index": 12 }, { "bbox": [ 126, 235, 505, 247 ], "spans": [ { "bbox": [ 126, 235, 505, 247 ], "score": 1.0, "content": "for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 13 }, { "bbox": [ 125, 245, 309, 256 ], "spans": [ { "bbox": [ 125, 245, 309, 256 ], "score": 1.0, "content": "Pattern Recognition, pages 17959–17968, 2022. 7", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 105, 223, 506, 256 ] }, { "type": "text", "bbox": [ 106, 263, 504, 294 ], "lines": [ { "bbox": [ 105, 262, 506, 276 ], "spans": [ { "bbox": [ 105, 262, 506, 276 ], "score": 1.0, "content": "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang,", "type": "text" } ], "index": 15 }, { "bbox": [ 126, 273, 505, 284 ], "spans": [ { "bbox": [ 126, 273, 505, 284 ], "score": 1.0, "content": "Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv", "type": "text" } ], "index": 16 }, { "bbox": [ 124, 283, 277, 295 ], "spans": [ { "bbox": [ 124, 283, 277, 295 ], "score": 1.0, "content": "preprint arXiv:2209.14792, 2022. 2, 3, 7", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 105, 262, 506, 295 ] }, { "type": "text", "bbox": [ 107, 301, 503, 332 ], "lines": [ { "bbox": [ 105, 300, 506, 313 ], "spans": [ { "bbox": [ 105, 300, 506, 313 ], "score": 1.0, "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", "type": "text" } ], "index": 18 }, { "bbox": [ 125, 311, 505, 324 ], "spans": [ { "bbox": [ 125, 311, 505, 324 ], "score": 1.0, "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages", "type": "text" } ], "index": 19 }, { "bbox": [ 126, 321, 241, 333 ], "spans": [ { "bbox": [ 126, 321, 241, 333 ], "score": 1.0, "content": "2256–2265. PMLR, 2015. 2, 4", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 105, 300, 506, 333 ] }, { "type": "text", "bbox": [ 106, 339, 503, 370 ], "lines": [ { "bbox": [ 105, 339, 505, 351 ], "spans": [ { "bbox": [ 105, 339, 505, 351 ], "score": 1.0, "content": "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole.", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 349, 505, 361 ], "spans": [ { "bbox": [ 126, 349, 505, 361 ], "score": 1.0, "content": "Score-based generative modeling through stochastic differential equations. In International Conference on", "type": "text" } ], "index": 22 }, { "bbox": [ 126, 359, 255, 371 ], "spans": [ { "bbox": [ 126, 359, 255, 371 ], "score": 1.0, "content": "Learning Representations, 2021. 3", "type": "text" } ], "index": 23 } ], "index": 22, "bbox_fs": [ 105, 339, 505, 371 ] }, { "type": "text", "bbox": [ 105, 376, 504, 408 ], "lines": [ { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "[47] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer.", "type": "text" } ], "index": 24 }, { "bbox": [ 126, 387, 505, 399 ], "spans": [ { "bbox": [ 126, 387, 505, 399 ], "score": 1.0, "content": "In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural", "type": "text" } ], "index": 25 }, { "bbox": [ 126, 397, 277, 409 ], "spans": [ { "bbox": [ 126, 397, 277, 409 ], "score": 1.0, "content": "Information Processing Systems, 2022. 3", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 105, 376, 506, 409 ] }, { "type": "text", "bbox": [ 106, 415, 504, 446 ], "lines": [ { "bbox": [ 105, 415, 506, 427 ], "spans": [ { "bbox": [ 105, 415, 506, 427 ], "score": 1.0, "content": "[48] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu,", "type": "text" } ], "index": 27 }, { "bbox": [ 125, 425, 505, 437 ], "spans": [ { "bbox": [ 125, 425, 505, 437 ], "score": 1.0, "content": "and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint", "type": "text" } ], "index": 28 }, { "bbox": [ 126, 435, 227, 446 ], "spans": [ { "bbox": [ 126, 435, 227, 446 ], "score": 1.0, "content": "arXiv:2205.14100, 2022. 7", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 105, 415, 506, 446 ] }, { "type": "text", "bbox": [ 105, 452, 504, 484 ], "lines": [ { "bbox": [ 105, 452, 506, 466 ], "spans": [ { "bbox": [ 105, 452, 506, 466 ], "score": 1.0, "content": "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren", "type": "text" } ], "index": 30 }, { "bbox": [ 126, 463, 506, 475 ], "spans": [ { "bbox": [ 126, 463, 506, 475 ], "score": 1.0, "content": "Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-to-", "type": "text" } ], "index": 31 }, { "bbox": [ 126, 474, 392, 484 ], "spans": [ { "bbox": [ 126, 474, 392, 484 ], "score": 1.0, "content": "sequence learning framework. arXiv preprint arXiv:2202.03052, 2022. 7", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 105, 452, 506, 484 ] }, { "type": "text", "bbox": [ 105, 491, 505, 522 ], "lines": [ { "bbox": [ 105, 490, 507, 504 ], "spans": [ { "bbox": [ 105, 490, 507, 504 ], "score": 1.0, "content": "[50] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan.", "type": "text" } ], "index": 33 }, { "bbox": [ 125, 501, 506, 513 ], "spans": [ { "bbox": [ 125, 501, 506, 513 ], "score": 1.0, "content": "Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806,", "type": "text" } ], "index": 34 }, { "bbox": [ 124, 510, 158, 523 ], "spans": [ { "bbox": [ 124, 510, 158, 523 ], "score": 1.0, "content": "2021. 7", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 105, 490, 507, 523 ] }, { "type": "text", "bbox": [ 106, 529, 507, 569 ], "lines": [ { "bbox": [ 106, 529, 505, 541 ], "spans": [ { "bbox": [ 106, 529, 505, 541 ], "score": 1.0, "content": "[51] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual", "type": "text" } ], "index": 36 }, { "bbox": [ 125, 538, 506, 552 ], "spans": [ { "bbox": [ 125, 538, 506, 552 ], "score": 1.0, "content": "synthesis pre-training for neural visual world creation. In Computer Vision–ECCV 2022: 17th European", "type": "text" } ], "index": 37 }, { "bbox": [ 126, 549, 506, 561 ], "spans": [ { "bbox": [ 126, 549, 506, 561 ], "score": 1.0, "content": "Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI, pages 720–736. Springer, 2022.", "type": "text" } ], "index": 38 }, { "bbox": [ 128, 562, 131, 567 ], "spans": [ { "bbox": [ 128, 562, 131, 567 ], "score": 1.0, "content": "7", "type": "text" } ], "index": 39 } ], "index": 37.5, "bbox_fs": [ 106, 529, 506, 567 ] }, { "type": "text", "bbox": [ 105, 577, 504, 608 ], "lines": [ { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian,", "type": "text" } ], "index": 40 }, { "bbox": [ 125, 587, 506, 599 ], "spans": [ { "bbox": [ 125, 587, 506, 599 ], "score": 1.0, "content": "Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video.", "type": "text" } ], "index": 41 }, { "bbox": [ 126, 597, 281, 609 ], "spans": [ { "bbox": [ 126, 597, 281, 609 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.00402, 2023. 7", "type": "text" } ], "index": 42 } ], "index": 41, "bbox_fs": [ 105, 577, 506, 609 ] }, { "type": "text", "bbox": [ 105, 615, 505, 636 ], "lines": [ { "bbox": [ 106, 615, 506, 626 ], "spans": [ { "bbox": [ 106, 615, 506, 626 ], "score": 1.0, "content": "[53] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text,", "type": "text" } ], "index": 43 }, { "bbox": [ 126, 625, 467, 636 ], "spans": [ { "bbox": [ 126, 625, 467, 636 ], "score": 1.0, "content": "images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 5, 7", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 106, 615, 506, 636 ] }, { "type": "text", "bbox": [ 106, 643, 505, 684 ], "lines": [ { "bbox": [ 106, 642, 506, 656 ], "spans": [ { "bbox": [ 106, 642, 506, 656 ], "score": 1.0, "content": "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining", "type": "text" } ], "index": 45 }, { "bbox": [ 125, 653, 505, 665 ], "spans": [ { "bbox": [ 125, 653, 505, 665 ], "score": 1.0, "content": "Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In", "type": "text" } ], "index": 46 }, { "bbox": [ 125, 662, 506, 675 ], "spans": [ { "bbox": [ 125, 662, 506, 675 ], "score": 1.0, "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036–5045,", "type": "text" } ], "index": 47 }, { "bbox": [ 126, 673, 180, 684 ], "spans": [ { "bbox": [ 126, 673, 180, 684 ], "score": 1.0, "content": "2022. 6, 7, 16", "type": "text" } ], "index": 48 } ], "index": 46.5, "bbox_fs": [ 106, 642, 506, 684 ] }, { "type": "text", "bbox": [ 105, 691, 504, 722 ], "lines": [ { "bbox": [ 105, 691, 506, 703 ], "spans": [ { "bbox": [ 105, 691, 506, 703 ], "score": 1.0, "content": "[55] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian,", "type": "text" } ], "index": 49 }, { "bbox": [ 126, 701, 506, 713 ], "spans": [ { "bbox": [ 126, 701, 506, 713 ], "score": 1.0, "content": "Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework.", "type": "text" } ], "index": 50 }, { "bbox": [ 126, 711, 290, 723 ], "spans": [ { "bbox": [ 126, 711, 290, 723 ], "score": 1.0, "content": "arXiv preprint arXiv:2205.01818, 2022. 2, 3", "type": "text" } ], "index": 51 } ], "index": 50, "bbox_fs": [ 105, 691, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 73, 506, 227 ], "lines": [ { "bbox": [ 106, 73, 506, 86 ], "spans": [ { "bbox": [ 106, 73, 506, 86 ], "score": 1.0, "content": "[56] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya", "type": "text" } ], "index": 0 }, { "bbox": [ 125, 82, 505, 96 ], "spans": [ { "bbox": [ 125, 82, 505, 96 ], "score": 1.0, "content": "Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through", "type": "text" } ], "index": 1 }, { "bbox": [ 126, 93, 506, 106 ], "spans": [ { "bbox": [ 126, 93, 506, 106 ], "score": 1.0, "content": "vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 2 }, { "bbox": [ 126, 104, 310, 115 ], "spans": [ { "bbox": [ 126, 104, 310, 115 ], "score": 1.0, "content": "Pattern Recognition, pages 16375–16387, 2022. 3", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 120, 505, 133 ], "spans": [ { "bbox": [ 106, 120, 505, 133 ], "score": 1.0, "content": "[57] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin", "type": "text" } ], "index": 4 }, { "bbox": [ 124, 130, 506, 144 ], "spans": [ { "bbox": [ 124, 130, 506, 144 ], "score": 1.0, "content": "Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 141, 255, 153 ], "spans": [ { "bbox": [ 126, 141, 255, 153 ], "score": 1.0, "content": "Systems, 34:23634–23651, 2021. 3", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "[58] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object", "type": "text" } ], "index": 7 }, { "bbox": [ 124, 168, 506, 181 ], "spans": [ { "bbox": [ 124, 168, 506, 181 ], "score": 1.0, "content": "relational graph with teacher-recommended learning for video captioning. In Proceedings of the IEEE/CVF", "type": "text" } ], "index": 8 }, { "bbox": [ 125, 179, 436, 191 ], "spans": [ { "bbox": [ 125, 179, 436, 191 ], "score": 1.0, "content": "conference on computer vision and pattern recognition, pages 13278–13288, 2020. 7", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 196, 506, 210 ], "spans": [ { "bbox": [ 105, 196, 506, 210 ], "score": 1.0, "content": "[59] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang,", "type": "text" } ], "index": 10 }, { "bbox": [ 125, 206, 506, 220 ], "spans": [ { "bbox": [ 125, 206, 506, 220 ], "score": 1.0, "content": "Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint", "type": "text" } ], "index": 11 }, { "bbox": [ 125, 217, 226, 228 ], "spans": [ { "bbox": [ 125, 217, 226, 228 ], "score": 1.0, "content": "arXiv:2211.11694, 2022. 7", "type": "text" } ], "index": 12 } ], "index": 6 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 105, 73, 506, 227 ], "lines": [ { "bbox": [ 106, 73, 506, 86 ], "spans": [ { "bbox": [ 106, 73, 506, 86 ], "score": 1.0, "content": "[56] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya", "type": "text" } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 125, 82, 505, 96 ], "spans": [ { "bbox": [ 125, 82, 505, 96 ], "score": 1.0, "content": "Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through", "type": "text" } ], "index": 1 }, { "bbox": [ 126, 93, 506, 106 ], "spans": [ { "bbox": [ 126, 93, 506, 106 ], "score": 1.0, "content": "vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 2 }, { "bbox": [ 126, 104, 310, 115 ], "spans": [ { "bbox": [ 126, 104, 310, 115 ], "score": 1.0, "content": "Pattern Recognition, pages 16375–16387, 2022. 3", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 106, 120, 505, 133 ], "spans": [ { "bbox": [ 106, 120, 505, 133 ], "score": 1.0, "content": "[57] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin", "type": "text" } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 124, 130, 506, 144 ], "spans": [ { "bbox": [ 124, 130, 506, 144 ], "score": 1.0, "content": "Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing", "type": "text" } ], "index": 5 }, { "bbox": [ 126, 141, 255, 153 ], "spans": [ { "bbox": [ 126, 141, 255, 153 ], "score": 1.0, "content": "Systems, 34:23634–23651, 2021. 3", "type": "text" } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "[58] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object", "type": "text" } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 124, 168, 506, 181 ], "spans": [ { "bbox": [ 124, 168, 506, 181 ], "score": 1.0, "content": "relational graph with teacher-recommended learning for video captioning. In Proceedings of the IEEE/CVF", "type": "text" } ], "index": 8 }, { "bbox": [ 125, 179, 436, 191 ], "spans": [ { "bbox": [ 125, 179, 436, 191 ], "score": 1.0, "content": "conference on computer vision and pattern recognition, pages 13278–13288, 2020. 7", "type": "text" } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 105, 196, 506, 210 ], "spans": [ { "bbox": [ 105, 196, 506, 210 ], "score": 1.0, "content": "[59] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang,", "type": "text" } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 125, 206, 506, 220 ], "spans": [ { "bbox": [ 125, 206, 506, 220 ], "score": 1.0, "content": "Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint", "type": "text" } ], "index": 11 }, { "bbox": [ 125, 217, 226, 228 ], "spans": [ { "bbox": [ 125, 217, 226, 228 ], "score": 1.0, "content": "arXiv:2211.11694, 2022. 7", "type": "text" } ], "index": 12, "is_list_end_line": true } ], "index": 6, "bbox_fs": [ 105, 73, 506, 228 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 71, 326, 85 ], "lines": [ { "bbox": [ 104, 68, 327, 88 ], "spans": [ { "bbox": [ 104, 68, 327, 88 ], "score": 1.0, "content": "A Model Architecture and Configuration", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 107, 97, 173, 109 ], "lines": [ { "bbox": [ 105, 95, 174, 111 ], "spans": [ { "bbox": [ 105, 95, 174, 111 ], "score": 1.0, "content": "A.1 Overview", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 107, 118, 505, 162 ], "lines": [ { "bbox": [ 106, 118, 505, 129 ], "spans": [ { "bbox": [ 106, 118, 505, 129 ], "score": 1.0, "content": "In this section, we provide more details on the model architecture as shown in Table 12, where each", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 129, 506, 140 ], "spans": [ { "bbox": [ 105, 129, 506, 140 ], "score": 1.0, "content": "modality specific diffuser is based on UNet architecture with different variations detailed in the table.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 140, 505, 152 ], "spans": [ { "bbox": [ 106, 140, 505, 152 ], "score": 1.0, "content": "Another notable difference is the video architecture where we add temporal attention and temporal", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 151, 430, 162 ], "spans": [ { "bbox": [ 106, 151, 430, 162 ], "score": 1.0, "content": "shift as discussed in Section 3.3 and we will discuss its detail in the next section.", "type": "text" } ], "index": 5 } ], "index": 3.5 }, { "type": "table", "bbox": [ 135, 199, 471, 488 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 172, 504, 195 ], "group_id": 0, "lines": [ { "bbox": [ 105, 171, 505, 185 ], "spans": [ { "bbox": [ 105, 171, 505, 185 ], "score": 1.0, "content": "Table 12: Hyperparameters for our diffusion models. Note the video and image generation uses the", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 183, 165, 195 ], "spans": [ { "bbox": [ 105, 183, 165, 195 ], "score": 1.0, "content": "same diffuser.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 135, 199, 471, 488 ], "group_id": 0, "lines": [ { "bbox": [ 135, 199, 471, 488 ], "spans": [ { "bbox": [ 135, 199, 471, 488 ], "score": 0.985, "html": "
ModalityVideo (Image) LDMAudio LDMText LDM
Hyperparameter
ArchitectureLDMLDMLDM
z-shape4× #frames × 64× 648× 256×16768×1×1
Channels320320320
Depth422
Channel multiplier1,2,4,41,2,4,41,2,4,4
Attention resolutions64,32,1664,32,1664,32,16
Head channels323232
Number of heads888
CA embed dim768768768
CA resolutions64,32,1664,32,1664,32,16
AutoencodersAutoKLAudioLDMOptimus
Weight initializationStable Diffusion-1.4-Versatile Diffusion
ParameterizationEEE
Learning rate2e-55e-65e-5
Total batch size25610241024
Diffusion Setup
Diffusion steps100010001000
Noise scheduleLinearLinearLinear
β0.000850.000850.00085
阳0.01200.01200.0120
Sampling Parameters
SamplerDDIMDDIMDDIM
Steps505050
n1.01.01.0
Guidance scale2.07.52.0
", "type": "table", "image_path": "0f1eb37557b3404ed3d97b4e9270dbc9d4725f2fb42c6ba33e4d45037b6bd75d.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 135, 199, 471, 295.3333333333333 ], "spans": [], "index": 8 }, { "bbox": [ 135, 295.3333333333333, 471, 391.66666666666663 ], "spans": [], "index": 9 }, { "bbox": [ 135, 391.66666666666663, 471, 487.99999999999994 ], "spans": [], "index": 10 } ] } ], "index": 7.75 }, { "type": "title", "bbox": [ 107, 514, 240, 525 ], "lines": [ { "bbox": [ 106, 513, 240, 527 ], "spans": [ { "bbox": [ 106, 513, 240, 527 ], "score": 1.0, "content": "A.2 Video LDM Architecture", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 106, 534, 505, 645 ], "lines": [ { "bbox": [ 105, 534, 505, 548 ], "spans": [ { "bbox": [ 105, 534, 505, 548 ], "score": 1.0, "content": "Except for the base image UNet architecture, we also add temporal attention and temporal shift [2]", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 546, 505, 558 ], "spans": [ { "bbox": [ 106, 546, 505, 558 ], "score": 1.0, "content": "before each residual block. Following VDM [21], the temporal attention is a transformer attention", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 557, 505, 569 ], "spans": [ { "bbox": [ 106, 557, 505, 569 ], "score": 1.0, "content": "module where we flatten the height and width dimension to batch size dimension and the self-attention", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 568, 505, 580 ], "spans": [ { "bbox": [ 106, 568, 505, 580 ], "score": 1.0, "content": "is performed on the time dimension. The temporal shift is illustrated in Fig. 6 where we first split", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 578, 506, 591 ], "spans": [ { "bbox": [ 106, 578, 163, 591 ], "score": 1.0, "content": "channels into", "type": "text" }, { "bbox": [ 164, 579, 171, 588 ], "score": 0.79, "content": "k", "type": "inline_equation" }, { "bbox": [ 171, 578, 427, 591 ], "score": 1.0, "content": "chunks. Then, we shift the channel dimension numbered 0 to", "type": "text" }, { "bbox": [ 427, 579, 452, 589 ], "score": 0.88, "content": "k - 1", "type": "inline_equation" }, { "bbox": [ 452, 578, 506, 591 ], "score": 1.0, "content": "by temporal", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 589, 505, 603 ], "spans": [ { "bbox": [ 105, 589, 193, 603 ], "score": 1.0, "content": "dimension from 0 to", "type": "text" }, { "bbox": [ 194, 590, 218, 600 ], "score": 0.87, "content": "k - 1", "type": "inline_equation" }, { "bbox": [ 219, 589, 505, 603 ], "score": 1.0, "content": "times respectively. Eventually, we concatenate the shifted chunks by", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 599, 506, 613 ], "spans": [ { "bbox": [ 105, 599, 273, 613 ], "score": 1.0, "content": "the hidden dimension. Note that we use", "type": "text" }, { "bbox": [ 274, 601, 301, 611 ], "score": 0.9, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 301, 599, 450, 613 ], "score": 1.0, "content": "in the illustration for simplicity but", "type": "text" }, { "bbox": [ 450, 601, 477, 611 ], "score": 0.91, "content": "k = 8", "type": "inline_equation" }, { "bbox": [ 477, 599, 506, 613 ], "score": 1.0, "content": "in our", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 611, 506, 624 ], "spans": [ { "bbox": [ 105, 611, 506, 624 ], "score": 1.0, "content": "implementation. We then add a convolution layer before the temporal shift module. Finally, we use", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "residual connection [18] and add the output to the input before the convolution layer. The complete", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 633, 253, 646 ], "spans": [ { "bbox": [ 105, 633, 253, 646 ], "score": 1.0, "content": "video UNet layer is shown in Fig. 7.", "type": "text" } ], "index": 21 } ], "index": 16.5 }, { "type": "title", "bbox": [ 107, 662, 208, 676 ], "lines": [ { "bbox": [ 104, 660, 210, 680 ], "spans": [ { "bbox": [ 104, 660, 210, 680 ], "score": 1.0, "content": "B Model Training", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 689, 504, 722 ], "lines": [ { "bbox": [ 106, 688, 505, 702 ], "spans": [ { "bbox": [ 106, 688, 505, 702 ], "score": 1.0, "content": "Prompt Encoders Training. As discussed in Section 3.2, we use bridging alignment to perform", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "contrastive learning between all prompt encoders. We use Adam [26] optimizer with learning rate", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 711, 221, 723 ], "spans": [ { "bbox": [ 106, 711, 221, 723 ], "score": 1.0, "content": "1e-4 and weight decay 1e-4.", "type": "text" } ], "index": 25 } ], "index": 24 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 754 ], "spans": [ { "bbox": [ 299, 740, 312, 754 ], "score": 1.0, "content": "15", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 71, 326, 85 ], "lines": [ { "bbox": [ 104, 68, 327, 88 ], "spans": [ { "bbox": [ 104, 68, 327, 88 ], "score": 1.0, "content": "A Model Architecture and Configuration", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 107, 97, 173, 109 ], "lines": [ { "bbox": [ 105, 95, 174, 111 ], "spans": [ { "bbox": [ 105, 95, 174, 111 ], "score": 1.0, "content": "A.1 Overview", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 107, 118, 505, 162 ], "lines": [ { "bbox": [ 106, 118, 505, 129 ], "spans": [ { "bbox": [ 106, 118, 505, 129 ], "score": 1.0, "content": "In this section, we provide more details on the model architecture as shown in Table 12, where each", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 129, 506, 140 ], "spans": [ { "bbox": [ 105, 129, 506, 140 ], "score": 1.0, "content": "modality specific diffuser is based on UNet architecture with different variations detailed in the table.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 140, 505, 152 ], "spans": [ { "bbox": [ 106, 140, 505, 152 ], "score": 1.0, "content": "Another notable difference is the video architecture where we add temporal attention and temporal", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 151, 430, 162 ], "spans": [ { "bbox": [ 106, 151, 430, 162 ], "score": 1.0, "content": "shift as discussed in Section 3.3 and we will discuss its detail in the next section.", "type": "text" } ], "index": 5 } ], "index": 3.5, "bbox_fs": [ 105, 118, 506, 162 ] }, { "type": "table", "bbox": [ 135, 199, 471, 488 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 172, 504, 195 ], "group_id": 0, "lines": [ { "bbox": [ 105, 171, 505, 185 ], "spans": [ { "bbox": [ 105, 171, 505, 185 ], "score": 1.0, "content": "Table 12: Hyperparameters for our diffusion models. Note the video and image generation uses the", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 183, 165, 195 ], "spans": [ { "bbox": [ 105, 183, 165, 195 ], "score": 1.0, "content": "same diffuser.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 135, 199, 471, 488 ], "group_id": 0, "lines": [ { "bbox": [ 135, 199, 471, 488 ], "spans": [ { "bbox": [ 135, 199, 471, 488 ], "score": 0.985, "html": "
ModalityVideo (Image) LDMAudio LDMText LDM
Hyperparameter
ArchitectureLDMLDMLDM
z-shape4× #frames × 64× 648× 256×16768×1×1
Channels320320320
Depth422
Channel multiplier1,2,4,41,2,4,41,2,4,4
Attention resolutions64,32,1664,32,1664,32,16
Head channels323232
Number of heads888
CA embed dim768768768
CA resolutions64,32,1664,32,1664,32,16
AutoencodersAutoKLAudioLDMOptimus
Weight initializationStable Diffusion-1.4-Versatile Diffusion
ParameterizationEEE
Learning rate2e-55e-65e-5
Total batch size25610241024
Diffusion Setup
Diffusion steps100010001000
Noise scheduleLinearLinearLinear
β0.000850.000850.00085
阳0.01200.01200.0120
Sampling Parameters
SamplerDDIMDDIMDDIM
Steps505050
n1.01.01.0
Guidance scale2.07.52.0
", "type": "table", "image_path": "0f1eb37557b3404ed3d97b4e9270dbc9d4725f2fb42c6ba33e4d45037b6bd75d.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 135, 199, 471, 295.3333333333333 ], "spans": [], "index": 8 }, { "bbox": [ 135, 295.3333333333333, 471, 391.66666666666663 ], "spans": [], "index": 9 }, { "bbox": [ 135, 391.66666666666663, 471, 487.99999999999994 ], "spans": [], "index": 10 } ] } ], "index": 7.75 }, { "type": "title", "bbox": [ 107, 514, 240, 525 ], "lines": [ { "bbox": [ 106, 513, 240, 527 ], "spans": [ { "bbox": [ 106, 513, 240, 527 ], "score": 1.0, "content": "A.2 Video LDM Architecture", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 106, 534, 505, 645 ], "lines": [ { "bbox": [ 105, 534, 505, 548 ], "spans": [ { "bbox": [ 105, 534, 505, 548 ], "score": 1.0, "content": "Except for the base image UNet architecture, we also add temporal attention and temporal shift [2]", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 546, 505, 558 ], "spans": [ { "bbox": [ 106, 546, 505, 558 ], "score": 1.0, "content": "before each residual block. Following VDM [21], the temporal attention is a transformer attention", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 557, 505, 569 ], "spans": [ { "bbox": [ 106, 557, 505, 569 ], "score": 1.0, "content": "module where we flatten the height and width dimension to batch size dimension and the self-attention", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 568, 505, 580 ], "spans": [ { "bbox": [ 106, 568, 505, 580 ], "score": 1.0, "content": "is performed on the time dimension. The temporal shift is illustrated in Fig. 6 where we first split", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 578, 506, 591 ], "spans": [ { "bbox": [ 106, 578, 163, 591 ], "score": 1.0, "content": "channels into", "type": "text" }, { "bbox": [ 164, 579, 171, 588 ], "score": 0.79, "content": "k", "type": "inline_equation" }, { "bbox": [ 171, 578, 427, 591 ], "score": 1.0, "content": "chunks. Then, we shift the channel dimension numbered 0 to", "type": "text" }, { "bbox": [ 427, 579, 452, 589 ], "score": 0.88, "content": "k - 1", "type": "inline_equation" }, { "bbox": [ 452, 578, 506, 591 ], "score": 1.0, "content": "by temporal", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 589, 505, 603 ], "spans": [ { "bbox": [ 105, 589, 193, 603 ], "score": 1.0, "content": "dimension from 0 to", "type": "text" }, { "bbox": [ 194, 590, 218, 600 ], "score": 0.87, "content": "k - 1", "type": "inline_equation" }, { "bbox": [ 219, 589, 505, 603 ], "score": 1.0, "content": "times respectively. Eventually, we concatenate the shifted chunks by", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 599, 506, 613 ], "spans": [ { "bbox": [ 105, 599, 273, 613 ], "score": 1.0, "content": "the hidden dimension. Note that we use", "type": "text" }, { "bbox": [ 274, 601, 301, 611 ], "score": 0.9, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 301, 599, 450, 613 ], "score": 1.0, "content": "in the illustration for simplicity but", "type": "text" }, { "bbox": [ 450, 601, 477, 611 ], "score": 0.91, "content": "k = 8", "type": "inline_equation" }, { "bbox": [ 477, 599, 506, 613 ], "score": 1.0, "content": "in our", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 611, 506, 624 ], "spans": [ { "bbox": [ 105, 611, 506, 624 ], "score": 1.0, "content": "implementation. We then add a convolution layer before the temporal shift module. Finally, we use", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "residual connection [18] and add the output to the input before the convolution layer. The complete", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 633, 253, 646 ], "spans": [ { "bbox": [ 105, 633, 253, 646 ], "score": 1.0, "content": "video UNet layer is shown in Fig. 7.", "type": "text" } ], "index": 21 } ], "index": 16.5, "bbox_fs": [ 105, 534, 506, 646 ] }, { "type": "title", "bbox": [ 107, 662, 208, 676 ], "lines": [ { "bbox": [ 104, 660, 210, 680 ], "spans": [ { "bbox": [ 104, 660, 210, 680 ], "score": 1.0, "content": "B Model Training", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 689, 504, 722 ], "lines": [ { "bbox": [ 106, 688, 505, 702 ], "spans": [ { "bbox": [ 106, 688, 505, 702 ], "score": 1.0, "content": "Prompt Encoders Training. As discussed in Section 3.2, we use bridging alignment to perform", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "contrastive learning between all prompt encoders. We use Adam [26] optimizer with learning rate", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 711, 221, 723 ], "spans": [ { "bbox": [ 106, 711, 221, 723 ], "score": 1.0, "content": "1e-4 and weight decay 1e-4.", "type": "text" } ], "index": 25 } ], "index": 24, "bbox_fs": [ 106, 688, 505, 723 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 136, 69, 473, 267 ], "blocks": [ { "type": "image_body", "bbox": [ 136, 69, 473, 267 ], "group_id": 0, "lines": [ { "bbox": [ 136, 69, 473, 267 ], "spans": [ { "bbox": [ 136, 69, 473, 267 ], "score": 0.972, "type": "image", "image_path": "e11082636d4b4499d75ad6f7451413683dac74d3ebb613e652b9ea9f4c995784.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 136, 69, 473, 135.0 ], "spans": [], "index": 0 }, { "bbox": [ 136, 135.0, 473, 201.0 ], "spans": [], "index": 1 }, { "bbox": [ 136, 201.0, 473, 267.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 276, 506, 299 ], "group_id": 0, "lines": [ { "bbox": [ 106, 276, 505, 289 ], "spans": [ { "bbox": [ 106, 276, 265, 289 ], "score": 1.0, "content": "Figure 6: Temporal shift [2] illustration.", "type": "text" }, { "bbox": [ 266, 276, 290, 288 ], "score": 0.26, "content": "C , H .", "type": "inline_equation" }, { "bbox": [ 290, 276, 292, 289 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 293, 277, 305, 287 ], "score": 0.48, "content": "W", "type": "inline_equation" }, { "bbox": [ 305, 276, 505, 289 ], "score": 1.0, "content": "represent channel, height, width, respectively. The", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 287, 506, 300 ], "spans": [ { "bbox": [ 106, 287, 260, 300 ], "score": 1.0, "content": "vertical line represents time steps from", "type": "text" }, { "bbox": [ 261, 288, 290, 298 ], "score": 0.68, "content": "t - 1 , t", "type": "inline_equation" }, { "bbox": [ 291, 287, 311, 300 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 311, 288, 333, 298 ], "score": 0.89, "content": "t + 1", "type": "inline_equation" }, { "bbox": [ 334, 287, 506, 300 ], "score": 1.0, "content": ". The grey blocks denote “padding tensors”.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "image", "bbox": [ 213, 308, 394, 396 ], "blocks": [ { "type": "image_body", "bbox": [ 213, 308, 394, 396 ], "group_id": 1, "lines": [ { "bbox": [ 213, 308, 394, 396 ], "spans": [ { "bbox": [ 213, 308, 394, 396 ], "score": 0.966, "type": "image", "image_path": "4409bc09bf832575915923690ca96fa4e7ccd04266897f8300b3a97df95ceeee.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 213, 308, 394, 322.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 213, 322.6666666666667, 394, 337.33333333333337 ], "spans": [], "index": 6 }, { "bbox": [ 213, 337.33333333333337, 394, 352.00000000000006 ], "spans": [], "index": 7 }, { "bbox": [ 213, 352.00000000000006, 394, 366.66666666666674 ], "spans": [], "index": 8 }, { "bbox": [ 213, 366.66666666666674, 394, 381.3333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 213, 381.3333333333334, 394, 396.0000000000001 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 106, 405, 504, 427 ], "group_id": 1, "lines": [ { "bbox": [ 106, 405, 505, 417 ], "spans": [ { "bbox": [ 106, 405, 505, 417 ], "score": 1.0, "content": "Figure 7: Video UNet layer architecture details including normalization & activation, 2D temporal", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 416, 366, 428 ], "spans": [ { "bbox": [ 106, 416, 366, 428 ], "score": 1.0, "content": "attention, followed by temporal shift and 1D spatial convolution.", "type": "text" } ], "index": 12 } ], "index": 11.5 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 447, 506, 547 ], "lines": [ { "bbox": [ 106, 448, 506, 461 ], "spans": [ { "bbox": [ 106, 448, 506, 461 ], "score": 1.0, "content": "Diffusion Model Training. We train diffusion model with training objectives and hyperparameters", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 458, 506, 472 ], "spans": [ { "bbox": [ 106, 458, 506, 472 ], "score": 1.0, "content": "detailed in Table 1 and Table 12. For video LDM, we adopt a more specific training curriculum.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "We adopt curriculum learning on frame resolution and frames-per-second (FPS). First, the diffuser", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 480, 507, 494 ], "spans": [ { "bbox": [ 105, 480, 507, 494 ], "score": 1.0, "content": "is trained on the WebVid dataset of a 256-frame resolution, with the training objective being text-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 492, 506, 504 ], "spans": [ { "bbox": [ 105, 492, 506, 504 ], "score": 1.0, "content": "conditioned video generation. The training clips are sampled from 2-second video chunks with 4 FPS.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 503, 505, 515 ], "spans": [ { "bbox": [ 106, 503, 505, 515 ], "score": 1.0, "content": "Second, the model is further trained on HDVILLA and ACAV datasets, with a 512-frame resolution", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 514, 505, 526 ], "spans": [ { "bbox": [ 106, 514, 505, 526 ], "score": 1.0, "content": "and 8 FPS, and the training objective is image-conditioned video generation (the image is a randomly", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 524, 506, 537 ], "spans": [ { "bbox": [ 106, 524, 506, 537 ], "score": 1.0, "content": "sampled frame of the clip). Each training clip contains 16 frames sampled from a 2-second video", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 536, 182, 547 ], "spans": [ { "bbox": [ 106, 536, 182, 547 ], "score": 1.0, "content": "chunk with 8 FPS.", "type": "text" } ], "index": 21 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 559, 504, 592 ], "lines": [ { "bbox": [ 105, 558, 505, 573 ], "spans": [ { "bbox": [ 105, 558, 505, 573 ], "score": 1.0, "content": "Joint Generation Training. As discussed in Section 3.2, we train joint generation by aligning", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 570, 505, 582 ], "spans": [ { "bbox": [ 105, 570, 505, 582 ], "score": 1.0, "content": "environment encoders and optimize cross-attention layers only in the diffusion models. We use Adam", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 581, 335, 593 ], "spans": [ { "bbox": [ 105, 581, 335, 593 ], "score": 1.0, "content": "optimizer with learning rate 1e-5 and weight decay 1e-4.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "title", "bbox": [ 107, 608, 219, 621 ], "lines": [ { "bbox": [ 105, 605, 221, 624 ], "spans": [ { "bbox": [ 105, 605, 221, 624 ], "score": 1.0, "content": "C Training Datasets", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 104, 632, 471, 644 ], "lines": [ { "bbox": [ 105, 631, 473, 646 ], "spans": [ { "bbox": [ 105, 631, 473, 646 ], "score": 1.0, "content": "In this section, we introduce more details about the video and audiovisual training datasets.", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 655, 505, 722 ], "lines": [ { "bbox": [ 106, 655, 505, 668 ], "spans": [ { "bbox": [ 106, 655, 505, 668 ], "score": 1.0, "content": "Video. WebVid [4] is a large-scale dataset of web videos with diverse content, spanning over 40", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 667, 506, 680 ], "spans": [ { "bbox": [ 105, 667, 506, 680 ], "score": 1.0, "content": "categories such as sports, cooking, and travel. It contains over 1.2 million video clips (all without", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 470, 691 ], "score": 1.0, "content": "sound) that are all at least 30 seconds in duration with video descriptions. We perform text", "type": "text" }, { "bbox": [ 471, 679, 481, 688 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 678, 505, 691 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 688, 505, 701 ], "spans": [ { "bbox": [ 106, 688, 505, 701 ], "score": 1.0, "content": "and video-text contrastive learning task with this dataset. HD-Villa-100M [54] is a large-scale video", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "dataset with over 100 million video clips sourced from YouTube. The dataset covers a wide range of", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "video categories and includes high-quality videos with a resolution of at least 720P. Since it lacks", "type": "text" } ], "index": 32 } ], "index": 29.5 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 742, 311, 750 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 136, 69, 473, 267 ], "blocks": [ { "type": "image_body", "bbox": [ 136, 69, 473, 267 ], "group_id": 0, "lines": [ { "bbox": [ 136, 69, 473, 267 ], "spans": [ { "bbox": [ 136, 69, 473, 267 ], "score": 0.972, "type": "image", "image_path": "e11082636d4b4499d75ad6f7451413683dac74d3ebb613e652b9ea9f4c995784.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 136, 69, 473, 135.0 ], "spans": [], "index": 0 }, { "bbox": [ 136, 135.0, 473, 201.0 ], "spans": [], "index": 1 }, { "bbox": [ 136, 201.0, 473, 267.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 276, 506, 299 ], "group_id": 0, "lines": [ { "bbox": [ 106, 276, 505, 289 ], "spans": [ { "bbox": [ 106, 276, 265, 289 ], "score": 1.0, "content": "Figure 6: Temporal shift [2] illustration.", "type": "text" }, { "bbox": [ 266, 276, 290, 288 ], "score": 0.26, "content": "C , H .", "type": "inline_equation" }, { "bbox": [ 290, 276, 292, 289 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 293, 277, 305, 287 ], "score": 0.48, "content": "W", "type": "inline_equation" }, { "bbox": [ 305, 276, 505, 289 ], "score": 1.0, "content": "represent channel, height, width, respectively. The", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 287, 506, 300 ], "spans": [ { "bbox": [ 106, 287, 260, 300 ], "score": 1.0, "content": "vertical line represents time steps from", "type": "text" }, { "bbox": [ 261, 288, 290, 298 ], "score": 0.68, "content": "t - 1 , t", "type": "inline_equation" }, { "bbox": [ 291, 287, 311, 300 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 311, 288, 333, 298 ], "score": 0.89, "content": "t + 1", "type": "inline_equation" }, { "bbox": [ 334, 287, 506, 300 ], "score": 1.0, "content": ". The grey blocks denote “padding tensors”.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "image", "bbox": [ 213, 308, 394, 396 ], "blocks": [ { "type": "image_body", "bbox": [ 213, 308, 394, 396 ], "group_id": 1, "lines": [ { "bbox": [ 213, 308, 394, 396 ], "spans": [ { "bbox": [ 213, 308, 394, 396 ], "score": 0.966, "type": "image", "image_path": "4409bc09bf832575915923690ca96fa4e7ccd04266897f8300b3a97df95ceeee.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 213, 308, 394, 322.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 213, 322.6666666666667, 394, 337.33333333333337 ], "spans": [], "index": 6 }, { "bbox": [ 213, 337.33333333333337, 394, 352.00000000000006 ], "spans": [], "index": 7 }, { "bbox": [ 213, 352.00000000000006, 394, 366.66666666666674 ], "spans": [], "index": 8 }, { "bbox": [ 213, 366.66666666666674, 394, 381.3333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 213, 381.3333333333334, 394, 396.0000000000001 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 106, 405, 504, 427 ], "group_id": 1, "lines": [ { "bbox": [ 106, 405, 505, 417 ], "spans": [ { "bbox": [ 106, 405, 505, 417 ], "score": 1.0, "content": "Figure 7: Video UNet layer architecture details including normalization & activation, 2D temporal", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 416, 366, 428 ], "spans": [ { "bbox": [ 106, 416, 366, 428 ], "score": 1.0, "content": "attention, followed by temporal shift and 1D spatial convolution.", "type": "text" } ], "index": 12 } ], "index": 11.5 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 447, 506, 547 ], "lines": [ { "bbox": [ 106, 448, 506, 461 ], "spans": [ { "bbox": [ 106, 448, 506, 461 ], "score": 1.0, "content": "Diffusion Model Training. We train diffusion model with training objectives and hyperparameters", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 458, 506, 472 ], "spans": [ { "bbox": [ 106, 458, 506, 472 ], "score": 1.0, "content": "detailed in Table 1 and Table 12. For video LDM, we adopt a more specific training curriculum.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "We adopt curriculum learning on frame resolution and frames-per-second (FPS). First, the diffuser", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 480, 507, 494 ], "spans": [ { "bbox": [ 105, 480, 507, 494 ], "score": 1.0, "content": "is trained on the WebVid dataset of a 256-frame resolution, with the training objective being text-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 492, 506, 504 ], "spans": [ { "bbox": [ 105, 492, 506, 504 ], "score": 1.0, "content": "conditioned video generation. The training clips are sampled from 2-second video chunks with 4 FPS.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 503, 505, 515 ], "spans": [ { "bbox": [ 106, 503, 505, 515 ], "score": 1.0, "content": "Second, the model is further trained on HDVILLA and ACAV datasets, with a 512-frame resolution", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 514, 505, 526 ], "spans": [ { "bbox": [ 106, 514, 505, 526 ], "score": 1.0, "content": "and 8 FPS, and the training objective is image-conditioned video generation (the image is a randomly", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 524, 506, 537 ], "spans": [ { "bbox": [ 106, 524, 506, 537 ], "score": 1.0, "content": "sampled frame of the clip). Each training clip contains 16 frames sampled from a 2-second video", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 536, 182, 547 ], "spans": [ { "bbox": [ 106, 536, 182, 547 ], "score": 1.0, "content": "chunk with 8 FPS.", "type": "text" } ], "index": 21 } ], "index": 17, "bbox_fs": [ 105, 448, 507, 547 ] }, { "type": "text", "bbox": [ 107, 559, 504, 592 ], "lines": [ { "bbox": [ 105, 558, 505, 573 ], "spans": [ { "bbox": [ 105, 558, 505, 573 ], "score": 1.0, "content": "Joint Generation Training. As discussed in Section 3.2, we train joint generation by aligning", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 570, 505, 582 ], "spans": [ { "bbox": [ 105, 570, 505, 582 ], "score": 1.0, "content": "environment encoders and optimize cross-attention layers only in the diffusion models. We use Adam", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 581, 335, 593 ], "spans": [ { "bbox": [ 105, 581, 335, 593 ], "score": 1.0, "content": "optimizer with learning rate 1e-5 and weight decay 1e-4.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 105, 558, 505, 593 ] }, { "type": "title", "bbox": [ 107, 608, 219, 621 ], "lines": [ { "bbox": [ 105, 605, 221, 624 ], "spans": [ { "bbox": [ 105, 605, 221, 624 ], "score": 1.0, "content": "C Training Datasets", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 104, 632, 471, 644 ], "lines": [ { "bbox": [ 105, 631, 473, 646 ], "spans": [ { "bbox": [ 105, 631, 473, 646 ], "score": 1.0, "content": "In this section, we introduce more details about the video and audiovisual training datasets.", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 105, 631, 473, 646 ] }, { "type": "text", "bbox": [ 107, 655, 505, 722 ], "lines": [ { "bbox": [ 106, 655, 505, 668 ], "spans": [ { "bbox": [ 106, 655, 505, 668 ], "score": 1.0, "content": "Video. WebVid [4] is a large-scale dataset of web videos with diverse content, spanning over 40", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 667, 506, 680 ], "spans": [ { "bbox": [ 105, 667, 506, 680 ], "score": 1.0, "content": "categories such as sports, cooking, and travel. It contains over 1.2 million video clips (all without", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 470, 691 ], "score": 1.0, "content": "sound) that are all at least 30 seconds in duration with video descriptions. We perform text", "type": "text" }, { "bbox": [ 471, 679, 481, 688 ], "score": 0.78, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 678, 505, 691 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 688, 505, 701 ], "spans": [ { "bbox": [ 106, 688, 505, 701 ], "score": 1.0, "content": "and video-text contrastive learning task with this dataset. HD-Villa-100M [54] is a large-scale video", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "dataset with over 100 million video clips sourced from YouTube. The dataset covers a wide range of", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "video categories and includes high-quality videos with a resolution of at least 720P. Since it lacks", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 72, 505, 86 ], "spans": [ { "bbox": [ 105, 72, 470, 86 ], "score": 1.0, "content": "curated video description and we use the middle frame as image input to perform image", "type": "text", "cross_page": true }, { "bbox": [ 471, 74, 481, 83 ], "score": 0.77, "content": "", "type": "inline_equation", "cross_page": true }, { "bbox": [ 481, 72, 505, 86 ], "score": 1.0, "content": "video", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 84, 154, 95 ], "spans": [ { "bbox": [ 105, 84, 154, 95 ], "score": 1.0, "content": "generation.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 29.5, "bbox_fs": [ 105, 655, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 72, 504, 95 ], "lines": [ { "bbox": [ 105, 72, 505, 86 ], "spans": [ { "bbox": [ 105, 72, 470, 86 ], "score": 1.0, "content": "curated video description and we use the middle frame as image input to perform image", "type": "text" }, { "bbox": [ 471, 74, 481, 83 ], "score": 0.77, "content": "", "type": "inline_equation" }, { "bbox": [ 481, 72, 505, 86 ], "score": 1.0, "content": "video", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 84, 154, 95 ], "spans": [ { "bbox": [ 105, 84, 154, 95 ], "score": 1.0, "content": "generation.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 107, 106, 505, 140 ], "lines": [ { "bbox": [ 106, 106, 506, 118 ], "spans": [ { "bbox": [ 106, 106, 506, 118 ], "score": 1.0, "content": "Audiovisual. SoundNet originally contains over two million sounds and spans a wide range of", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 118, 505, 129 ], "spans": [ { "bbox": [ 105, 118, 505, 129 ], "score": 1.0, "content": "categories including music, animal sounds, natural sounds, and environmental sounds. We collected", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 129, 248, 141 ], "spans": [ { "bbox": [ 106, 129, 248, 141 ], "score": 1.0, "content": "all currently accessible 1M videos.", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "title", "bbox": [ 108, 155, 290, 169 ], "lines": [ { "bbox": [ 104, 153, 291, 172 ], "spans": [ { "bbox": [ 104, 153, 291, 172 ], "score": 1.0, "content": "D Limitations & Broader Impacts", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 108, 180, 504, 213 ], "lines": [ { "bbox": [ 106, 180, 505, 192 ], "spans": [ { "bbox": [ 106, 180, 505, 192 ], "score": 1.0, "content": "While the paper primarily focuses on the technical advancements and potential applications of", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 191, 505, 203 ], "spans": [ { "bbox": [ 106, 191, 505, 203 ], "score": 1.0, "content": "CoDi, we also consider potential negative social impacts that could arise from the development and", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 201, 348, 215 ], "spans": [ { "bbox": [ 106, 201, 348, 215 ], "score": 1.0, "content": "deployment of such technology. These impacts can include:", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 225, 505, 280 ], "lines": [ { "bbox": [ 105, 224, 506, 238 ], "spans": [ { "bbox": [ 105, 224, 506, 238 ], "score": 1.0, "content": "Deepfakes and Misinformation. As part of a common issue for generative AI models, the ability", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 235, 505, 249 ], "spans": [ { "bbox": [ 105, 235, 505, 249 ], "score": 1.0, "content": "of CoDi to generate realistic and synchronized multimodal outputs also raises concerns about the", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "creation and dissemination of deepfakes. Malicious actors could exploit this technology to create", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "highly convincing fake content, such as fabricated videos or audio clips, which can be used for", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 268, 307, 281 ], "spans": [ { "bbox": [ 106, 268, 307, 281 ], "score": 1.0, "content": "misinformation, fraud, or other harmful purposes.", "type": "text" } ], "index": 13 } ], "index": 11 }, { "type": "text", "bbox": [ 106, 291, 504, 314 ], "lines": [ { "bbox": [ 105, 290, 505, 305 ], "spans": [ { "bbox": [ 105, 290, 505, 305 ], "score": 1.0, "content": "Bias and Stereotyping. If the training data used for CoDi is biased or contains stereotypes, the", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 302, 319, 316 ], "spans": [ { "bbox": [ 106, 302, 319, 316 ], "score": 1.0, "content": "generated multimodal outputs may also reflect these.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "title", "bbox": [ 107, 329, 167, 343 ], "lines": [ { "bbox": [ 105, 328, 168, 344 ], "spans": [ { "bbox": [ 105, 328, 168, 344 ], "score": 1.0, "content": "E License", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 108, 354, 504, 376 ], "lines": [ { "bbox": [ 106, 353, 506, 366 ], "spans": [ { "bbox": [ 106, 353, 506, 366 ], "score": 1.0, "content": "We will publicly release our code and checkpoints. We cite licenses from the individual dataset or", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 365, 443, 378 ], "spans": [ { "bbox": [ 105, 365, 443, 378 ], "score": 1.0, "content": "package we use from the community and provide the following links for references.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "text", "bbox": [ 108, 381, 296, 392 ], "lines": [ { "bbox": [ 106, 380, 297, 394 ], "spans": [ { "bbox": [ 106, 380, 297, 394 ], "score": 1.0, "content": "LAION-400M: Creative Common CC-BY 4.0", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 108, 398, 275, 409 ], "lines": [ { "bbox": [ 106, 396, 275, 410 ], "spans": [ { "bbox": [ 106, 396, 275, 410 ], "score": 1.0, "content": "AudioSet: Creative Common CC-BY 4.0", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 414, 180, 425 ], "lines": [ { "bbox": [ 105, 412, 181, 427 ], "spans": [ { "bbox": [ 105, 412, 181, 427 ], "score": 1.0, "content": "AudioCaps: MIT", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 430, 237, 442 ], "lines": [ { "bbox": [ 106, 430, 237, 443 ], "spans": [ { "bbox": [ 106, 430, 237, 443 ], "score": 1.0, "content": "Freesound: Creative Commons", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 108, 446, 306, 458 ], "lines": [ { "bbox": [ 106, 445, 306, 459 ], "spans": [ { "bbox": [ 106, 445, 306, 459 ], "score": 1.0, "content": "BBC Sound Effect: The BBC’s Content Licence", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 463, 175, 474 ], "lines": [ { "bbox": [ 105, 462, 176, 475 ], "spans": [ { "bbox": [ 105, 462, 176, 475 ], "score": 1.0, "content": "SoundNet: MIT", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 479, 196, 491 ], "lines": [ { "bbox": [ 106, 478, 198, 492 ], "spans": [ { "bbox": [ 106, 478, 198, 492 ], "score": 1.0, "content": "Webvid10M: Webvid", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 109, 496, 333, 507 ], "lines": [ { "bbox": [ 106, 494, 335, 509 ], "spans": [ { "bbox": [ 106, 494, 335, 509 ], "score": 1.0, "content": "HD-Villa-100M: Research Use of Data Agreement v1.0", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 512, 193, 524 ], "lines": [ { "bbox": [ 106, 511, 193, 525 ], "spans": [ { "bbox": [ 106, 511, 193, 525 ], "score": 1.0, "content": "PyTorch: BSD-style", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 529, 260, 541 ], "lines": [ { "bbox": [ 105, 528, 261, 541 ], "spans": [ { "bbox": [ 105, 528, 261, 541 ], "score": 1.0, "content": "Huggingface Transformers: Apache", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 545, 221, 556 ], "lines": [ { "bbox": [ 106, 545, 222, 558 ], "spans": [ { "bbox": [ 106, 545, 222, 558 ], "score": 1.0, "content": "Torchvision: BSD 3-Clause", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 561, 221, 573 ], "lines": [ { "bbox": [ 105, 560, 222, 574 ], "spans": [ { "bbox": [ 105, 560, 222, 574 ], "score": 1.0, "content": "Torchaudio: BSD 2-Clause", "type": "text" } ], "index": 30 } ], "index": 30 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 754 ], "spans": [ { "bbox": [ 299, 740, 312, 754 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 72, 504, 95 ], "lines": [], "index": 0.5, "bbox_fs": [ 105, 72, 505, 95 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 106, 505, 140 ], "lines": [ { "bbox": [ 106, 106, 506, 118 ], "spans": [ { "bbox": [ 106, 106, 506, 118 ], "score": 1.0, "content": "Audiovisual. SoundNet originally contains over two million sounds and spans a wide range of", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 118, 505, 129 ], "spans": [ { "bbox": [ 105, 118, 505, 129 ], "score": 1.0, "content": "categories including music, animal sounds, natural sounds, and environmental sounds. We collected", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 129, 248, 141 ], "spans": [ { "bbox": [ 106, 129, 248, 141 ], "score": 1.0, "content": "all currently accessible 1M videos.", "type": "text" } ], "index": 4 } ], "index": 3, "bbox_fs": [ 105, 106, 506, 141 ] }, { "type": "title", "bbox": [ 108, 155, 290, 169 ], "lines": [ { "bbox": [ 104, 153, 291, 172 ], "spans": [ { "bbox": [ 104, 153, 291, 172 ], "score": 1.0, "content": "D Limitations & Broader Impacts", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 108, 180, 504, 213 ], "lines": [ { "bbox": [ 106, 180, 505, 192 ], "spans": [ { "bbox": [ 106, 180, 505, 192 ], "score": 1.0, "content": "While the paper primarily focuses on the technical advancements and potential applications of", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 191, 505, 203 ], "spans": [ { "bbox": [ 106, 191, 505, 203 ], "score": 1.0, "content": "CoDi, we also consider potential negative social impacts that could arise from the development and", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 201, 348, 215 ], "spans": [ { "bbox": [ 106, 201, 348, 215 ], "score": 1.0, "content": "deployment of such technology. These impacts can include:", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 106, 180, 505, 215 ] }, { "type": "text", "bbox": [ 107, 225, 505, 280 ], "lines": [ { "bbox": [ 105, 224, 506, 238 ], "spans": [ { "bbox": [ 105, 224, 506, 238 ], "score": 1.0, "content": "Deepfakes and Misinformation. As part of a common issue for generative AI models, the ability", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 235, 505, 249 ], "spans": [ { "bbox": [ 105, 235, 505, 249 ], "score": 1.0, "content": "of CoDi to generate realistic and synchronized multimodal outputs also raises concerns about the", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "creation and dissemination of deepfakes. Malicious actors could exploit this technology to create", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "highly convincing fake content, such as fabricated videos or audio clips, which can be used for", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 268, 307, 281 ], "spans": [ { "bbox": [ 106, 268, 307, 281 ], "score": 1.0, "content": "misinformation, fraud, or other harmful purposes.", "type": "text" } ], "index": 13 } ], "index": 11, "bbox_fs": [ 105, 224, 506, 281 ] }, { "type": "text", "bbox": [ 106, 291, 504, 314 ], "lines": [ { "bbox": [ 105, 290, 505, 305 ], "spans": [ { "bbox": [ 105, 290, 505, 305 ], "score": 1.0, "content": "Bias and Stereotyping. If the training data used for CoDi is biased or contains stereotypes, the", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 302, 319, 316 ], "spans": [ { "bbox": [ 106, 302, 319, 316 ], "score": 1.0, "content": "generated multimodal outputs may also reflect these.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 105, 290, 505, 316 ] }, { "type": "title", "bbox": [ 107, 329, 167, 343 ], "lines": [ { "bbox": [ 105, 328, 168, 344 ], "spans": [ { "bbox": [ 105, 328, 168, 344 ], "score": 1.0, "content": "E License", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 108, 354, 504, 376 ], "lines": [ { "bbox": [ 106, 353, 506, 366 ], "spans": [ { "bbox": [ 106, 353, 506, 366 ], "score": 1.0, "content": "We will publicly release our code and checkpoints. We cite licenses from the individual dataset or", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 365, 443, 378 ], "spans": [ { "bbox": [ 105, 365, 443, 378 ], "score": 1.0, "content": "package we use from the community and provide the following links for references.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 105, 353, 506, 378 ] }, { "type": "text", "bbox": [ 108, 381, 296, 392 ], "lines": [ { "bbox": [ 106, 380, 297, 394 ], "spans": [ { "bbox": [ 106, 380, 297, 394 ], "score": 1.0, "content": "LAION-400M: Creative Common CC-BY 4.0", "type": "text" } ], "index": 19 } ], "index": 19, "bbox_fs": [ 106, 380, 297, 394 ] }, { "type": "text", "bbox": [ 108, 398, 275, 409 ], "lines": [ { "bbox": [ 106, 396, 275, 410 ], "spans": [ { "bbox": [ 106, 396, 275, 410 ], "score": 1.0, "content": "AudioSet: Creative Common CC-BY 4.0", "type": "text" } ], "index": 20 } ], "index": 20, "bbox_fs": [ 106, 396, 275, 410 ] }, { "type": "text", "bbox": [ 107, 414, 180, 425 ], "lines": [ { "bbox": [ 105, 412, 181, 427 ], "spans": [ { "bbox": [ 105, 412, 181, 427 ], "score": 1.0, "content": "AudioCaps: MIT", "type": "text" } ], "index": 21 } ], "index": 21, "bbox_fs": [ 105, 412, 181, 427 ] }, { "type": "text", "bbox": [ 107, 430, 237, 442 ], "lines": [ { "bbox": [ 106, 430, 237, 443 ], "spans": [ { "bbox": [ 106, 430, 237, 443 ], "score": 1.0, "content": "Freesound: Creative Commons", "type": "text" } ], "index": 22 } ], "index": 22, "bbox_fs": [ 106, 430, 237, 443 ] }, { "type": "text", "bbox": [ 108, 446, 306, 458 ], "lines": [ { "bbox": [ 106, 445, 306, 459 ], "spans": [ { "bbox": [ 106, 445, 306, 459 ], "score": 1.0, "content": "BBC Sound Effect: The BBC’s Content Licence", "type": "text" } ], "index": 23 } ], "index": 23, "bbox_fs": [ 106, 445, 306, 459 ] }, { "type": "text", "bbox": [ 107, 463, 175, 474 ], "lines": [ { "bbox": [ 105, 462, 176, 475 ], "spans": [ { "bbox": [ 105, 462, 176, 475 ], "score": 1.0, "content": "SoundNet: MIT", "type": "text" } ], "index": 24 } ], "index": 24, "bbox_fs": [ 105, 462, 176, 475 ] }, { "type": "text", "bbox": [ 107, 479, 196, 491 ], "lines": [ { "bbox": [ 106, 478, 198, 492 ], "spans": [ { "bbox": [ 106, 478, 198, 492 ], "score": 1.0, "content": "Webvid10M: Webvid", "type": "text" } ], "index": 25 } ], "index": 25, "bbox_fs": [ 106, 478, 198, 492 ] }, { "type": "text", "bbox": [ 109, 496, 333, 507 ], "lines": [ { "bbox": [ 106, 494, 335, 509 ], "spans": [ { "bbox": [ 106, 494, 335, 509 ], "score": 1.0, "content": "HD-Villa-100M: Research Use of Data Agreement v1.0", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 106, 494, 335, 509 ] }, { "type": "text", "bbox": [ 107, 512, 193, 524 ], "lines": [ { "bbox": [ 106, 511, 193, 525 ], "spans": [ { "bbox": [ 106, 511, 193, 525 ], "score": 1.0, "content": "PyTorch: BSD-style", "type": "text" } ], "index": 27 } ], "index": 27, "bbox_fs": [ 106, 511, 193, 525 ] }, { "type": "text", "bbox": [ 107, 529, 260, 541 ], "lines": [ { "bbox": [ 105, 528, 261, 541 ], "spans": [ { "bbox": [ 105, 528, 261, 541 ], "score": 1.0, "content": "Huggingface Transformers: Apache", "type": "text" } ], "index": 28 } ], "index": 28, "bbox_fs": [ 105, 528, 261, 541 ] }, { "type": "text", "bbox": [ 107, 545, 221, 556 ], "lines": [ { "bbox": [ 106, 545, 222, 558 ], "spans": [ { "bbox": [ 106, 545, 222, 558 ], "score": 1.0, "content": "Torchvision: BSD 3-Clause", "type": "text" } ], "index": 29 } ], "index": 29, "bbox_fs": [ 106, 545, 222, 558 ] }, { "type": "text", "bbox": [ 107, 561, 221, 573 ], "lines": [ { "bbox": [ 105, 560, 222, 574 ], "spans": [ { "bbox": [ 105, 560, 222, 574 ], "score": 1.0, "content": "Torchaudio: BSD 2-Clause", "type": "text" } ], "index": 30 } ], "index": 30, "bbox_fs": [ 105, 560, 222, 574 ] } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }