{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 78, 503, 115 ], "lines": [ { "bbox": [ 105, 77, 505, 98 ], "spans": [ { "bbox": [ 105, 77, 505, 98 ], "score": 1.0, "content": "DREAMIX: VIDEO DIFFUSION MODELS ARE GENERAL", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 98, 229, 118 ], "spans": [ { "bbox": [ 107, 98, 229, 118 ], "score": 1.0, "content": "VIDEO EDITORS", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 113, 136, 244, 158 ], "lines": [ { "bbox": [ 113, 136, 201, 146 ], "spans": [ { "bbox": [ 113, 136, 201, 146 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 111, 146, 245, 158 ], "spans": [ { "bbox": [ 111, 146, 245, 158 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 284, 163, 325, 173 ], "lines": [ { "bbox": [ 283, 162, 327, 174 ], "spans": [ { "bbox": [ 283, 162, 327, 174 ], "score": 1.0, "content": "Input Video", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "image", "bbox": [ 114, 174, 496, 284 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 174, 496, 284 ], "group_id": 0, "lines": [ { "bbox": [ 114, 174, 496, 284 ], "spans": [ { "bbox": [ 114, 174, 496, 284 ], "score": 0.936, "type": "image", "image_path": "34229e9ec884d9428958ad82247223d154fca3b98d39396d79b0b5839354fa44.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 114, 174, 496, 210.66666666666666 ], "spans": [], "index": 5 }, { "bbox": [ 114, 210.66666666666666, 496, 247.33333333333331 ], "spans": [], "index": 6 }, { "bbox": [ 114, 247.33333333333331, 496, 284.0 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 106, 288, 505, 343 ], "group_id": 0, "lines": [ { "bbox": [ 106, 288, 505, 300 ], "spans": [ { "bbox": [ 106, 288, 505, 300 ], "score": 1.0, "content": "Figure 1: Video Editing with Dreamix: By conditioning on the text prompt “A bear dancing and", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 298, 505, 312 ], "spans": [ { "bbox": [ 105, 298, 505, 312 ], "score": 1.0, "content": "jumping to upbeat music, moving his whole body“, Dreamix transforms the eating monkey (top", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 309, 506, 323 ], "spans": [ { "bbox": [ 105, 309, 506, 323 ], "score": 1.0, "content": "row) into a dancing bear (bottom row), affecting motion and appearance. It maintains fidelity to", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 318, 506, 335 ], "spans": [ { "bbox": [ 105, 318, 506, 335 ], "score": 1.0, "content": "color, object and camera pose, and results in a temporally consistent video. We strongly encourage", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 332, 292, 344 ], "spans": [ { "bbox": [ 106, 332, 292, 344 ], "score": 1.0, "content": "the reviewer to view the supplementary videos", "type": "text" } ], "index": 12 } ], "index": 10 } ], "index": 8.0 }, { "type": "title", "bbox": [ 278, 356, 333, 367 ], "lines": [ { "bbox": [ 276, 354, 336, 370 ], "spans": [ { "bbox": [ 276, 354, 336, 370 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 143, 383, 468, 547 ], "lines": [ { "bbox": [ 141, 382, 469, 394 ], "spans": [ { "bbox": [ 141, 382, 469, 394 ], "score": 1.0, "content": "Text-driven image and video diffusion models have recently achieved unprece-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 393, 470, 406 ], "spans": [ { "bbox": [ 141, 393, 470, 406 ], "score": 1.0, "content": "dented generation realism. While diffusion models have been successfully applied", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 403, 470, 417 ], "spans": [ { "bbox": [ 141, 403, 470, 417 ], "score": 1.0, "content": "for image editing, none can edit motion in video. We present the first diffusion-", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 415, 470, 428 ], "spans": [ { "bbox": [ 141, 415, 470, 428 ], "score": 1.0, "content": "based method that is able to perform text-based motion and appearance editing of", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 427, 469, 438 ], "spans": [ { "bbox": [ 141, 427, 469, 438 ], "score": 1.0, "content": "general, real-world videos. Our approach uses a video diffusion model to com-", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 437, 470, 449 ], "spans": [ { "bbox": [ 141, 437, 470, 449 ], "score": 1.0, "content": "bine, at inference time, the low-resolution spatio-temporal information from the", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 448, 470, 461 ], "spans": [ { "bbox": [ 141, 448, 470, 461 ], "score": 1.0, "content": "original video with new, high resolution information that it synthesized to align", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 459, 470, 472 ], "spans": [ { "bbox": [ 141, 459, 470, 472 ], "score": 1.0, "content": "with the guiding text prompt. As maintaining high-fidelity to the original video", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 470, 469, 483 ], "spans": [ { "bbox": [ 141, 470, 469, 483 ], "score": 1.0, "content": "requires retaining some of its high-resolution information, we add a preliminary", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 481, 469, 493 ], "spans": [ { "bbox": [ 141, 481, 469, 493 ], "score": 1.0, "content": "stage of finetuning the model on the original video, significantly boosting fidelity.", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 492, 469, 505 ], "spans": [ { "bbox": [ 141, 492, 469, 505 ], "score": 1.0, "content": "We propose to improve motion editability by using a mixed objective that jointly", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 503, 469, 515 ], "spans": [ { "bbox": [ 141, 503, 469, 515 ], "score": 1.0, "content": "finetunes with full temporal attention and with temporal attention masking. We", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 514, 470, 527 ], "spans": [ { "bbox": [ 141, 514, 470, 527 ], "score": 1.0, "content": "extend our method for animating images, bringing them to life by adding mo-", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 525, 469, 537 ], "spans": [ { "bbox": [ 141, 525, 469, 537 ], "score": 1.0, "content": "tion to existing or new objects, and camera movements. Extensive experiments", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 536, 412, 548 ], "spans": [ { "bbox": [ 141, 536, 412, 548 ], "score": 1.0, "content": "showcase our method’s remarkable ability to edit motion in videos.", "type": "text" } ], "index": 28 } ], "index": 21 }, { "type": "title", "bbox": [ 108, 573, 206, 586 ], "lines": [ { "bbox": [ 105, 572, 208, 588 ], "spans": [ { "bbox": [ 105, 572, 208, 588 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 106, 600, 505, 732 ], "lines": [ { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "Recent advancements in generative models Ho et al. (2020); Chang et al. (2022); Yu et al. (2022a);", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 609, 506, 623 ], "spans": [ { "bbox": [ 105, 609, 506, 623 ], "score": 1.0, "content": "Chang et al. (2023) and multimodal vision-language models Radford et al. (2021), paved the way to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 505, 633 ], "spans": [ { "bbox": [ 105, 622, 505, 633 ], "score": 1.0, "content": "large-scale text-to-image models capable of unprecedented generation realism and diversity Ramesh", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 633, 504, 644 ], "spans": [ { "bbox": [ 106, 633, 504, 644 ], "score": 1.0, "content": "et al. (2022); Rombach et al. (2022); Saharia et al. (2022b); Nichol et al. (2021); Avrahami et al.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 643, 505, 656 ], "spans": [ { "bbox": [ 105, 643, 505, 656 ], "score": 1.0, "content": "(2022b). These models have ushered in a new era of creativity, applications, and research. Although", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 654, 505, 667 ], "spans": [ { "bbox": [ 105, 654, 505, 667 ], "score": 1.0, "content": "these models offer new creative processes, they are limited to synthesizing new images rather than", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "editing existing ones. To bridge this gap, intuitive image editing methods offer text-based editing", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "of generated and real images while maintaining some of their original attributes Hertz et al. (2022);", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 686, 506, 700 ], "spans": [ { "bbox": [ 105, 686, 506, 700 ], "score": 1.0, "content": "Tumanyan et al. (2022); Brooks et al. (2022); Kawar et al. (2022); Valevski et al. (2022). Similarly to", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 698, 505, 711 ], "spans": [ { "bbox": [ 105, 698, 505, 711 ], "score": 1.0, "content": "images, text-to-video models have recently been proposed Ho et al. (2022c;a); Singer et al. (2022);", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 709, 505, 721 ], "spans": [ { "bbox": [ 106, 709, 505, 721 ], "score": 1.0, "content": "Yu et al. (2022b), but very few methods use them for video editing and none can edit the motion in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 718, 139, 733 ], "spans": [ { "bbox": [ 105, 718, 139, 733 ], "score": 1.0, "content": "videos.", "type": "text" } ], "index": 41 } ], "index": 35.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "1", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 78, 503, 115 ], "lines": [ { "bbox": [ 105, 77, 505, 98 ], "spans": [ { "bbox": [ 105, 77, 505, 98 ], "score": 1.0, "content": "DREAMIX: VIDEO DIFFUSION MODELS ARE GENERAL", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 98, 229, 118 ], "spans": [ { "bbox": [ 107, 98, 229, 118 ], "score": 1.0, "content": "VIDEO EDITORS", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 113, 136, 244, 158 ], "lines": [ { "bbox": [ 113, 136, 201, 146 ], "spans": [ { "bbox": [ 113, 136, 201, 146 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 111, 146, 245, 158 ], "spans": [ { "bbox": [ 111, 146, 245, 158 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 111, 136, 245, 158 ] }, { "type": "text", "bbox": [ 284, 163, 325, 173 ], "lines": [ { "bbox": [ 283, 162, 327, 174 ], "spans": [ { "bbox": [ 283, 162, 327, 174 ], "score": 1.0, "content": "Input Video", "type": "text" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 283, 162, 327, 174 ] }, { "type": "image", "bbox": [ 114, 174, 496, 284 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 174, 496, 284 ], "group_id": 0, "lines": [ { "bbox": [ 114, 174, 496, 284 ], "spans": [ { "bbox": [ 114, 174, 496, 284 ], "score": 0.936, "type": "image", "image_path": "34229e9ec884d9428958ad82247223d154fca3b98d39396d79b0b5839354fa44.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 114, 174, 496, 210.66666666666666 ], "spans": [], "index": 5 }, { "bbox": [ 114, 210.66666666666666, 496, 247.33333333333331 ], "spans": [], "index": 6 }, { "bbox": [ 114, 247.33333333333331, 496, 284.0 ], "spans": [], "index": 7 } ] }, { "type": "image_caption", "bbox": [ 106, 288, 505, 343 ], "group_id": 0, "lines": [ { "bbox": [ 106, 288, 505, 300 ], "spans": [ { "bbox": [ 106, 288, 505, 300 ], "score": 1.0, "content": "Figure 1: Video Editing with Dreamix: By conditioning on the text prompt “A bear dancing and", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 298, 505, 312 ], "spans": [ { "bbox": [ 105, 298, 505, 312 ], "score": 1.0, "content": "jumping to upbeat music, moving his whole body“, Dreamix transforms the eating monkey (top", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 309, 506, 323 ], "spans": [ { "bbox": [ 105, 309, 506, 323 ], "score": 1.0, "content": "row) into a dancing bear (bottom row), affecting motion and appearance. It maintains fidelity to", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 318, 506, 335 ], "spans": [ { "bbox": [ 105, 318, 506, 335 ], "score": 1.0, "content": "color, object and camera pose, and results in a temporally consistent video. We strongly encourage", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 332, 292, 344 ], "spans": [ { "bbox": [ 106, 332, 292, 344 ], "score": 1.0, "content": "the reviewer to view the supplementary videos", "type": "text" } ], "index": 12 } ], "index": 10 } ], "index": 8.0 }, { "type": "title", "bbox": [ 278, 356, 333, 367 ], "lines": [ { "bbox": [ 276, 354, 336, 370 ], "spans": [ { "bbox": [ 276, 354, 336, 370 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 143, 383, 468, 547 ], "lines": [ { "bbox": [ 141, 382, 469, 394 ], "spans": [ { "bbox": [ 141, 382, 469, 394 ], "score": 1.0, "content": "Text-driven image and video diffusion models have recently achieved unprece-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 393, 470, 406 ], "spans": [ { "bbox": [ 141, 393, 470, 406 ], "score": 1.0, "content": "dented generation realism. While diffusion models have been successfully applied", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 403, 470, 417 ], "spans": [ { "bbox": [ 141, 403, 470, 417 ], "score": 1.0, "content": "for image editing, none can edit motion in video. We present the first diffusion-", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 415, 470, 428 ], "spans": [ { "bbox": [ 141, 415, 470, 428 ], "score": 1.0, "content": "based method that is able to perform text-based motion and appearance editing of", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 427, 469, 438 ], "spans": [ { "bbox": [ 141, 427, 469, 438 ], "score": 1.0, "content": "general, real-world videos. Our approach uses a video diffusion model to com-", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 437, 470, 449 ], "spans": [ { "bbox": [ 141, 437, 470, 449 ], "score": 1.0, "content": "bine, at inference time, the low-resolution spatio-temporal information from the", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 448, 470, 461 ], "spans": [ { "bbox": [ 141, 448, 470, 461 ], "score": 1.0, "content": "original video with new, high resolution information that it synthesized to align", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 459, 470, 472 ], "spans": [ { "bbox": [ 141, 459, 470, 472 ], "score": 1.0, "content": "with the guiding text prompt. As maintaining high-fidelity to the original video", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 470, 469, 483 ], "spans": [ { "bbox": [ 141, 470, 469, 483 ], "score": 1.0, "content": "requires retaining some of its high-resolution information, we add a preliminary", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 481, 469, 493 ], "spans": [ { "bbox": [ 141, 481, 469, 493 ], "score": 1.0, "content": "stage of finetuning the model on the original video, significantly boosting fidelity.", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 492, 469, 505 ], "spans": [ { "bbox": [ 141, 492, 469, 505 ], "score": 1.0, "content": "We propose to improve motion editability by using a mixed objective that jointly", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 503, 469, 515 ], "spans": [ { "bbox": [ 141, 503, 469, 515 ], "score": 1.0, "content": "finetunes with full temporal attention and with temporal attention masking. We", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 514, 470, 527 ], "spans": [ { "bbox": [ 141, 514, 470, 527 ], "score": 1.0, "content": "extend our method for animating images, bringing them to life by adding mo-", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 525, 469, 537 ], "spans": [ { "bbox": [ 141, 525, 469, 537 ], "score": 1.0, "content": "tion to existing or new objects, and camera movements. Extensive experiments", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 536, 412, 548 ], "spans": [ { "bbox": [ 141, 536, 412, 548 ], "score": 1.0, "content": "showcase our method’s remarkable ability to edit motion in videos.", "type": "text" } ], "index": 28 } ], "index": 21, "bbox_fs": [ 141, 382, 470, 548 ] }, { "type": "title", "bbox": [ 108, 573, 206, 586 ], "lines": [ { "bbox": [ 105, 572, 208, 588 ], "spans": [ { "bbox": [ 105, 572, 208, 588 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 106, 600, 505, 732 ], "lines": [ { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "Recent advancements in generative models Ho et al. (2020); Chang et al. (2022); Yu et al. (2022a);", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 609, 506, 623 ], "spans": [ { "bbox": [ 105, 609, 506, 623 ], "score": 1.0, "content": "Chang et al. (2023) and multimodal vision-language models Radford et al. (2021), paved the way to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 505, 633 ], "spans": [ { "bbox": [ 105, 622, 505, 633 ], "score": 1.0, "content": "large-scale text-to-image models capable of unprecedented generation realism and diversity Ramesh", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 633, 504, 644 ], "spans": [ { "bbox": [ 106, 633, 504, 644 ], "score": 1.0, "content": "et al. (2022); Rombach et al. (2022); Saharia et al. (2022b); Nichol et al. (2021); Avrahami et al.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 643, 505, 656 ], "spans": [ { "bbox": [ 105, 643, 505, 656 ], "score": 1.0, "content": "(2022b). These models have ushered in a new era of creativity, applications, and research. Although", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 654, 505, 667 ], "spans": [ { "bbox": [ 105, 654, 505, 667 ], "score": 1.0, "content": "these models offer new creative processes, they are limited to synthesizing new images rather than", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "editing existing ones. To bridge this gap, intuitive image editing methods offer text-based editing", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "of generated and real images while maintaining some of their original attributes Hertz et al. (2022);", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 686, 506, 700 ], "spans": [ { "bbox": [ 105, 686, 506, 700 ], "score": 1.0, "content": "Tumanyan et al. (2022); Brooks et al. (2022); Kawar et al. (2022); Valevski et al. (2022). Similarly to", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 698, 505, 711 ], "spans": [ { "bbox": [ 105, 698, 505, 711 ], "score": 1.0, "content": "images, text-to-video models have recently been proposed Ho et al. (2022c;a); Singer et al. (2022);", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 709, 505, 721 ], "spans": [ { "bbox": [ 106, 709, 505, 721 ], "score": 1.0, "content": "Yu et al. (2022b), but very few methods use them for video editing and none can edit the motion in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 718, 139, 733 ], "spans": [ { "bbox": [ 105, 718, 139, 733 ], "score": 1.0, "content": "videos.", "type": "text" } ], "index": 41 } ], "index": 35.5, "bbox_fs": [ 105, 599, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 170 ], "lines": [ { "bbox": [ 105, 83, 505, 95 ], "spans": [ { "bbox": [ 105, 83, 505, 95 ], "score": 1.0, "content": "In text-guided video editing, the user provides an input video and a text prompt describing the", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "desired attributes of the resulting video (Fig. 1). The objectives are three-fold: i) alignment: the", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "edited video should conform with the input text prompt ii) fidelity: the edited video should preserve", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 127 ], "spans": [ { "bbox": [ 105, 115, 505, 127 ], "score": 1.0, "content": "the content of the original input iii) quality: the edited video should be of high-quality. Video", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 505, 140 ], "spans": [ { "bbox": [ 105, 126, 505, 140 ], "score": 1.0, "content": "editing is more challenging than image editing, as it requires synthesizing new motion, not merely", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 505, 150 ], "spans": [ { "bbox": [ 105, 137, 505, 150 ], "score": 1.0, "content": "modifying appearance. It also requires temporal consistency. As a result, applying image-level", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "editing methods e.g. SDEdit Meng et al. (2021) or Prompt-to-Prompt Hertz et al. (2022) sequentially", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 246, 171 ], "spans": [ { "bbox": [ 105, 159, 246, 171 ], "score": 1.0, "content": "on the video frames is insufficient.", "type": "text" } ], "index": 7 } ], "index": 3.5 }, { "type": "text", "bbox": [ 106, 176, 505, 330 ], "lines": [ { "bbox": [ 106, 176, 505, 187 ], "spans": [ { "bbox": [ 106, 176, 505, 187 ], "score": 1.0, "content": "We present a new method, Dreamix, to adapt a text-conditioned video diffusion model (VDM) for", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 187, 506, 200 ], "spans": [ { "bbox": [ 106, 187, 506, 200 ], "score": 1.0, "content": "video editing, in a manner inspired by UniTune Valevski et al. (2022). The core of our method is", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 198, 505, 210 ], "spans": [ { "bbox": [ 106, 198, 505, 210 ], "score": 1.0, "content": "enabling a text-conditioned VDM to maintain high fidelity to an input video via two main ideas.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 209, 505, 222 ], "spans": [ { "bbox": [ 105, 209, 505, 222 ], "score": 1.0, "content": "First, instead of using pure-noise as initialization for the model, we use a degraded version of the", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 220, 505, 233 ], "spans": [ { "bbox": [ 105, 220, 505, 233 ], "score": 1.0, "content": "original video, keeping only low spatio-temporal information by downscaling it and adding noise.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 230, 505, 243 ], "spans": [ { "bbox": [ 105, 230, 505, 243 ], "score": 1.0, "content": "This is similar to SDEdit but the degradation includes not merely noise, but also downscaling. Sec-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 241, 505, 254 ], "spans": [ { "bbox": [ 105, 241, 505, 254 ], "score": 1.0, "content": "ond, we further improve the fidelity to the original video by finetuning the model on the original", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 253, 505, 265 ], "spans": [ { "bbox": [ 106, 253, 505, 265 ], "score": 1.0, "content": "video. Finetuning ensures the model has knowledge of the high-resolution attributes of the original", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 263, 505, 276 ], "spans": [ { "bbox": [ 106, 263, 505, 276 ], "score": 1.0, "content": "video. Naively finetuning on the input video results in relatively low motion editability as the model", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 275, 505, 287 ], "spans": [ { "bbox": [ 106, 275, 505, 287 ], "score": 1.0, "content": "learns to prefer the original motion instead of following the text prompt. We propose a novel use", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 505, 299 ], "spans": [ { "bbox": [ 105, 285, 505, 299 ], "score": 1.0, "content": "for the mixed finetuning approach, suggested in VDM Ho et al. (2022c) and Imagen-Video Ho et al.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "(2022a), in which the VDMs are trained on both images and video. In our approach, we finetune the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 308, 505, 319 ], "spans": [ { "bbox": [ 106, 308, 505, 319 ], "score": 1.0, "content": "model on both the original video but also on its (unordered) frames individually. This allows us to", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 319, 427, 331 ], "spans": [ { "bbox": [ 105, 319, 427, 331 ], "score": 1.0, "content": "perform significantly larger motion edits with high fidelity to the original video.", "type": "text" } ], "index": 21 } ], "index": 14.5 }, { "type": "text", "bbox": [ 107, 335, 505, 413 ], "lines": [ { "bbox": [ 105, 336, 505, 348 ], "spans": [ { "bbox": [ 105, 336, 505, 348 ], "score": 1.0, "content": "As a further contribution, we leverage our video editing model to add motion to still images (see", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 345, 505, 360 ], "spans": [ { "bbox": [ 104, 345, 505, 360 ], "score": 1.0, "content": "Fig. 2) e.g., animating the objects and background in an image or creating dynamic camera motion.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 357, 505, 370 ], "spans": [ { "bbox": [ 105, 357, 505, 370 ], "score": 1.0, "content": "To do so, we first create a coarse video by simple image processing operations, e.g., frame replication", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 369, 505, 380 ], "spans": [ { "bbox": [ 105, 369, 505, 380 ], "score": 1.0, "content": "or geometric image transformation. We then edit it with our Dreamix video editor. Our framework", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 379, 505, 393 ], "spans": [ { "bbox": [ 105, 379, 505, 393 ], "score": 1.0, "content": "can also perform subject-driven video generation (see Fig. 2), extending the scope of current image-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 390, 505, 403 ], "spans": [ { "bbox": [ 105, 390, 505, 403 ], "score": 1.0, "content": "based methods e.g., DreamboothRuiz et al. (2022) to video and motion editing. We evaluate our", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 401, 502, 414 ], "spans": [ { "bbox": [ 105, 401, 502, 414 ], "score": 1.0, "content": "method extensively, demonstrating its remarkable capabilities unmatched by the baseline methods.", "type": "text" } ], "index": 28 } ], "index": 25 }, { "type": "text", "bbox": [ 108, 418, 276, 429 ], "lines": [ { "bbox": [ 106, 416, 279, 431 ], "spans": [ { "bbox": [ 106, 416, 279, 431 ], "score": 1.0, "content": "To summarize, our main contributions are:", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 130, 439, 505, 533 ], "lines": [ { "bbox": [ 129, 439, 505, 452 ], "spans": [ { "bbox": [ 129, 439, 505, 452 ], "score": 1.0, "content": "1. Proposing the first method for text-based editing of real-world videos that can edit their", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 451, 308, 464 ], "spans": [ { "bbox": [ 141, 451, 308, 464 ], "score": 1.0, "content": "motion and not merely their appearance.", "type": "text" } ], "index": 31 }, { "bbox": [ 128, 465, 505, 480 ], "spans": [ { "bbox": [ 128, 465, 505, 480 ], "score": 1.0, "content": "2. Repurposing mixed training as a finetuning objective that significantly improves motion", "type": "text" } ], "index": 32 }, { "bbox": [ 141, 476, 176, 492 ], "spans": [ { "bbox": [ 141, 476, 176, 492 ], "score": 1.0, "content": "editing.", "type": "text" } ], "index": 33 }, { "bbox": [ 128, 492, 505, 507 ], "spans": [ { "bbox": [ 128, 492, 505, 507 ], "score": 1.0, "content": "3. Presenting a new framework for text-guided image animation, by applying our video editor", "type": "text" } ], "index": 34 }, { "bbox": [ 142, 505, 373, 518 ], "spans": [ { "bbox": [ 142, 505, 373, 518 ], "score": 1.0, "content": "method on top of simple image preprocessing operations.", "type": "text" } ], "index": 35 }, { "bbox": [ 129, 520, 464, 533 ], "spans": [ { "bbox": [ 129, 520, 464, 533 ], "score": 1.0, "content": "4. Extending the scope of subject-driven generation methods to motion generation.", "type": "text" } ], "index": 36 } ], "index": 33 }, { "type": "title", "bbox": [ 108, 553, 211, 566 ], "lines": [ { "bbox": [ 105, 552, 213, 568 ], "spans": [ { "bbox": [ 105, 552, 213, 568 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "title", "bbox": [ 108, 578, 289, 590 ], "lines": [ { "bbox": [ 106, 578, 290, 591 ], "spans": [ { "bbox": [ 106, 578, 290, 591 ], "score": 1.0, "content": "2.1 DIFFUSION MODELS FOR SYNTHESIS", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 106, 600, 505, 731 ], "lines": [ { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "Deep diffusion models recently emerged as a powerful new paradigm for image generation Ho et al.", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 610, 505, 623 ], "spans": [ { "bbox": [ 106, 610, 505, 623 ], "score": 1.0, "content": "(2020); Song et al. (2020), and have their roots in score-matching Hyvarinen & Dayan ¨ (2005); Vin-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 621, 505, 633 ], "spans": [ { "bbox": [ 105, 621, 505, 633 ], "score": 1.0, "content": "cent (2011); Sohl-Dickstein et al. (2015). They outperform Dhariwal & Nichol (2021) the previous", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 632, 505, 644 ], "spans": [ { "bbox": [ 105, 632, 505, 644 ], "score": 1.0, "content": "state-of-the-art approach, generative adversarial networks (GANs) Goodfellow et al. (2020). While", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 505, 656 ], "score": 1.0, "content": "they have multiple formulations, EDM Karras et al. (2022) showed they are equivalent. Outstand-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 655, 504, 667 ], "spans": [ { "bbox": [ 105, 655, 504, 667 ], "score": 1.0, "content": "ing progress was made in text-to-image generation Saharia et al. (2022b); Ramesh et al. (2022);", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "Rombach et al. (2022); Avrahami et al. (2022b), where new images are sampled conditioned on an", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 676, 506, 689 ], "spans": [ { "bbox": [ 105, 676, 506, 689 ], "score": 1.0, "content": "input text prompt. Extending diffusion models to video generation is a challenging computational", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 688, 504, 699 ], "spans": [ { "bbox": [ 106, 688, 504, 699 ], "score": 1.0, "content": "and algorithmic task. Early work include Ho et al. (2022c) and text-to-video extensions by Ho et al.", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 698, 505, 711 ], "spans": [ { "bbox": [ 106, 698, 505, 711 ], "score": 1.0, "content": "(2022a); Singer et al. (2022). Another line of work extends synthesis to various image reconstruc-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 708, 505, 722 ], "spans": [ { "bbox": [ 105, 708, 505, 722 ], "score": 1.0, "content": "tion tasks Saharia et al. (2022c;a); Ho et al. (2022b); Lugmayr et al. (2022); Chung et al. (2022),", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 719, 428, 732 ], "spans": [ { "bbox": [ 106, 719, 428, 732 ], "score": 1.0, "content": "Horwitz & Hoshen (2022) extracts confidence intervals for reconstruction tasks.", "type": "text" } ], "index": 50 } ], "index": 44.5 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 763 ], "spans": [ { "bbox": [ 301, 750, 310, 763 ], "score": 1.0, "content": "2", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 170 ], "lines": [ { "bbox": [ 105, 83, 505, 95 ], "spans": [ { "bbox": [ 105, 83, 505, 95 ], "score": 1.0, "content": "In text-guided video editing, the user provides an input video and a text prompt describing the", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "desired attributes of the resulting video (Fig. 1). The objectives are three-fold: i) alignment: the", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "edited video should conform with the input text prompt ii) fidelity: the edited video should preserve", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 127 ], "spans": [ { "bbox": [ 105, 115, 505, 127 ], "score": 1.0, "content": "the content of the original input iii) quality: the edited video should be of high-quality. Video", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 505, 140 ], "spans": [ { "bbox": [ 105, 126, 505, 140 ], "score": 1.0, "content": "editing is more challenging than image editing, as it requires synthesizing new motion, not merely", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 505, 150 ], "spans": [ { "bbox": [ 105, 137, 505, 150 ], "score": 1.0, "content": "modifying appearance. It also requires temporal consistency. As a result, applying image-level", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "editing methods e.g. SDEdit Meng et al. (2021) or Prompt-to-Prompt Hertz et al. (2022) sequentially", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 246, 171 ], "spans": [ { "bbox": [ 105, 159, 246, 171 ], "score": 1.0, "content": "on the video frames is insufficient.", "type": "text" } ], "index": 7 } ], "index": 3.5, "bbox_fs": [ 105, 83, 505, 171 ] }, { "type": "text", "bbox": [ 106, 176, 505, 330 ], "lines": [ { "bbox": [ 106, 176, 505, 187 ], "spans": [ { "bbox": [ 106, 176, 505, 187 ], "score": 1.0, "content": "We present a new method, Dreamix, to adapt a text-conditioned video diffusion model (VDM) for", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 187, 506, 200 ], "spans": [ { "bbox": [ 106, 187, 506, 200 ], "score": 1.0, "content": "video editing, in a manner inspired by UniTune Valevski et al. (2022). The core of our method is", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 198, 505, 210 ], "spans": [ { "bbox": [ 106, 198, 505, 210 ], "score": 1.0, "content": "enabling a text-conditioned VDM to maintain high fidelity to an input video via two main ideas.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 209, 505, 222 ], "spans": [ { "bbox": [ 105, 209, 505, 222 ], "score": 1.0, "content": "First, instead of using pure-noise as initialization for the model, we use a degraded version of the", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 220, 505, 233 ], "spans": [ { "bbox": [ 105, 220, 505, 233 ], "score": 1.0, "content": "original video, keeping only low spatio-temporal information by downscaling it and adding noise.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 230, 505, 243 ], "spans": [ { "bbox": [ 105, 230, 505, 243 ], "score": 1.0, "content": "This is similar to SDEdit but the degradation includes not merely noise, but also downscaling. Sec-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 241, 505, 254 ], "spans": [ { "bbox": [ 105, 241, 505, 254 ], "score": 1.0, "content": "ond, we further improve the fidelity to the original video by finetuning the model on the original", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 253, 505, 265 ], "spans": [ { "bbox": [ 106, 253, 505, 265 ], "score": 1.0, "content": "video. Finetuning ensures the model has knowledge of the high-resolution attributes of the original", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 263, 505, 276 ], "spans": [ { "bbox": [ 106, 263, 505, 276 ], "score": 1.0, "content": "video. Naively finetuning on the input video results in relatively low motion editability as the model", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 275, 505, 287 ], "spans": [ { "bbox": [ 106, 275, 505, 287 ], "score": 1.0, "content": "learns to prefer the original motion instead of following the text prompt. We propose a novel use", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 505, 299 ], "spans": [ { "bbox": [ 105, 285, 505, 299 ], "score": 1.0, "content": "for the mixed finetuning approach, suggested in VDM Ho et al. (2022c) and Imagen-Video Ho et al.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "(2022a), in which the VDMs are trained on both images and video. In our approach, we finetune the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 308, 505, 319 ], "spans": [ { "bbox": [ 106, 308, 505, 319 ], "score": 1.0, "content": "model on both the original video but also on its (unordered) frames individually. This allows us to", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 319, 427, 331 ], "spans": [ { "bbox": [ 105, 319, 427, 331 ], "score": 1.0, "content": "perform significantly larger motion edits with high fidelity to the original video.", "type": "text" } ], "index": 21 } ], "index": 14.5, "bbox_fs": [ 105, 176, 506, 331 ] }, { "type": "text", "bbox": [ 107, 335, 505, 413 ], "lines": [ { "bbox": [ 105, 336, 505, 348 ], "spans": [ { "bbox": [ 105, 336, 505, 348 ], "score": 1.0, "content": "As a further contribution, we leverage our video editing model to add motion to still images (see", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 345, 505, 360 ], "spans": [ { "bbox": [ 104, 345, 505, 360 ], "score": 1.0, "content": "Fig. 2) e.g., animating the objects and background in an image or creating dynamic camera motion.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 357, 505, 370 ], "spans": [ { "bbox": [ 105, 357, 505, 370 ], "score": 1.0, "content": "To do so, we first create a coarse video by simple image processing operations, e.g., frame replication", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 369, 505, 380 ], "spans": [ { "bbox": [ 105, 369, 505, 380 ], "score": 1.0, "content": "or geometric image transformation. We then edit it with our Dreamix video editor. Our framework", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 379, 505, 393 ], "spans": [ { "bbox": [ 105, 379, 505, 393 ], "score": 1.0, "content": "can also perform subject-driven video generation (see Fig. 2), extending the scope of current image-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 390, 505, 403 ], "spans": [ { "bbox": [ 105, 390, 505, 403 ], "score": 1.0, "content": "based methods e.g., DreamboothRuiz et al. (2022) to video and motion editing. We evaluate our", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 401, 502, 414 ], "spans": [ { "bbox": [ 105, 401, 502, 414 ], "score": 1.0, "content": "method extensively, demonstrating its remarkable capabilities unmatched by the baseline methods.", "type": "text" } ], "index": 28 } ], "index": 25, "bbox_fs": [ 104, 336, 505, 414 ] }, { "type": "text", "bbox": [ 108, 418, 276, 429 ], "lines": [ { "bbox": [ 106, 416, 279, 431 ], "spans": [ { "bbox": [ 106, 416, 279, 431 ], "score": 1.0, "content": "To summarize, our main contributions are:", "type": "text" } ], "index": 29 } ], "index": 29, "bbox_fs": [ 106, 416, 279, 431 ] }, { "type": "list", "bbox": [ 130, 439, 505, 533 ], "lines": [ { "bbox": [ 129, 439, 505, 452 ], "spans": [ { "bbox": [ 129, 439, 505, 452 ], "score": 1.0, "content": "1. Proposing the first method for text-based editing of real-world videos that can edit their", "type": "text" } ], "index": 30, "is_list_start_line": true }, { "bbox": [ 141, 451, 308, 464 ], "spans": [ { "bbox": [ 141, 451, 308, 464 ], "score": 1.0, "content": "motion and not merely their appearance.", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 128, 465, 505, 480 ], "spans": [ { "bbox": [ 128, 465, 505, 480 ], "score": 1.0, "content": "2. Repurposing mixed training as a finetuning objective that significantly improves motion", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 141, 476, 176, 492 ], "spans": [ { "bbox": [ 141, 476, 176, 492 ], "score": 1.0, "content": "editing.", "type": "text" } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 128, 492, 505, 507 ], "spans": [ { "bbox": [ 128, 492, 505, 507 ], "score": 1.0, "content": "3. Presenting a new framework for text-guided image animation, by applying our video editor", "type": "text" } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 142, 505, 373, 518 ], "spans": [ { "bbox": [ 142, 505, 373, 518 ], "score": 1.0, "content": "method on top of simple image preprocessing operations.", "type": "text" } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 129, 520, 464, 533 ], "spans": [ { "bbox": [ 129, 520, 464, 533 ], "score": 1.0, "content": "4. Extending the scope of subject-driven generation methods to motion generation.", "type": "text" } ], "index": 36, "is_list_start_line": true, "is_list_end_line": true } ], "index": 33, "bbox_fs": [ 128, 439, 505, 533 ] }, { "type": "title", "bbox": [ 108, 553, 211, 566 ], "lines": [ { "bbox": [ 105, 552, 213, 568 ], "spans": [ { "bbox": [ 105, 552, 213, 568 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "title", "bbox": [ 108, 578, 289, 590 ], "lines": [ { "bbox": [ 106, 578, 290, 591 ], "spans": [ { "bbox": [ 106, 578, 290, 591 ], "score": 1.0, "content": "2.1 DIFFUSION MODELS FOR SYNTHESIS", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 106, 600, 505, 731 ], "lines": [ { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "Deep diffusion models recently emerged as a powerful new paradigm for image generation Ho et al.", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 610, 505, 623 ], "spans": [ { "bbox": [ 106, 610, 505, 623 ], "score": 1.0, "content": "(2020); Song et al. (2020), and have their roots in score-matching Hyvarinen & Dayan ¨ (2005); Vin-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 621, 505, 633 ], "spans": [ { "bbox": [ 105, 621, 505, 633 ], "score": 1.0, "content": "cent (2011); Sohl-Dickstein et al. (2015). They outperform Dhariwal & Nichol (2021) the previous", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 632, 505, 644 ], "spans": [ { "bbox": [ 105, 632, 505, 644 ], "score": 1.0, "content": "state-of-the-art approach, generative adversarial networks (GANs) Goodfellow et al. (2020). While", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 505, 656 ], "score": 1.0, "content": "they have multiple formulations, EDM Karras et al. (2022) showed they are equivalent. Outstand-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 655, 504, 667 ], "spans": [ { "bbox": [ 105, 655, 504, 667 ], "score": 1.0, "content": "ing progress was made in text-to-image generation Saharia et al. (2022b); Ramesh et al. (2022);", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "Rombach et al. (2022); Avrahami et al. (2022b), where new images are sampled conditioned on an", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 676, 506, 689 ], "spans": [ { "bbox": [ 105, 676, 506, 689 ], "score": 1.0, "content": "input text prompt. Extending diffusion models to video generation is a challenging computational", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 688, 504, 699 ], "spans": [ { "bbox": [ 106, 688, 504, 699 ], "score": 1.0, "content": "and algorithmic task. Early work include Ho et al. (2022c) and text-to-video extensions by Ho et al.", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 698, 505, 711 ], "spans": [ { "bbox": [ 106, 698, 505, 711 ], "score": 1.0, "content": "(2022a); Singer et al. (2022). Another line of work extends synthesis to various image reconstruc-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 708, 505, 722 ], "spans": [ { "bbox": [ 105, 708, 505, 722 ], "score": 1.0, "content": "tion tasks Saharia et al. (2022c;a); Ho et al. (2022b); Lugmayr et al. (2022); Chung et al. (2022),", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 719, 428, 732 ], "spans": [ { "bbox": [ 106, 719, 428, 732 ], "score": 1.0, "content": "Horwitz & Hoshen (2022) extracts confidence intervals for reconstruction tasks.", "type": "text" } ], "index": 50 } ], "index": 44.5, "bbox_fs": [ 105, 599, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 108, 85, 502, 208 ], "blocks": [ { "type": "image_caption", "bbox": [ 207, 82, 471, 93 ], "group_id": 0, "lines": [ { "bbox": [ 207, 82, 471, 93 ], "spans": [], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 108, 85, 502, 208 ], "group_id": 0, "lines": [ { "bbox": [ 108, 85, 502, 208 ], "spans": [ { "bbox": [ 108, 85, 502, 208 ], "score": 0.883, "type": "image", "image_path": "a2fe01488e8b94937a9ce87bc47ec2c8dbd5a9d3474809fe87c4e0b5fe903286.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 85, 502, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 126.0, 502, 167.0 ], "spans": [], "index": 2 }, { "bbox": [ 108, 167.0, 502, 208.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 106, 223, 505, 279 ], "group_id": 0, "lines": [ { "bbox": [ 105, 223, 505, 237 ], "spans": [ { "bbox": [ 105, 223, 505, 237 ], "score": 1.0, "content": "Figure 2: Image-to-Video editing with Dreamix: Dreamix instills complex motion in a static image", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 234, 506, 248 ], "spans": [ { "bbox": [ 106, 234, 506, 248 ], "score": 1.0, "content": "(first row), adding a moving shark and making the turtle swim. In this case, visual fidelity to object", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 246, 505, 258 ], "spans": [ { "bbox": [ 106, 246, 505, 258 ], "score": 1.0, "content": "location and background was preserved but the turtle direction was flipped. In the subject-driven", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "case (second row), Dreamix extracts the visual features of a subject given multiple images and", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 267, 324, 281 ], "spans": [ { "bbox": [ 105, 267, 324, 281 ], "score": 1.0, "content": "animates it in different scenarios such as weightlifting", "type": "text" } ], "index": 8 } ], "index": 6 } ], "index": 2 }, { "type": "title", "bbox": [ 108, 300, 277, 311 ], "lines": [ { "bbox": [ 106, 300, 279, 313 ], "spans": [ { "bbox": [ 106, 300, 279, 313 ], "score": 1.0, "content": "2.2 DIFFUSION MODELS FOR EDITING", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 106, 320, 505, 562 ], "lines": [ { "bbox": [ 105, 319, 505, 334 ], "spans": [ { "bbox": [ 105, 319, 505, 334 ], "score": 1.0, "content": "Image editing with generative models has been studied extensively, in past years many of the mod-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 331, 505, 344 ], "spans": [ { "bbox": [ 105, 331, 505, 344 ], "score": 1.0, "content": "els were based on GANs Vinker et al. (2021); Patashnik et al. (2021); Gal et al. (2021); Roich", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 341, 505, 356 ], "spans": [ { "bbox": [ 104, 341, 505, 356 ], "score": 1.0, "content": "et al. (2022); Wang et al. (2018b); Park et al. (2019); Bau et al. (2020); Skorokhodov et al. (2022);", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 353, 504, 366 ], "spans": [ { "bbox": [ 106, 353, 504, 366 ], "score": 1.0, "content": "Jamriska et al. ˇ (2019); Wang et al. (2018a); Tzaban et al. (2022); Xu et al. (2022); Liu et al. (2022).", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 363, 505, 378 ], "spans": [ { "bbox": [ 105, 363, 505, 378 ], "score": 1.0, "content": "Another recent line of works demonstrated preliminary generation and editing capabilities using", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 375, 504, 387 ], "spans": [ { "bbox": [ 106, 375, 504, 387 ], "score": 1.0, "content": "masked image models Yu et al. (2022b); Villegas et al. (2022); Yao et al. (2021); Nash et al. (2022).", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 386, 505, 398 ], "spans": [ { "bbox": [ 105, 386, 505, 398 ], "score": 1.0, "content": "However, most of the recent editing methods adopt diffusion models Avrahami et al. (2022c;a);", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 396, 506, 411 ], "spans": [ { "bbox": [ 105, 396, 506, 411 ], "score": 1.0, "content": "Voynov et al. (2022). SDEdit Meng et al. (2021) proposed to add targeted noise to an input image,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 407, 504, 421 ], "spans": [ { "bbox": [ 105, 407, 504, 421 ], "score": 1.0, "content": "and then use diffusion models for reversing the process. Prompt-to-Prompt Hertz et al. (2022); Tu-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 419, 505, 432 ], "spans": [ { "bbox": [ 105, 419, 505, 432 ], "score": 1.0, "content": "manyan et al. (2022); Mokady et al. (2022) perform semantic edits by mixing activations extracted", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 430, 505, 442 ], "spans": [ { "bbox": [ 106, 430, 505, 442 ], "score": 1.0, "content": "with the original and target prompts. For InstructPix2Pix Brooks et al. (2022) this is only needed", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 441, 506, 454 ], "spans": [ { "bbox": [ 106, 441, 506, 454 ], "score": 1.0, "content": "for constructing the training dataset. Other works (e.g. Gal et al. (2022); Ruiz et al. (2022)) use", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 452, 505, 465 ], "spans": [ { "bbox": [ 106, 452, 505, 465 ], "score": 1.0, "content": "finetuning and optimization to allow for personalization of the model, learning a special token de-", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 462, 506, 475 ], "spans": [ { "bbox": [ 106, 462, 506, 475 ], "score": 1.0, "content": "scribing the content. UniTune Valevski et al. (2022) and Imagic Kawar et al. (2022) finetune on", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 473, 506, 487 ], "spans": [ { "bbox": [ 104, 473, 506, 487 ], "score": 1.0, "content": "a single image, allowing better editability while maintaining good fidelity. However, the methods", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 484, 506, 497 ], "spans": [ { "bbox": [ 105, 484, 506, 497 ], "score": 1.0, "content": "are image-centric and do not use temporal information. Neural Atlases Kasten et al. (2021) and", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 496, 506, 508 ], "spans": [ { "bbox": [ 106, 496, 506, 508 ], "score": 1.0, "content": "Text2Live Bar-Tal et al. (2022) allow some texture-based video editing, however, unlike our method", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 505, 506, 520 ], "spans": [ { "bbox": [ 105, 505, 506, 520 ], "score": 1.0, "content": "they cannot edit the motion of a video. A concurrent paper, Tune-a-Video Wu et al. (2022) pre-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "forms video editing by inflating a text-to-image model to learn temporal consistency. Despite their", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 529, 504, 541 ], "spans": [ { "bbox": [ 106, 529, 504, 541 ], "score": 1.0, "content": "promising results, they use a text-to-image backbone that can edit video appearance but not motion.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 540, 505, 552 ], "spans": [ { "bbox": [ 106, 540, 505, 552 ], "score": 1.0, "content": "Their results are also not fully temporally consistent. In contrast, our method uses a text-to-video", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 549, 476, 565 ], "spans": [ { "bbox": [ 105, 549, 476, 565 ], "score": 1.0, "content": "backbone, enabling motion editing while maintaining smoothness and temporal consistency.", "type": "text" } ], "index": 31 } ], "index": 20.5 }, { "type": "title", "bbox": [ 107, 578, 350, 591 ], "lines": [ { "bbox": [ 104, 577, 351, 593 ], "spans": [ { "bbox": [ 104, 577, 351, 593 ], "score": 1.0, "content": "3 BACKGROUND: VIDEO DIFFUSION MODELS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 106, 602, 505, 693 ], "lines": [ { "bbox": [ 105, 602, 505, 616 ], "spans": [ { "bbox": [ 105, 602, 505, 616 ], "score": 1.0, "content": "Denoising Model Training. Diffusion models rely on a deep denoising neural network denoted by", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 614, 505, 627 ], "spans": [ { "bbox": [ 106, 614, 120, 625 ], "score": 0.87, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 120, 614, 281, 627 ], "score": 1.0, "content": ". Let us denote the ground truth video as", "type": "text" }, { "bbox": [ 281, 617, 288, 624 ], "score": 0.75, "content": "v", "type": "inline_equation" }, { "bbox": [ 288, 614, 505, 627 ], "score": 1.0, "content": ", an i.i.d Gaussian noise tensor of the same dimensions", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 625, 505, 638 ], "spans": [ { "bbox": [ 105, 625, 172, 638 ], "score": 1.0, "content": "as the video as", "type": "text" }, { "bbox": [ 172, 625, 225, 637 ], "score": 0.92, "content": "\\epsilon \\sim N ( 0 , { \\bf I } )", "type": "inline_equation" }, { "bbox": [ 225, 625, 344, 638 ], "score": 1.0, "content": ", and the noise level at time", "type": "text" }, { "bbox": [ 344, 627, 351, 635 ], "score": 0.67, "content": "s", "type": "inline_equation" }, { "bbox": [ 351, 625, 364, 638 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 364, 627, 375, 636 ], "score": 0.84, "content": "\\sigma _ { s }", "type": "inline_equation" }, { "bbox": [ 376, 625, 505, 638 ], "score": 1.0, "content": ". The noisy video is given by:", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 636, 504, 651 ], "spans": [ { "bbox": [ 106, 639, 172, 650 ], "score": 0.89, "content": "z _ { s } = \\gamma _ { s } v + \\sigma _ { s } \\epsilon", "type": "inline_equation" }, { "bbox": [ 172, 637, 203, 651 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 203, 636, 265, 650 ], "score": 0.93, "content": "\\gamma _ { s } = \\sqrt { 1 - \\sigma _ { s } ^ { 2 } }", "type": "inline_equation" }, { "bbox": [ 266, 637, 498, 651 ], "score": 1.0, "content": ". Furthermore, let us denote a conditioning text prompt as", "type": "text" }, { "bbox": [ 499, 639, 504, 648 ], "score": 0.65, "content": "t", "type": "inline_equation" } ], "index": 36 }, { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 209, 662 ], "score": 1.0, "content": "and a conditioning video", "type": "text" }, { "bbox": [ 209, 651, 215, 659 ], "score": 0.74, "content": "c", "type": "inline_equation" }, { "bbox": [ 216, 648, 304, 662 ], "score": 1.0, "content": "(for super-resolution,", "type": "text" }, { "bbox": [ 305, 651, 311, 659 ], "score": 0.75, "content": "c", "type": "inline_equation" }, { "bbox": [ 311, 648, 433, 662 ], "score": 1.0, "content": "is a low-resolution version of", "type": "text" }, { "bbox": [ 433, 651, 439, 659 ], "score": 0.62, "content": "v", "type": "inline_equation" }, { "bbox": [ 439, 648, 506, 662 ], "score": 1.0, "content": "). The objective", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 658, 506, 674 ], "spans": [ { "bbox": [ 105, 658, 209, 674 ], "score": 1.0, "content": "of the denoising network", "type": "text" }, { "bbox": [ 210, 660, 223, 671 ], "score": 0.89, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 224, 658, 370, 674 ], "score": 1.0, "content": "is to recover the ground truth video", "type": "text" }, { "bbox": [ 371, 662, 378, 670 ], "score": 0.63, "content": "v", "type": "inline_equation" }, { "bbox": [ 378, 658, 491, 674 ], "score": 1.0, "content": "given the noisy input video", "type": "text" }, { "bbox": [ 491, 662, 501, 671 ], "score": 0.82, "content": "z _ { s }", "type": "inline_equation" }, { "bbox": [ 501, 658, 506, 674 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 671, 504, 684 ], "spans": [ { "bbox": [ 106, 671, 142, 684 ], "score": 1.0, "content": "the time", "type": "text" }, { "bbox": [ 142, 673, 148, 681 ], "score": 0.68, "content": "s", "type": "inline_equation" }, { "bbox": [ 148, 671, 183, 684 ], "score": 1.0, "content": ", prompt", "type": "text" }, { "bbox": [ 184, 672, 189, 681 ], "score": 0.76, "content": "t", "type": "inline_equation" }, { "bbox": [ 189, 671, 286, 684 ], "score": 1.0, "content": "and conditioning video", "type": "text" }, { "bbox": [ 286, 674, 291, 680 ], "score": 0.71, "content": "c", "type": "inline_equation" }, { "bbox": [ 292, 671, 495, 684 ], "score": 1.0, "content": ". The model is trained on a (large) training corpus", "type": "text" }, { "bbox": [ 496, 672, 504, 681 ], "score": 0.78, "content": "\\nu", "type": "inline_equation" } ], "index": 39 }, { "bbox": [ 106, 682, 303, 694 ], "spans": [ { "bbox": [ 106, 682, 217, 694 ], "score": 1.0, "content": "consisting of pairs of video", "type": "text" }, { "bbox": [ 218, 684, 224, 691 ], "score": 0.77, "content": "v", "type": "inline_equation" }, { "bbox": [ 224, 682, 295, 694 ], "score": 1.0, "content": "and text prompts", "type": "text" }, { "bbox": [ 295, 684, 299, 691 ], "score": 0.82, "content": "t", "type": "inline_equation" }, { "bbox": [ 299, 682, 303, 694 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40 } ], "index": 36.5 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 698, 506, 712 ], "spans": [ { "bbox": [ 106, 698, 506, 712 ], "score": 1.0, "content": "Sampling from Diffusion Models. The key challenge in diffusion models is to use the denoiser", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 142, 722 ], "score": 1.0, "content": "network", "type": "text" }, { "bbox": [ 142, 710, 156, 721 ], "score": 0.89, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 156, 710, 453, 722 ], "score": 1.0, "content": "to sample from the distribution of videos conditioned on the text prompt", "type": "text" }, { "bbox": [ 453, 711, 458, 720 ], "score": 0.72, "content": "t", "type": "inline_equation" }, { "bbox": [ 459, 710, 505, 722 ], "score": 1.0, "content": "and condi-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 721, 506, 733 ], "spans": [ { "bbox": [ 106, 721, 162, 733 ], "score": 1.0, "content": "tioning video", "type": "text" }, { "bbox": [ 162, 722, 168, 731 ], "score": 0.43, "content": "c", "type": "inline_equation" }, { "bbox": [ 168, 721, 171, 733 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 171, 721, 208, 732 ], "score": 0.92, "content": "P ( v | t , c )", "type": "inline_equation" }, { "bbox": [ 208, 721, 506, 733 ], "score": 1.0, "content": ". While the derivation of such sampling rule is non-trivial (see e.g. Karras", "type": "text" } ], "index": 43 } ], "index": 42 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 307, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 108, 85, 502, 208 ], "blocks": [ { "type": "image_caption", "bbox": [ 207, 82, 471, 93 ], "group_id": 0, "lines": [ { "bbox": [ 207, 82, 471, 93 ], "spans": [], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 108, 85, 502, 208 ], "group_id": 0, "lines": [ { "bbox": [ 108, 85, 502, 208 ], "spans": [ { "bbox": [ 108, 85, 502, 208 ], "score": 0.883, "type": "image", "image_path": "a2fe01488e8b94937a9ce87bc47ec2c8dbd5a9d3474809fe87c4e0b5fe903286.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 85, 502, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 126.0, 502, 167.0 ], "spans": [], "index": 2 }, { "bbox": [ 108, 167.0, 502, 208.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 106, 223, 505, 279 ], "group_id": 0, "lines": [ { "bbox": [ 105, 223, 505, 237 ], "spans": [ { "bbox": [ 105, 223, 505, 237 ], "score": 1.0, "content": "Figure 2: Image-to-Video editing with Dreamix: Dreamix instills complex motion in a static image", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 234, 506, 248 ], "spans": [ { "bbox": [ 106, 234, 506, 248 ], "score": 1.0, "content": "(first row), adding a moving shark and making the turtle swim. In this case, visual fidelity to object", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 246, 505, 258 ], "spans": [ { "bbox": [ 106, 246, 505, 258 ], "score": 1.0, "content": "location and background was preserved but the turtle direction was flipped. In the subject-driven", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "case (second row), Dreamix extracts the visual features of a subject given multiple images and", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 267, 324, 281 ], "spans": [ { "bbox": [ 105, 267, 324, 281 ], "score": 1.0, "content": "animates it in different scenarios such as weightlifting", "type": "text" } ], "index": 8 } ], "index": 6 } ], "index": 2 }, { "type": "title", "bbox": [ 108, 300, 277, 311 ], "lines": [ { "bbox": [ 106, 300, 279, 313 ], "spans": [ { "bbox": [ 106, 300, 279, 313 ], "score": 1.0, "content": "2.2 DIFFUSION MODELS FOR EDITING", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 106, 320, 505, 562 ], "lines": [ { "bbox": [ 105, 319, 505, 334 ], "spans": [ { "bbox": [ 105, 319, 505, 334 ], "score": 1.0, "content": "Image editing with generative models has been studied extensively, in past years many of the mod-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 331, 505, 344 ], "spans": [ { "bbox": [ 105, 331, 505, 344 ], "score": 1.0, "content": "els were based on GANs Vinker et al. (2021); Patashnik et al. (2021); Gal et al. (2021); Roich", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 341, 505, 356 ], "spans": [ { "bbox": [ 104, 341, 505, 356 ], "score": 1.0, "content": "et al. (2022); Wang et al. (2018b); Park et al. (2019); Bau et al. (2020); Skorokhodov et al. (2022);", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 353, 504, 366 ], "spans": [ { "bbox": [ 106, 353, 504, 366 ], "score": 1.0, "content": "Jamriska et al. ˇ (2019); Wang et al. (2018a); Tzaban et al. (2022); Xu et al. (2022); Liu et al. (2022).", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 363, 505, 378 ], "spans": [ { "bbox": [ 105, 363, 505, 378 ], "score": 1.0, "content": "Another recent line of works demonstrated preliminary generation and editing capabilities using", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 375, 504, 387 ], "spans": [ { "bbox": [ 106, 375, 504, 387 ], "score": 1.0, "content": "masked image models Yu et al. (2022b); Villegas et al. (2022); Yao et al. (2021); Nash et al. (2022).", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 386, 505, 398 ], "spans": [ { "bbox": [ 105, 386, 505, 398 ], "score": 1.0, "content": "However, most of the recent editing methods adopt diffusion models Avrahami et al. (2022c;a);", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 396, 506, 411 ], "spans": [ { "bbox": [ 105, 396, 506, 411 ], "score": 1.0, "content": "Voynov et al. (2022). SDEdit Meng et al. (2021) proposed to add targeted noise to an input image,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 407, 504, 421 ], "spans": [ { "bbox": [ 105, 407, 504, 421 ], "score": 1.0, "content": "and then use diffusion models for reversing the process. Prompt-to-Prompt Hertz et al. (2022); Tu-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 419, 505, 432 ], "spans": [ { "bbox": [ 105, 419, 505, 432 ], "score": 1.0, "content": "manyan et al. (2022); Mokady et al. (2022) perform semantic edits by mixing activations extracted", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 430, 505, 442 ], "spans": [ { "bbox": [ 106, 430, 505, 442 ], "score": 1.0, "content": "with the original and target prompts. For InstructPix2Pix Brooks et al. (2022) this is only needed", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 441, 506, 454 ], "spans": [ { "bbox": [ 106, 441, 506, 454 ], "score": 1.0, "content": "for constructing the training dataset. Other works (e.g. Gal et al. (2022); Ruiz et al. (2022)) use", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 452, 505, 465 ], "spans": [ { "bbox": [ 106, 452, 505, 465 ], "score": 1.0, "content": "finetuning and optimization to allow for personalization of the model, learning a special token de-", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 462, 506, 475 ], "spans": [ { "bbox": [ 106, 462, 506, 475 ], "score": 1.0, "content": "scribing the content. UniTune Valevski et al. (2022) and Imagic Kawar et al. (2022) finetune on", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 473, 506, 487 ], "spans": [ { "bbox": [ 104, 473, 506, 487 ], "score": 1.0, "content": "a single image, allowing better editability while maintaining good fidelity. However, the methods", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 484, 506, 497 ], "spans": [ { "bbox": [ 105, 484, 506, 497 ], "score": 1.0, "content": "are image-centric and do not use temporal information. Neural Atlases Kasten et al. (2021) and", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 496, 506, 508 ], "spans": [ { "bbox": [ 106, 496, 506, 508 ], "score": 1.0, "content": "Text2Live Bar-Tal et al. (2022) allow some texture-based video editing, however, unlike our method", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 505, 506, 520 ], "spans": [ { "bbox": [ 105, 505, 506, 520 ], "score": 1.0, "content": "they cannot edit the motion of a video. A concurrent paper, Tune-a-Video Wu et al. (2022) pre-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "forms video editing by inflating a text-to-image model to learn temporal consistency. Despite their", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 529, 504, 541 ], "spans": [ { "bbox": [ 106, 529, 504, 541 ], "score": 1.0, "content": "promising results, they use a text-to-image backbone that can edit video appearance but not motion.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 540, 505, 552 ], "spans": [ { "bbox": [ 106, 540, 505, 552 ], "score": 1.0, "content": "Their results are also not fully temporally consistent. In contrast, our method uses a text-to-video", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 549, 476, 565 ], "spans": [ { "bbox": [ 105, 549, 476, 565 ], "score": 1.0, "content": "backbone, enabling motion editing while maintaining smoothness and temporal consistency.", "type": "text" } ], "index": 31 } ], "index": 20.5, "bbox_fs": [ 104, 319, 506, 565 ] }, { "type": "title", "bbox": [ 107, 578, 350, 591 ], "lines": [ { "bbox": [ 104, 577, 351, 593 ], "spans": [ { "bbox": [ 104, 577, 351, 593 ], "score": 1.0, "content": "3 BACKGROUND: VIDEO DIFFUSION MODELS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 106, 602, 505, 693 ], "lines": [ { "bbox": [ 105, 602, 505, 616 ], "spans": [ { "bbox": [ 105, 602, 505, 616 ], "score": 1.0, "content": "Denoising Model Training. Diffusion models rely on a deep denoising neural network denoted by", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 614, 505, 627 ], "spans": [ { "bbox": [ 106, 614, 120, 625 ], "score": 0.87, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 120, 614, 281, 627 ], "score": 1.0, "content": ". Let us denote the ground truth video as", "type": "text" }, { "bbox": [ 281, 617, 288, 624 ], "score": 0.75, "content": "v", "type": "inline_equation" }, { "bbox": [ 288, 614, 505, 627 ], "score": 1.0, "content": ", an i.i.d Gaussian noise tensor of the same dimensions", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 625, 505, 638 ], "spans": [ { "bbox": [ 105, 625, 172, 638 ], "score": 1.0, "content": "as the video as", "type": "text" }, { "bbox": [ 172, 625, 225, 637 ], "score": 0.92, "content": "\\epsilon \\sim N ( 0 , { \\bf I } )", "type": "inline_equation" }, { "bbox": [ 225, 625, 344, 638 ], "score": 1.0, "content": ", and the noise level at time", "type": "text" }, { "bbox": [ 344, 627, 351, 635 ], "score": 0.67, "content": "s", "type": "inline_equation" }, { "bbox": [ 351, 625, 364, 638 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 364, 627, 375, 636 ], "score": 0.84, "content": "\\sigma _ { s }", "type": "inline_equation" }, { "bbox": [ 376, 625, 505, 638 ], "score": 1.0, "content": ". The noisy video is given by:", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 636, 504, 651 ], "spans": [ { "bbox": [ 106, 639, 172, 650 ], "score": 0.89, "content": "z _ { s } = \\gamma _ { s } v + \\sigma _ { s } \\epsilon", "type": "inline_equation" }, { "bbox": [ 172, 637, 203, 651 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 203, 636, 265, 650 ], "score": 0.93, "content": "\\gamma _ { s } = \\sqrt { 1 - \\sigma _ { s } ^ { 2 } }", "type": "inline_equation" }, { "bbox": [ 266, 637, 498, 651 ], "score": 1.0, "content": ". Furthermore, let us denote a conditioning text prompt as", "type": "text" }, { "bbox": [ 499, 639, 504, 648 ], "score": 0.65, "content": "t", "type": "inline_equation" } ], "index": 36 }, { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 209, 662 ], "score": 1.0, "content": "and a conditioning video", "type": "text" }, { "bbox": [ 209, 651, 215, 659 ], "score": 0.74, "content": "c", "type": "inline_equation" }, { "bbox": [ 216, 648, 304, 662 ], "score": 1.0, "content": "(for super-resolution,", "type": "text" }, { "bbox": [ 305, 651, 311, 659 ], "score": 0.75, "content": "c", "type": "inline_equation" }, { "bbox": [ 311, 648, 433, 662 ], "score": 1.0, "content": "is a low-resolution version of", "type": "text" }, { "bbox": [ 433, 651, 439, 659 ], "score": 0.62, "content": "v", "type": "inline_equation" }, { "bbox": [ 439, 648, 506, 662 ], "score": 1.0, "content": "). The objective", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 658, 506, 674 ], "spans": [ { "bbox": [ 105, 658, 209, 674 ], "score": 1.0, "content": "of the denoising network", "type": "text" }, { "bbox": [ 210, 660, 223, 671 ], "score": 0.89, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 224, 658, 370, 674 ], "score": 1.0, "content": "is to recover the ground truth video", "type": "text" }, { "bbox": [ 371, 662, 378, 670 ], "score": 0.63, "content": "v", "type": "inline_equation" }, { "bbox": [ 378, 658, 491, 674 ], "score": 1.0, "content": "given the noisy input video", "type": "text" }, { "bbox": [ 491, 662, 501, 671 ], "score": 0.82, "content": "z _ { s }", "type": "inline_equation" }, { "bbox": [ 501, 658, 506, 674 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 671, 504, 684 ], "spans": [ { "bbox": [ 106, 671, 142, 684 ], "score": 1.0, "content": "the time", "type": "text" }, { "bbox": [ 142, 673, 148, 681 ], "score": 0.68, "content": "s", "type": "inline_equation" }, { "bbox": [ 148, 671, 183, 684 ], "score": 1.0, "content": ", prompt", "type": "text" }, { "bbox": [ 184, 672, 189, 681 ], "score": 0.76, "content": "t", "type": "inline_equation" }, { "bbox": [ 189, 671, 286, 684 ], "score": 1.0, "content": "and conditioning video", "type": "text" }, { "bbox": [ 286, 674, 291, 680 ], "score": 0.71, "content": "c", "type": "inline_equation" }, { "bbox": [ 292, 671, 495, 684 ], "score": 1.0, "content": ". The model is trained on a (large) training corpus", "type": "text" }, { "bbox": [ 496, 672, 504, 681 ], "score": 0.78, "content": "\\nu", "type": "inline_equation" } ], "index": 39 }, { "bbox": [ 106, 682, 303, 694 ], "spans": [ { "bbox": [ 106, 682, 217, 694 ], "score": 1.0, "content": "consisting of pairs of video", "type": "text" }, { "bbox": [ 218, 684, 224, 691 ], "score": 0.77, "content": "v", "type": "inline_equation" }, { "bbox": [ 224, 682, 295, 694 ], "score": 1.0, "content": "and text prompts", "type": "text" }, { "bbox": [ 295, 684, 299, 691 ], "score": 0.82, "content": "t", "type": "inline_equation" }, { "bbox": [ 299, 682, 303, 694 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40 } ], "index": 36.5, "bbox_fs": [ 105, 602, 506, 694 ] }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 698, 506, 712 ], "spans": [ { "bbox": [ 106, 698, 506, 712 ], "score": 1.0, "content": "Sampling from Diffusion Models. The key challenge in diffusion models is to use the denoiser", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 142, 722 ], "score": 1.0, "content": "network", "type": "text" }, { "bbox": [ 142, 710, 156, 721 ], "score": 0.89, "content": "D _ { \\theta }", "type": "inline_equation" }, { "bbox": [ 156, 710, 453, 722 ], "score": 1.0, "content": "to sample from the distribution of videos conditioned on the text prompt", "type": "text" }, { "bbox": [ 453, 711, 458, 720 ], "score": 0.72, "content": "t", "type": "inline_equation" }, { "bbox": [ 459, 710, 505, 722 ], "score": 1.0, "content": "and condi-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 721, 506, 733 ], "spans": [ { "bbox": [ 106, 721, 162, 733 ], "score": 1.0, "content": "tioning video", "type": "text" }, { "bbox": [ 162, 722, 168, 731 ], "score": 0.43, "content": "c", "type": "inline_equation" }, { "bbox": [ 168, 721, 171, 733 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 171, 721, 208, 732 ], "score": 0.92, "content": "P ( v | t , c )", "type": "inline_equation" }, { "bbox": [ 208, 721, 506, 733 ], "score": 1.0, "content": ". While the derivation of such sampling rule is non-trivial (see e.g. Karras", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 260, 505, 272 ], "spans": [ { "bbox": [ 105, 260, 505, 272 ], "score": 1.0, "content": "et al. (2022)), the implementation of such sampling is relatively simple in practice. We follow Ho", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 105, 270, 505, 284 ], "spans": [ { "bbox": [ 105, 270, 505, 284 ], "score": 1.0, "content": "et al. (2022a) in using stochastic DDIM sampling. At a heuristic level, at each step, we first use the", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 106, 282, 505, 294 ], "spans": [ { "bbox": [ 106, 282, 505, 294 ], "score": 1.0, "content": "denoiser network to estimate the noise. We then remove a fraction of the estimated noise and fi-", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 293, 505, 306 ], "spans": [ { "bbox": [ 105, 293, 505, 306 ], "score": 1.0, "content": "nally add randomly generated Gaussian noise, with magnitude corresponding to half of the removed", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 303, 134, 317 ], "spans": [ { "bbox": [ 105, 303, 134, 317 ], "score": 1.0, "content": "noise.", "type": "text", "cross_page": true } ], "index": 11 } ], "index": 42, "bbox_fs": [ 106, 698, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 108, 92, 502, 194 ], "blocks": [ { "type": "image_caption", "bbox": [ 279, 81, 331, 92 ], "group_id": 0, "lines": [ { "bbox": [ 279, 79, 333, 95 ], "spans": [ { "bbox": [ 279, 79, 333, 95 ], "score": 1.0, "content": "Input Video", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 108, 92, 502, 194 ], "group_id": 0, "lines": [ { "bbox": [ 108, 92, 502, 194 ], "spans": [ { "bbox": [ 108, 92, 502, 194 ], "score": 0.967, "type": "image", "image_path": "aa7aef3582ee7697ccb8bb2ff1f9d931e9a0f0f65e6c26b3afb8f47276346237.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 92, 502, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 126.0, 502, 160.0 ], "spans": [], "index": 2 }, { "bbox": [ 108, 160.0, 502, 194.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 107, 204, 505, 237 ], "group_id": 0, "lines": [ { "bbox": [ 105, 204, 505, 217 ], "spans": [ { "bbox": [ 105, 204, 505, 217 ], "score": 1.0, "content": "Figure 3: Video Motion Editing: Dreamix can significantly change the actions and motions of sub-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 214, 505, 229 ], "spans": [ { "bbox": [ 105, 214, 505, 229 ], "score": 1.0, "content": "jects in a video (e.g. making a puppy leap) while maintaining temporal consistency and preserving", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 226, 187, 237 ], "spans": [ { "bbox": [ 106, 226, 187, 237 ], "score": 1.0, "content": "the unedited details", "type": "text" } ], "index": 6 } ], "index": 5 } ], "index": 2 }, { "type": "text", "bbox": [ 107, 260, 505, 315 ], "lines": [ { "bbox": [ 105, 260, 505, 272 ], "spans": [ { "bbox": [ 105, 260, 505, 272 ], "score": 1.0, "content": "et al. (2022)), the implementation of such sampling is relatively simple in practice. We follow Ho", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 270, 505, 284 ], "spans": [ { "bbox": [ 105, 270, 505, 284 ], "score": 1.0, "content": "et al. (2022a) in using stochastic DDIM sampling. At a heuristic level, at each step, we first use the", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 282, 505, 294 ], "spans": [ { "bbox": [ 106, 282, 505, 294 ], "score": 1.0, "content": "denoiser network to estimate the noise. We then remove a fraction of the estimated noise and fi-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 293, 505, 306 ], "spans": [ { "bbox": [ 105, 293, 505, 306 ], "score": 1.0, "content": "nally add randomly generated Gaussian noise, with magnitude corresponding to half of the removed", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 303, 134, 317 ], "spans": [ { "bbox": [ 105, 303, 134, 317 ], "score": 1.0, "content": "noise.", "type": "text" } ], "index": 11 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 321, 505, 387 ], "lines": [ { "bbox": [ 106, 321, 504, 333 ], "spans": [ { "bbox": [ 106, 321, 504, 333 ], "score": 1.0, "content": "Cascaded Video Diffusion Models. Training high-resolution text-to-video models is very chal-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 331, 506, 346 ], "spans": [ { "bbox": [ 105, 331, 506, 346 ], "score": 1.0, "content": "lenging due to the high computational complexity. Several diffusion models overcome this by using", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 343, 506, 355 ], "spans": [ { "bbox": [ 105, 343, 506, 355 ], "score": 1.0, "content": "cascaded architectures. We use a model that follows the architecture of Ho et al. (2022a), which", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 354, 505, 367 ], "spans": [ { "bbox": [ 105, 354, 505, 367 ], "score": 1.0, "content": "consists of a cascade of 7 models. The base model maps the input text prompt into a 5-second video", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 364, 505, 378 ], "spans": [ { "bbox": [ 105, 364, 118, 378 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 118, 365, 176, 376 ], "score": 0.91, "content": "2 4 \\times 4 0 \\times 1 6", "type": "inline_equation" }, { "bbox": [ 177, 364, 505, 378 ], "score": 1.0, "content": "frames. It is then followed by 3 spatial super-resolution models and 3 temporal", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 376, 388, 388 ], "spans": [ { "bbox": [ 105, 376, 388, 388 ], "score": 1.0, "content": "super-resolution models. For implementation details, see Appendix C.", "type": "text" } ], "index": 17 } ], "index": 14.5 }, { "type": "title", "bbox": [ 108, 405, 333, 417 ], "lines": [ { "bbox": [ 104, 404, 334, 420 ], "spans": [ { "bbox": [ 104, 404, 334, 420 ], "score": 1.0, "content": "4 EDITING BY VIDEO DIFFUSION MODELS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 430, 504, 453 ], "lines": [ { "bbox": [ 106, 430, 505, 443 ], "spans": [ { "bbox": [ 106, 430, 505, 443 ], "score": 1.0, "content": "We propose a new method for video editing using text-guided video diffusion models. We extended", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 441, 234, 454 ], "spans": [ { "bbox": [ 105, 441, 234, 454 ], "score": 1.0, "content": "it to image animation in Sec. 5.", "type": "text" } ], "index": 20 } ], "index": 19.5 }, { "type": "title", "bbox": [ 106, 468, 329, 479 ], "lines": [ { "bbox": [ 106, 468, 330, 480 ], "spans": [ { "bbox": [ 106, 468, 330, 480 ], "score": 1.0, "content": "4.1 VIDEO EDITING BY INVERTING CORRUPTIONS", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 489, 505, 599 ], "lines": [ { "bbox": [ 106, 489, 505, 502 ], "spans": [ { "bbox": [ 106, 489, 376, 502 ], "score": 1.0, "content": "We wish to edit an input video using the guidance of a text prompt", "type": "text" }, { "bbox": [ 377, 491, 382, 499 ], "score": 0.64, "content": "t", "type": "inline_equation" }, { "bbox": [ 383, 489, 505, 502 ], "score": 1.0, "content": "describing the video after the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "edit. In order to do so we leverage the power of a cascade of VDMs. The key idea is to first corrupt", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 511, 505, 524 ], "spans": [ { "bbox": [ 105, 511, 505, 524 ], "score": 1.0, "content": "the video by downsampling followed by adding noise. We then apply the sampling process of the", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 522, 506, 535 ], "spans": [ { "bbox": [ 105, 522, 496, 535 ], "score": 1.0, "content": "cascaded diffusion models from the time step corresponding to the noise level, conditioned on", "type": "text" }, { "bbox": [ 496, 523, 502, 533 ], "score": 0.6, "content": "t .", "type": "inline_equation" }, { "bbox": [ 502, 522, 506, 535 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 533, 505, 545 ], "spans": [ { "bbox": [ 105, 533, 505, 545 ], "score": 1.0, "content": "which upscales the video to the final spatio-temporal resolution. The effect is that the VDM will", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 544, 505, 557 ], "spans": [ { "bbox": [ 105, 544, 505, 557 ], "score": 1.0, "content": "use the low-resolution details provided by the degraded input video, but synthesize new high spatio-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 555, 506, 568 ], "spans": [ { "bbox": [ 105, 555, 506, 568 ], "score": 1.0, "content": "temporal resolution information using the text prompt guidance. While this procedure is essentially a", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 565, 506, 579 ], "spans": [ { "bbox": [ 105, 565, 506, 579 ], "score": 1.0, "content": "text-guided version of SDEdit Meng et al. (2021), for complex edits e.g., motion editing this by itself", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 575, 505, 591 ], "spans": [ { "bbox": [ 105, 575, 505, 591 ], "score": 1.0, "content": "does not result in sufficiently high-fidelity videos. To mitigate this issue, we use a mixed-finetuning", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 588, 232, 600 ], "spans": [ { "bbox": [ 105, 588, 232, 600 ], "score": 1.0, "content": "objective described in Sec. 4.2.", "type": "text" } ], "index": 31 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 605, 505, 671 ], "lines": [ { "bbox": [ 106, 605, 505, 616 ], "spans": [ { "bbox": [ 106, 605, 505, 616 ], "score": 1.0, "content": "Input Video Degradation. We downsample the input video to the resolution of the base model (16", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 615, 506, 628 ], "spans": [ { "bbox": [ 105, 615, 147, 628 ], "score": 1.0, "content": "frames of", "type": "text" }, { "bbox": [ 147, 616, 182, 627 ], "score": 0.89, "content": "2 4 \\times 4 0", "type": "inline_equation" }, { "bbox": [ 182, 615, 381, 628 ], "score": 1.0, "content": "). We then add i.i.d Gaussian noise with variance", "type": "text" }, { "bbox": [ 381, 615, 393, 628 ], "score": 0.9, "content": "\\sigma _ { s } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 394, 615, 506, 628 ], "score": 1.0, "content": "to further corrupt the input", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 627, 505, 639 ], "spans": [ { "bbox": [ 106, 627, 301, 639 ], "score": 1.0, "content": "video. The noise strength is equivalent to time", "type": "text" }, { "bbox": [ 301, 629, 307, 637 ], "score": 0.37, "content": "s", "type": "inline_equation" }, { "bbox": [ 308, 627, 505, 639 ], "score": 1.0, "content": "in the diffusion process of the base model. For", "type": "text" } ], "index": 34 }, { "bbox": [ 107, 638, 506, 650 ], "spans": [ { "bbox": [ 107, 639, 131, 648 ], "score": 0.87, "content": "s = 0", "type": "inline_equation" }, { "bbox": [ 132, 638, 249, 650 ], "score": 1.0, "content": ", no noise is added, while for", "type": "text" }, { "bbox": [ 249, 639, 274, 648 ], "score": 0.9, "content": "s = 1", "type": "inline_equation" }, { "bbox": [ 274, 638, 506, 650 ], "score": 1.0, "content": ", the video is replaced by pure Gaussian noise. Note, that", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 649, 505, 662 ], "spans": [ { "bbox": [ 105, 649, 505, 662 ], "score": 1.0, "content": "even when no noise is added, the input video is highly corrupted due to the extreme downsampling", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 659, 130, 672 ], "spans": [ { "bbox": [ 105, 659, 130, 672 ], "score": 1.0, "content": "ratio.", "type": "text" } ], "index": 37 } ], "index": 34.5 }, { "type": "text", "bbox": [ 107, 677, 504, 732 ], "lines": [ { "bbox": [ 106, 676, 505, 689 ], "spans": [ { "bbox": [ 106, 676, 505, 689 ], "score": 1.0, "content": "Text-Guided Corruption Inversion. We can now use the cascaded VDMs to map the corrupted,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "low-resolution video into a high-resolution video that aligns with the text. The core idea here is", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "that given a noisy, very low spatio-temporal resolution video, there are many perfectly feasible,", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 412, 723 ], "score": 1.0, "content": "high-resolution videos that correspond to it. We use the target text prompt", "type": "text" }, { "bbox": [ 413, 711, 418, 720 ], "score": 0.71, "content": "t", "type": "inline_equation" }, { "bbox": [ 418, 709, 506, 723 ], "score": 1.0, "content": "to select the feasible", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 720, 506, 734 ], "spans": [ { "bbox": [ 106, 720, 506, 734 ], "score": 1.0, "content": "outputs that not only correspond to the low-resolution of the original video but are also aligned to", "type": "text" } ], "index": 42 } ], "index": 40 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 759 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 108, 92, 502, 194 ], "blocks": [ { "type": "image_caption", "bbox": [ 279, 81, 331, 92 ], "group_id": 0, "lines": [ { "bbox": [ 279, 79, 333, 95 ], "spans": [ { "bbox": [ 279, 79, 333, 95 ], "score": 1.0, "content": "Input Video", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 108, 92, 502, 194 ], "group_id": 0, "lines": [ { "bbox": [ 108, 92, 502, 194 ], "spans": [ { "bbox": [ 108, 92, 502, 194 ], "score": 0.967, "type": "image", "image_path": "aa7aef3582ee7697ccb8bb2ff1f9d931e9a0f0f65e6c26b3afb8f47276346237.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 92, 502, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 126.0, 502, 160.0 ], "spans": [], "index": 2 }, { "bbox": [ 108, 160.0, 502, 194.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 107, 204, 505, 237 ], "group_id": 0, "lines": [ { "bbox": [ 105, 204, 505, 217 ], "spans": [ { "bbox": [ 105, 204, 505, 217 ], "score": 1.0, "content": "Figure 3: Video Motion Editing: Dreamix can significantly change the actions and motions of sub-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 214, 505, 229 ], "spans": [ { "bbox": [ 105, 214, 505, 229 ], "score": 1.0, "content": "jects in a video (e.g. making a puppy leap) while maintaining temporal consistency and preserving", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 226, 187, 237 ], "spans": [ { "bbox": [ 106, 226, 187, 237 ], "score": 1.0, "content": "the unedited details", "type": "text" } ], "index": 6 } ], "index": 5 } ], "index": 2 }, { "type": "text", "bbox": [ 107, 260, 505, 315 ], "lines": [], "index": 9, "bbox_fs": [ 105, 260, 505, 317 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 321, 505, 387 ], "lines": [ { "bbox": [ 106, 321, 504, 333 ], "spans": [ { "bbox": [ 106, 321, 504, 333 ], "score": 1.0, "content": "Cascaded Video Diffusion Models. Training high-resolution text-to-video models is very chal-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 331, 506, 346 ], "spans": [ { "bbox": [ 105, 331, 506, 346 ], "score": 1.0, "content": "lenging due to the high computational complexity. Several diffusion models overcome this by using", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 343, 506, 355 ], "spans": [ { "bbox": [ 105, 343, 506, 355 ], "score": 1.0, "content": "cascaded architectures. We use a model that follows the architecture of Ho et al. (2022a), which", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 354, 505, 367 ], "spans": [ { "bbox": [ 105, 354, 505, 367 ], "score": 1.0, "content": "consists of a cascade of 7 models. The base model maps the input text prompt into a 5-second video", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 364, 505, 378 ], "spans": [ { "bbox": [ 105, 364, 118, 378 ], "score": 1.0, "content": "of", "type": "text" }, { "bbox": [ 118, 365, 176, 376 ], "score": 0.91, "content": "2 4 \\times 4 0 \\times 1 6", "type": "inline_equation" }, { "bbox": [ 177, 364, 505, 378 ], "score": 1.0, "content": "frames. It is then followed by 3 spatial super-resolution models and 3 temporal", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 376, 388, 388 ], "spans": [ { "bbox": [ 105, 376, 388, 388 ], "score": 1.0, "content": "super-resolution models. For implementation details, see Appendix C.", "type": "text" } ], "index": 17 } ], "index": 14.5, "bbox_fs": [ 105, 321, 506, 388 ] }, { "type": "title", "bbox": [ 108, 405, 333, 417 ], "lines": [ { "bbox": [ 104, 404, 334, 420 ], "spans": [ { "bbox": [ 104, 404, 334, 420 ], "score": 1.0, "content": "4 EDITING BY VIDEO DIFFUSION MODELS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 430, 504, 453 ], "lines": [ { "bbox": [ 106, 430, 505, 443 ], "spans": [ { "bbox": [ 106, 430, 505, 443 ], "score": 1.0, "content": "We propose a new method for video editing using text-guided video diffusion models. We extended", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 441, 234, 454 ], "spans": [ { "bbox": [ 105, 441, 234, 454 ], "score": 1.0, "content": "it to image animation in Sec. 5.", "type": "text" } ], "index": 20 } ], "index": 19.5, "bbox_fs": [ 105, 430, 505, 454 ] }, { "type": "title", "bbox": [ 106, 468, 329, 479 ], "lines": [ { "bbox": [ 106, 468, 330, 480 ], "spans": [ { "bbox": [ 106, 468, 330, 480 ], "score": 1.0, "content": "4.1 VIDEO EDITING BY INVERTING CORRUPTIONS", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 489, 505, 599 ], "lines": [ { "bbox": [ 106, 489, 505, 502 ], "spans": [ { "bbox": [ 106, 489, 376, 502 ], "score": 1.0, "content": "We wish to edit an input video using the guidance of a text prompt", "type": "text" }, { "bbox": [ 377, 491, 382, 499 ], "score": 0.64, "content": "t", "type": "inline_equation" }, { "bbox": [ 383, 489, 505, 502 ], "score": 1.0, "content": "describing the video after the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "edit. In order to do so we leverage the power of a cascade of VDMs. The key idea is to first corrupt", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 511, 505, 524 ], "spans": [ { "bbox": [ 105, 511, 505, 524 ], "score": 1.0, "content": "the video by downsampling followed by adding noise. We then apply the sampling process of the", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 522, 506, 535 ], "spans": [ { "bbox": [ 105, 522, 496, 535 ], "score": 1.0, "content": "cascaded diffusion models from the time step corresponding to the noise level, conditioned on", "type": "text" }, { "bbox": [ 496, 523, 502, 533 ], "score": 0.6, "content": "t .", "type": "inline_equation" }, { "bbox": [ 502, 522, 506, 535 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 533, 505, 545 ], "spans": [ { "bbox": [ 105, 533, 505, 545 ], "score": 1.0, "content": "which upscales the video to the final spatio-temporal resolution. The effect is that the VDM will", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 544, 505, 557 ], "spans": [ { "bbox": [ 105, 544, 505, 557 ], "score": 1.0, "content": "use the low-resolution details provided by the degraded input video, but synthesize new high spatio-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 555, 506, 568 ], "spans": [ { "bbox": [ 105, 555, 506, 568 ], "score": 1.0, "content": "temporal resolution information using the text prompt guidance. While this procedure is essentially a", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 565, 506, 579 ], "spans": [ { "bbox": [ 105, 565, 506, 579 ], "score": 1.0, "content": "text-guided version of SDEdit Meng et al. (2021), for complex edits e.g., motion editing this by itself", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 575, 505, 591 ], "spans": [ { "bbox": [ 105, 575, 505, 591 ], "score": 1.0, "content": "does not result in sufficiently high-fidelity videos. To mitigate this issue, we use a mixed-finetuning", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 588, 232, 600 ], "spans": [ { "bbox": [ 105, 588, 232, 600 ], "score": 1.0, "content": "objective described in Sec. 4.2.", "type": "text" } ], "index": 31 } ], "index": 26.5, "bbox_fs": [ 105, 489, 506, 600 ] }, { "type": "text", "bbox": [ 107, 605, 505, 671 ], "lines": [ { "bbox": [ 106, 605, 505, 616 ], "spans": [ { "bbox": [ 106, 605, 505, 616 ], "score": 1.0, "content": "Input Video Degradation. We downsample the input video to the resolution of the base model (16", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 615, 506, 628 ], "spans": [ { "bbox": [ 105, 615, 147, 628 ], "score": 1.0, "content": "frames of", "type": "text" }, { "bbox": [ 147, 616, 182, 627 ], "score": 0.89, "content": "2 4 \\times 4 0", "type": "inline_equation" }, { "bbox": [ 182, 615, 381, 628 ], "score": 1.0, "content": "). We then add i.i.d Gaussian noise with variance", "type": "text" }, { "bbox": [ 381, 615, 393, 628 ], "score": 0.9, "content": "\\sigma _ { s } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 394, 615, 506, 628 ], "score": 1.0, "content": "to further corrupt the input", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 627, 505, 639 ], "spans": [ { "bbox": [ 106, 627, 301, 639 ], "score": 1.0, "content": "video. The noise strength is equivalent to time", "type": "text" }, { "bbox": [ 301, 629, 307, 637 ], "score": 0.37, "content": "s", "type": "inline_equation" }, { "bbox": [ 308, 627, 505, 639 ], "score": 1.0, "content": "in the diffusion process of the base model. For", "type": "text" } ], "index": 34 }, { "bbox": [ 107, 638, 506, 650 ], "spans": [ { "bbox": [ 107, 639, 131, 648 ], "score": 0.87, "content": "s = 0", "type": "inline_equation" }, { "bbox": [ 132, 638, 249, 650 ], "score": 1.0, "content": ", no noise is added, while for", "type": "text" }, { "bbox": [ 249, 639, 274, 648 ], "score": 0.9, "content": "s = 1", "type": "inline_equation" }, { "bbox": [ 274, 638, 506, 650 ], "score": 1.0, "content": ", the video is replaced by pure Gaussian noise. Note, that", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 649, 505, 662 ], "spans": [ { "bbox": [ 105, 649, 505, 662 ], "score": 1.0, "content": "even when no noise is added, the input video is highly corrupted due to the extreme downsampling", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 659, 130, 672 ], "spans": [ { "bbox": [ 105, 659, 130, 672 ], "score": 1.0, "content": "ratio.", "type": "text" } ], "index": 37 } ], "index": 34.5, "bbox_fs": [ 105, 605, 506, 672 ] }, { "type": "text", "bbox": [ 107, 677, 504, 732 ], "lines": [ { "bbox": [ 106, 676, 505, 689 ], "spans": [ { "bbox": [ 106, 676, 505, 689 ], "score": 1.0, "content": "Text-Guided Corruption Inversion. We can now use the cascaded VDMs to map the corrupted,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "low-resolution video into a high-resolution video that aligns with the text. The core idea here is", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "that given a noisy, very low spatio-temporal resolution video, there are many perfectly feasible,", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 412, 723 ], "score": 1.0, "content": "high-resolution videos that correspond to it. We use the target text prompt", "type": "text" }, { "bbox": [ 413, 711, 418, 720 ], "score": 0.71, "content": "t", "type": "inline_equation" }, { "bbox": [ 418, 709, 506, 723 ], "score": 1.0, "content": "to select the feasible", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 720, 506, 734 ], "spans": [ { "bbox": [ 106, 720, 506, 734 ], "score": 1.0, "content": "outputs that not only correspond to the low-resolution of the original video but are also aligned to", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 294, 505, 306 ], "spans": [ { "bbox": [ 105, 294, 505, 306 ], "score": 1.0, "content": "edits desired by the user. The base model starts with the corrupted video, which has the same noise", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 304, 505, 318 ], "spans": [ { "bbox": [ 105, 304, 235, 318 ], "score": 1.0, "content": "as the diffusion process at time", "type": "text", "cross_page": true }, { "bbox": [ 235, 307, 241, 315 ], "score": 0.64, "content": "s", "type": "inline_equation", "cross_page": true }, { "bbox": [ 241, 304, 505, 318 ], "score": 1.0, "content": ". We use the model to reverse the diffusion process up to time 0.", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 315, 505, 329 ], "spans": [ { "bbox": [ 105, 315, 505, 329 ], "score": 1.0, "content": "We then upscale the video through the entire cascade of super-resolution models (see Appendix C).", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 326, 284, 339 ], "spans": [ { "bbox": [ 105, 326, 275, 339 ], "score": 1.0, "content": "All models are conditioned on the prompt", "type": "text", "cross_page": true }, { "bbox": [ 276, 328, 280, 337 ], "score": 0.68, "content": "t", "type": "inline_equation", "cross_page": true }, { "bbox": [ 281, 326, 284, 339 ], "score": 1.0, "content": ".", "type": "text", "cross_page": true } ], "index": 11 } ], "index": 40, "bbox_fs": [ 105, 676, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 151, 83, 458, 209 ], "blocks": [ { "type": "image_body", "bbox": [ 151, 83, 458, 209 ], "group_id": 0, "lines": [ { "bbox": [ 151, 83, 458, 209 ], "spans": [ { "bbox": [ 151, 83, 458, 209 ], "score": 0.973, "type": "image", "image_path": "12e0623b2a593ab2c48d73176d55b0335fa925ea264e914d3a903c218e455fb9.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 151, 83, 458, 125.0 ], "spans": [], "index": 0 }, { "bbox": [ 151, 125.0, 458, 167.0 ], "spans": [], "index": 1 }, { "bbox": [ 151, 167.0, 458, 209.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 217, 506, 273 ], "group_id": 0, "lines": [ { "bbox": [ 105, 217, 506, 231 ], "spans": [ { "bbox": [ 105, 217, 506, 231 ], "score": 1.0, "content": "Figure 4: Mixed Video-Image Finetuning: Finetuning the VDM on the input video alone limits", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 229, 505, 241 ], "spans": [ { "bbox": [ 106, 229, 505, 241 ], "score": 1.0, "content": "the extent of motion change. Instead, we use a mixed objective that beside the original objective", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 239, 506, 252 ], "spans": [ { "bbox": [ 105, 239, 506, 252 ], "score": 1.0, "content": "(bottom left) also finetunes on the unordered set of frames. We use “masked temporal attention“ to", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 250, 506, 264 ], "spans": [ { "bbox": [ 105, 250, 506, 264 ], "score": 1.0, "content": "prevent the temporal attention and convolution from changing (bottom right). This allows adding", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 263, 202, 273 ], "spans": [ { "bbox": [ 106, 263, 202, 273 ], "score": 1.0, "content": "motion to a static video", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 293, 505, 339 ], "lines": [ { "bbox": [ 105, 294, 505, 306 ], "spans": [ { "bbox": [ 105, 294, 505, 306 ], "score": 1.0, "content": "edits desired by the user. The base model starts with the corrupted video, which has the same noise", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 304, 505, 318 ], "spans": [ { "bbox": [ 105, 304, 235, 318 ], "score": 1.0, "content": "as the diffusion process at time", "type": "text" }, { "bbox": [ 235, 307, 241, 315 ], "score": 0.64, "content": "s", "type": "inline_equation" }, { "bbox": [ 241, 304, 505, 318 ], "score": 1.0, "content": ". We use the model to reverse the diffusion process up to time 0.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 315, 505, 329 ], "spans": [ { "bbox": [ 105, 315, 505, 329 ], "score": 1.0, "content": "We then upscale the video through the entire cascade of super-resolution models (see Appendix C).", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 326, 284, 339 ], "spans": [ { "bbox": [ 105, 326, 275, 339 ], "score": 1.0, "content": "All models are conditioned on the prompt", "type": "text" }, { "bbox": [ 276, 328, 280, 337 ], "score": 0.68, "content": "t", "type": "inline_equation" }, { "bbox": [ 281, 326, 284, 339 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 11 } ], "index": 9.5 }, { "type": "title", "bbox": [ 108, 351, 283, 363 ], "lines": [ { "bbox": [ 106, 351, 284, 364 ], "spans": [ { "bbox": [ 106, 351, 284, 364 ], "score": 1.0, "content": "4.2 MIXED VIDEO-IMAGE FINETUNING", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 372, 505, 483 ], "lines": [ { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "The naive method presented in Sec. 4.1 relies on a corrupted version of the input video which does", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 383, 506, 396 ], "spans": [ { "bbox": [ 105, 383, 506, 396 ], "score": 1.0, "content": "not include enough information to preserve high-resolution details such as fine textures or object", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 506, 407 ], "score": 1.0, "content": "identity. We tackle this by adding a preliminary stage of finetuning the model on the input video", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 403, 506, 419 ], "spans": [ { "bbox": [ 106, 407, 113, 415 ], "score": 0.6, "content": "v", "type": "inline_equation" }, { "bbox": [ 113, 403, 506, 419 ], "score": 1.0, "content": ". Note that this only needs to be done once for the video, which can then be edited by many", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 416, 506, 428 ], "spans": [ { "bbox": [ 105, 416, 506, 428 ], "score": 1.0, "content": "prompts without further finetuning. We would like the model to separately update its prior both on", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 427, 504, 438 ], "spans": [ { "bbox": [ 106, 427, 504, 438 ], "score": 1.0, "content": "the appearance and the motion of the input video. Our approach therefore treats the input video,", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 437, 505, 451 ], "spans": [ { "bbox": [ 105, 437, 323, 451 ], "score": 1.0, "content": "both as a single video clip and as an unordered set of", "type": "text" }, { "bbox": [ 324, 438, 335, 448 ], "score": 0.82, "content": "M", "type": "inline_equation" }, { "bbox": [ 336, 437, 416, 451 ], "score": 1.0, "content": "frames, denoted by", "type": "text" }, { "bbox": [ 417, 438, 501, 450 ], "score": 0.94, "content": "\\boldsymbol { u } = \\{ x _ { 1 } , x _ { 2 } , . . , x _ { M } \\}", "type": "inline_equation" }, { "bbox": [ 501, 437, 505, 451 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 447, 506, 463 ], "spans": [ { "bbox": [ 105, 447, 189, 463 ], "score": 1.0, "content": "We use a rare string", "type": "text" }, { "bbox": [ 190, 450, 199, 459 ], "score": 0.83, "content": "t ^ { * }", "type": "inline_equation" }, { "bbox": [ 199, 447, 506, 463 ], "score": 1.0, "content": "as the text prompt, following Ruiz et al. (2022). We finetune the denoising", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 460, 505, 472 ], "spans": [ { "bbox": [ 105, 460, 505, 472 ], "score": 1.0, "content": "models by a combination of two objectives. The first objective updates the model prior on both", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 470, 498, 484 ], "spans": [ { "bbox": [ 105, 470, 380, 484 ], "score": 1.0, "content": "motion and appearance by requiring it to reconstruct the input video", "type": "text" }, { "bbox": [ 380, 473, 387, 481 ], "score": 0.77, "content": "v", "type": "inline_equation" }, { "bbox": [ 387, 470, 484, 484 ], "score": 1.0, "content": "given its noisy versions", "type": "text" }, { "bbox": [ 484, 473, 493, 482 ], "score": 0.79, "content": "z _ { s }", "type": "inline_equation" }, { "bbox": [ 494, 470, 498, 484 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 22 } ], "index": 17.5 }, { "type": "interline_equation", "bbox": [ 198, 496, 412, 512 ], "lines": [ { "bbox": [ 198, 496, 412, 512 ], "spans": [ { "bbox": [ 198, 496, 412, 512 ], "score": 0.9, "content": "\\mathcal { L } _ { \\theta } ^ { v i d } ( v ) = \\mathbb { E } _ { \\epsilon \\sim N ( 0 , \\mathbf { I } ) , s \\in \\mathcal { U } ( 0 , 1 ) } \\Vert D _ { \\theta ^ { \\prime } } ( z _ { s } , s , t ^ { * } , c ) - v \\Vert ^ { 2 }", "type": "interline_equation", "image_path": "f3b76add60efaeb879d08b45df45ad23a9cb0ac3e703e38d3102542af9ac7a09.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 198, 496, 412, 512 ], "spans": [], "index": 23 } ] }, { "type": "text", "bbox": [ 107, 518, 504, 585 ], "lines": [ { "bbox": [ 105, 518, 505, 532 ], "spans": [ { "bbox": [ 105, 518, 505, 532 ], "score": 1.0, "content": "Additionally, we train the model to reconstruct each of the frames individually given their noisy", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 530, 505, 543 ], "spans": [ { "bbox": [ 105, 530, 505, 543 ], "score": 1.0, "content": "version. This enhances the appearance prior of the model, separately from the motion. Technically,", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 541, 506, 554 ], "spans": [ { "bbox": [ 105, 541, 288, 554 ], "score": 1.0, "content": "the model is trained on a sequence of frames", "type": "text" }, { "bbox": [ 289, 543, 295, 551 ], "score": 0.75, "content": "u", "type": "inline_equation" }, { "bbox": [ 296, 541, 506, 554 ], "score": 1.0, "content": "by replacing the temporal attention layers by trivial", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "fixed masks ensuring the model only pays attention within each frame, and also by masking the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 562, 506, 577 ], "spans": [ { "bbox": [ 105, 562, 468, 577 ], "score": 1.0, "content": "residual temporal convolution blocks. We denote the attention masked denoising model as", "type": "text" }, { "bbox": [ 468, 563, 482, 575 ], "score": 0.88, "content": "D _ { \\theta } ^ { a }", "type": "inline_equation" }, { "bbox": [ 483, 562, 506, 577 ], "score": 1.0, "content": ". The", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 575, 227, 586 ], "spans": [ { "bbox": [ 105, 575, 227, 586 ], "score": 1.0, "content": "masked attention objective is:", "type": "text" } ], "index": 29 } ], "index": 26.5 }, { "type": "interline_equation", "bbox": [ 191, 589, 420, 606 ], "lines": [ { "bbox": [ 191, 589, 420, 606 ], "spans": [ { "bbox": [ 191, 589, 420, 606 ], "score": 0.9, "content": "\\mathcal { L } _ { \\theta } ^ { f r a m e } ( u ) = \\mathbb { E } _ { \\epsilon \\sim N ( 0 , \\mathbf { I } ) , s \\in \\mathcal { U } ( 0 , 1 ) } \\Vert D _ { \\theta ^ { \\prime } } ^ { a } ( z _ { s } , s , t ^ { * } , c ) - u \\Vert ^ { 2 }", "type": "interline_equation", "image_path": "5d0e2d8a8405e33a7e2c9e266ad35a26643675bde618e04f6a2f67efdf68e03d.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 191, 589, 420, 606 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 108, 615, 218, 626 ], "lines": [ { "bbox": [ 106, 613, 219, 628 ], "spans": [ { "bbox": [ 106, 613, 219, 628 ], "score": 1.0, "content": "We train the joint objective:", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "interline_equation", "bbox": [ 214, 631, 396, 650 ], "lines": [ { "bbox": [ 214, 631, 396, 650 ], "spans": [ { "bbox": [ 214, 631, 396, 650 ], "score": 0.92, "content": "\\theta = a r g \\operatorname* { m i n } _ { \\theta ^ { \\prime } } \\alpha \\mathcal { L } _ { \\theta ^ { \\prime } } ^ { v i d } ( v ) + ( 1 - \\alpha ) \\mathcal { L } _ { \\theta ^ { \\prime } } ^ { f r a m e } ( u )", "type": "interline_equation", "image_path": "a0ad186089757028efe5bd5852b31b09de661dbdcaf60df444e5bfebae73b46e.jpg" } ] } ], "index": 32, "virtual_lines": [ { "bbox": [ 214, 631, 396, 650 ], "spans": [], "index": 32 } ] }, { "type": "text", "bbox": [ 106, 654, 505, 732 ], "lines": [ { "bbox": [ 106, 655, 504, 666 ], "spans": [ { "bbox": [ 106, 655, 135, 666 ], "score": 1.0, "content": "Where", "type": "text" }, { "bbox": [ 135, 657, 143, 665 ], "score": 0.78, "content": "\\alpha", "type": "inline_equation" }, { "bbox": [ 144, 655, 504, 666 ], "score": 1.0, "content": "is a constant factor, see Fig. 4. Training on a single video or a handful of frames can easily", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 666, 506, 678 ], "spans": [ { "bbox": [ 106, 666, 506, 678 ], "score": 1.0, "content": "lead to overfitting, reducing the editing ability of the original model. To mitigate overfitting, we use", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "a small number of finetuning iterations and a low learning rate (see Appendix C). Note that while", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 686, 505, 701 ], "spans": [ { "bbox": [ 104, 686, 505, 701 ], "score": 1.0, "content": "such a training objective was used by Imagen-VideoHo et al. (2022a) and VDMHo et al. (2022c),", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 699, 506, 711 ], "spans": [ { "bbox": [ 106, 699, 506, 711 ], "score": 1.0, "content": "its purpose was different. There, the aim was to increase dataset size and diversity by training on", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "large image datasets. Here, the aim is to enforce the style of the video in the model, while allowing", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 721, 169, 734 ], "spans": [ { "bbox": [ 105, 721, 169, 734 ], "score": 1.0, "content": "motion editing.", "type": "text" } ], "index": 39 } ], "index": 36 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 763 ], "spans": [ { "bbox": [ 302, 750, 309, 763 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 151, 83, 458, 209 ], "blocks": [ { "type": "image_body", "bbox": [ 151, 83, 458, 209 ], "group_id": 0, "lines": [ { "bbox": [ 151, 83, 458, 209 ], "spans": [ { "bbox": [ 151, 83, 458, 209 ], "score": 0.973, "type": "image", "image_path": "12e0623b2a593ab2c48d73176d55b0335fa925ea264e914d3a903c218e455fb9.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 151, 83, 458, 125.0 ], "spans": [], "index": 0 }, { "bbox": [ 151, 125.0, 458, 167.0 ], "spans": [], "index": 1 }, { "bbox": [ 151, 167.0, 458, 209.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 217, 506, 273 ], "group_id": 0, "lines": [ { "bbox": [ 105, 217, 506, 231 ], "spans": [ { "bbox": [ 105, 217, 506, 231 ], "score": 1.0, "content": "Figure 4: Mixed Video-Image Finetuning: Finetuning the VDM on the input video alone limits", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 229, 505, 241 ], "spans": [ { "bbox": [ 106, 229, 505, 241 ], "score": 1.0, "content": "the extent of motion change. Instead, we use a mixed objective that beside the original objective", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 239, 506, 252 ], "spans": [ { "bbox": [ 105, 239, 506, 252 ], "score": 1.0, "content": "(bottom left) also finetunes on the unordered set of frames. We use “masked temporal attention“ to", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 250, 506, 264 ], "spans": [ { "bbox": [ 105, 250, 506, 264 ], "score": 1.0, "content": "prevent the temporal attention and convolution from changing (bottom right). This allows adding", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 263, 202, 273 ], "spans": [ { "bbox": [ 106, 263, 202, 273 ], "score": 1.0, "content": "motion to a static video", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 293, 505, 339 ], "lines": [], "index": 9.5, "bbox_fs": [ 105, 294, 505, 339 ], "lines_deleted": true }, { "type": "title", "bbox": [ 108, 351, 283, 363 ], "lines": [ { "bbox": [ 106, 351, 284, 364 ], "spans": [ { "bbox": [ 106, 351, 284, 364 ], "score": 1.0, "content": "4.2 MIXED VIDEO-IMAGE FINETUNING", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 372, 505, 483 ], "lines": [ { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "The naive method presented in Sec. 4.1 relies on a corrupted version of the input video which does", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 383, 506, 396 ], "spans": [ { "bbox": [ 105, 383, 506, 396 ], "score": 1.0, "content": "not include enough information to preserve high-resolution details such as fine textures or object", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 506, 407 ], "score": 1.0, "content": "identity. We tackle this by adding a preliminary stage of finetuning the model on the input video", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 403, 506, 419 ], "spans": [ { "bbox": [ 106, 407, 113, 415 ], "score": 0.6, "content": "v", "type": "inline_equation" }, { "bbox": [ 113, 403, 506, 419 ], "score": 1.0, "content": ". Note that this only needs to be done once for the video, which can then be edited by many", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 416, 506, 428 ], "spans": [ { "bbox": [ 105, 416, 506, 428 ], "score": 1.0, "content": "prompts without further finetuning. We would like the model to separately update its prior both on", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 427, 504, 438 ], "spans": [ { "bbox": [ 106, 427, 504, 438 ], "score": 1.0, "content": "the appearance and the motion of the input video. Our approach therefore treats the input video,", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 437, 505, 451 ], "spans": [ { "bbox": [ 105, 437, 323, 451 ], "score": 1.0, "content": "both as a single video clip and as an unordered set of", "type": "text" }, { "bbox": [ 324, 438, 335, 448 ], "score": 0.82, "content": "M", "type": "inline_equation" }, { "bbox": [ 336, 437, 416, 451 ], "score": 1.0, "content": "frames, denoted by", "type": "text" }, { "bbox": [ 417, 438, 501, 450 ], "score": 0.94, "content": "\\boldsymbol { u } = \\{ x _ { 1 } , x _ { 2 } , . . , x _ { M } \\}", "type": "inline_equation" }, { "bbox": [ 501, 437, 505, 451 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 447, 506, 463 ], "spans": [ { "bbox": [ 105, 447, 189, 463 ], "score": 1.0, "content": "We use a rare string", "type": "text" }, { "bbox": [ 190, 450, 199, 459 ], "score": 0.83, "content": "t ^ { * }", "type": "inline_equation" }, { "bbox": [ 199, 447, 506, 463 ], "score": 1.0, "content": "as the text prompt, following Ruiz et al. (2022). We finetune the denoising", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 460, 505, 472 ], "spans": [ { "bbox": [ 105, 460, 505, 472 ], "score": 1.0, "content": "models by a combination of two objectives. The first objective updates the model prior on both", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 470, 498, 484 ], "spans": [ { "bbox": [ 105, 470, 380, 484 ], "score": 1.0, "content": "motion and appearance by requiring it to reconstruct the input video", "type": "text" }, { "bbox": [ 380, 473, 387, 481 ], "score": 0.77, "content": "v", "type": "inline_equation" }, { "bbox": [ 387, 470, 484, 484 ], "score": 1.0, "content": "given its noisy versions", "type": "text" }, { "bbox": [ 484, 473, 493, 482 ], "score": 0.79, "content": "z _ { s }", "type": "inline_equation" }, { "bbox": [ 494, 470, 498, 484 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 22 } ], "index": 17.5, "bbox_fs": [ 105, 371, 506, 484 ] }, { "type": "interline_equation", "bbox": [ 198, 496, 412, 512 ], "lines": [ { "bbox": [ 198, 496, 412, 512 ], "spans": [ { "bbox": [ 198, 496, 412, 512 ], "score": 0.9, "content": "\\mathcal { L } _ { \\theta } ^ { v i d } ( v ) = \\mathbb { E } _ { \\epsilon \\sim N ( 0 , \\mathbf { I } ) , s \\in \\mathcal { U } ( 0 , 1 ) } \\Vert D _ { \\theta ^ { \\prime } } ( z _ { s } , s , t ^ { * } , c ) - v \\Vert ^ { 2 }", "type": "interline_equation", "image_path": "f3b76add60efaeb879d08b45df45ad23a9cb0ac3e703e38d3102542af9ac7a09.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 198, 496, 412, 512 ], "spans": [], "index": 23 } ] }, { "type": "text", "bbox": [ 107, 518, 504, 585 ], "lines": [ { "bbox": [ 105, 518, 505, 532 ], "spans": [ { "bbox": [ 105, 518, 505, 532 ], "score": 1.0, "content": "Additionally, we train the model to reconstruct each of the frames individually given their noisy", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 530, 505, 543 ], "spans": [ { "bbox": [ 105, 530, 505, 543 ], "score": 1.0, "content": "version. This enhances the appearance prior of the model, separately from the motion. Technically,", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 541, 506, 554 ], "spans": [ { "bbox": [ 105, 541, 288, 554 ], "score": 1.0, "content": "the model is trained on a sequence of frames", "type": "text" }, { "bbox": [ 289, 543, 295, 551 ], "score": 0.75, "content": "u", "type": "inline_equation" }, { "bbox": [ 296, 541, 506, 554 ], "score": 1.0, "content": "by replacing the temporal attention layers by trivial", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "fixed masks ensuring the model only pays attention within each frame, and also by masking the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 562, 506, 577 ], "spans": [ { "bbox": [ 105, 562, 468, 577 ], "score": 1.0, "content": "residual temporal convolution blocks. We denote the attention masked denoising model as", "type": "text" }, { "bbox": [ 468, 563, 482, 575 ], "score": 0.88, "content": "D _ { \\theta } ^ { a }", "type": "inline_equation" }, { "bbox": [ 483, 562, 506, 577 ], "score": 1.0, "content": ". The", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 575, 227, 586 ], "spans": [ { "bbox": [ 105, 575, 227, 586 ], "score": 1.0, "content": "masked attention objective is:", "type": "text" } ], "index": 29 } ], "index": 26.5, "bbox_fs": [ 105, 518, 506, 586 ] }, { "type": "interline_equation", "bbox": [ 191, 589, 420, 606 ], "lines": [ { "bbox": [ 191, 589, 420, 606 ], "spans": [ { "bbox": [ 191, 589, 420, 606 ], "score": 0.9, "content": "\\mathcal { L } _ { \\theta } ^ { f r a m e } ( u ) = \\mathbb { E } _ { \\epsilon \\sim N ( 0 , \\mathbf { I } ) , s \\in \\mathcal { U } ( 0 , 1 ) } \\Vert D _ { \\theta ^ { \\prime } } ^ { a } ( z _ { s } , s , t ^ { * } , c ) - u \\Vert ^ { 2 }", "type": "interline_equation", "image_path": "5d0e2d8a8405e33a7e2c9e266ad35a26643675bde618e04f6a2f67efdf68e03d.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 191, 589, 420, 606 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 108, 615, 218, 626 ], "lines": [ { "bbox": [ 106, 613, 219, 628 ], "spans": [ { "bbox": [ 106, 613, 219, 628 ], "score": 1.0, "content": "We train the joint objective:", "type": "text" } ], "index": 31 } ], "index": 31, "bbox_fs": [ 106, 613, 219, 628 ] }, { "type": "interline_equation", "bbox": [ 214, 631, 396, 650 ], "lines": [ { "bbox": [ 214, 631, 396, 650 ], "spans": [ { "bbox": [ 214, 631, 396, 650 ], "score": 0.92, "content": "\\theta = a r g \\operatorname* { m i n } _ { \\theta ^ { \\prime } } \\alpha \\mathcal { L } _ { \\theta ^ { \\prime } } ^ { v i d } ( v ) + ( 1 - \\alpha ) \\mathcal { L } _ { \\theta ^ { \\prime } } ^ { f r a m e } ( u )", "type": "interline_equation", "image_path": "a0ad186089757028efe5bd5852b31b09de661dbdcaf60df444e5bfebae73b46e.jpg" } ] } ], "index": 32, "virtual_lines": [ { "bbox": [ 214, 631, 396, 650 ], "spans": [], "index": 32 } ] }, { "type": "text", "bbox": [ 106, 654, 505, 732 ], "lines": [ { "bbox": [ 106, 655, 504, 666 ], "spans": [ { "bbox": [ 106, 655, 135, 666 ], "score": 1.0, "content": "Where", "type": "text" }, { "bbox": [ 135, 657, 143, 665 ], "score": 0.78, "content": "\\alpha", "type": "inline_equation" }, { "bbox": [ 144, 655, 504, 666 ], "score": 1.0, "content": "is a constant factor, see Fig. 4. Training on a single video or a handful of frames can easily", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 666, 506, 678 ], "spans": [ { "bbox": [ 106, 666, 506, 678 ], "score": 1.0, "content": "lead to overfitting, reducing the editing ability of the original model. To mitigate overfitting, we use", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "a small number of finetuning iterations and a low learning rate (see Appendix C). Note that while", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 686, 505, 701 ], "spans": [ { "bbox": [ 104, 686, 505, 701 ], "score": 1.0, "content": "such a training objective was used by Imagen-VideoHo et al. (2022a) and VDMHo et al. (2022c),", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 699, 506, 711 ], "spans": [ { "bbox": [ 106, 699, 506, 711 ], "score": 1.0, "content": "its purpose was different. There, the aim was to increase dataset size and diversity by training on", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "large image datasets. Here, the aim is to enforce the style of the video in the model, while allowing", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 721, 169, 734 ], "spans": [ { "bbox": [ 105, 721, 169, 734 ], "score": 1.0, "content": "motion editing.", "type": "text" } ], "index": 39 } ], "index": 36, "bbox_fs": [ 104, 655, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 79, 503, 219 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 79, 503, 219 ], "group_id": 0, "lines": [ { "bbox": [ 107, 79, 503, 219 ], "spans": [ { "bbox": [ 107, 79, 503, 219 ], "score": 0.97, "type": "image", "image_path": "0f571953d5039d4088edb017708dbd10e6a2e17c5f758968ef5a7ae46ba2e726.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 79, 503, 125.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 107, 125.66666666666666, 503, 172.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 107, 172.33333333333331, 503, 218.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 224, 505, 302 ], "group_id": 0, "lines": [ { "bbox": [ 105, 224, 505, 237 ], "spans": [ { "bbox": [ 105, 224, 505, 237 ], "score": 1.0, "content": "Figure 5: Inference Overview: Our method supports multiple applications by converting the input", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 236, 505, 248 ], "spans": [ { "bbox": [ 105, 236, 505, 248 ], "score": 1.0, "content": "into a uniform video format (left). For image-to-video, the input image is duplicated and trans-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "formed using perspective transformations, synthesizing a coarse video with some camera motion.", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "For subject-driven video generation, the input is omitted - finetuning alone takes care of the fidelity.", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 267, 505, 281 ], "spans": [ { "bbox": [ 105, 267, 505, 281 ], "score": 1.0, "content": "This coarse video is then edited using our general “Dreamix Video Editor“ (right): we first corrupt", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 279, 506, 293 ], "spans": [ { "bbox": [ 105, 279, 506, 293 ], "score": 1.0, "content": "the video by downsampling followed by adding noise. We then apply the finetuned text-guided", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 290, 390, 303 ], "spans": [ { "bbox": [ 106, 290, 390, 303 ], "score": 1.0, "content": "VDM, which upscales the video to the final spatio-temporal resolution", "type": "text" } ], "index": 9 } ], "index": 6 } ], "index": 3.5 }, { "type": "title", "bbox": [ 108, 326, 271, 338 ], "lines": [ { "bbox": [ 105, 326, 273, 340 ], "spans": [ { "bbox": [ 105, 326, 273, 340 ], "score": 1.0, "content": "5 APPLICATIONS OF DREAMIX", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 108, 353, 505, 386 ], "lines": [ { "bbox": [ 106, 353, 505, 365 ], "spans": [ { "bbox": [ 106, 353, 505, 365 ], "score": 1.0, "content": "The method proposed in Sec. 4, can edit motion and appearance in real-world videos. In this section,", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 365, 505, 376 ], "spans": [ { "bbox": [ 106, 365, 505, 376 ], "score": 1.0, "content": "we propose a framework for using our Dreamix video editor for general, text-conditioned image-to-", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 374, 271, 388 ], "spans": [ { "bbox": [ 106, 374, 271, 388 ], "score": 1.0, "content": "video editing, see Fig. 5 for an overview.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 392, 505, 546 ], "lines": [ { "bbox": [ 105, 391, 506, 405 ], "spans": [ { "bbox": [ 105, 391, 506, 405 ], "score": 1.0, "content": "Dreamix for Single Images. Provided our general video editing method, Dreamix, we now propose", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 402, 505, 417 ], "spans": [ { "bbox": [ 105, 402, 505, 417 ], "score": 1.0, "content": "a framework for image animation conditioned on a text prompt. The idea is to transform the image", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "or a set of images into a coarse, corrupted video and edit it using Dreamix. For example, given a", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 425, 505, 438 ], "spans": [ { "bbox": [ 106, 425, 161, 438 ], "score": 1.0, "content": "single image", "type": "text" }, { "bbox": [ 161, 427, 168, 435 ], "score": 0.68, "content": "x", "type": "inline_equation" }, { "bbox": [ 168, 425, 505, 438 ], "score": 1.0, "content": "as input, we can transform it to a video by replicating it 16 times to form a static", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 435, 506, 450 ], "spans": [ { "bbox": [ 105, 435, 131, 450 ], "score": 1.0, "content": "video", "type": "text" }, { "bbox": [ 132, 436, 198, 448 ], "score": 0.92, "content": "v = \\mathbf { \\bar { [ } } x , x , x . . . \\bar { x } ]", "type": "inline_equation" }, { "bbox": [ 198, 435, 506, 450 ], "score": 1.0, "content": ". We can then edit its appearance and motion using Dreamix conditioned on", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "a text prompt. Here, we do not wish to incorporate the motion of the input video (as it is static", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 457, 505, 471 ], "spans": [ { "bbox": [ 105, 457, 450, 471 ], "score": 1.0, "content": "and meaningless) and therefore use only the masked temporal attention finetuning", "type": "text" }, { "bbox": [ 450, 459, 482, 469 ], "score": 0.85, "content": "( \\alpha = 0", "type": "inline_equation" }, { "bbox": [ 482, 457, 505, 471 ], "score": 1.0, "content": "). To", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 469, 505, 482 ], "spans": [ { "bbox": [ 106, 469, 505, 482 ], "score": 1.0, "content": "create “cinematic” effects, we can further control the output video by simulating camera motion,", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 480, 506, 493 ], "spans": [ { "bbox": [ 106, 480, 506, 493 ], "score": 1.0, "content": "such as panning and zoom. We perform this by sampling a smooth sequence of 16 perspective", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 490, 505, 504 ], "spans": [ { "bbox": [ 105, 490, 172, 504 ], "score": 1.0, "content": "transformations", "type": "text" }, { "bbox": [ 172, 491, 218, 502 ], "score": 0.92, "content": "T _ { 1 } , T _ { 2 } . . T _ { 1 6 }", "type": "inline_equation" }, { "bbox": [ 219, 490, 505, 504 ], "score": 1.0, "content": "and apply each on the original image. When the perspective requires", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 502, 506, 515 ], "spans": [ { "bbox": [ 106, 502, 506, 515 ], "score": 1.0, "content": "pixels outside the input image, we simply outpaint them using reflection padding. We concatenate", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 512, 504, 525 ], "spans": [ { "bbox": [ 106, 512, 386, 525 ], "score": 1.0, "content": "the sequence of transformed images into a low quality input video", "type": "text" }, { "bbox": [ 386, 513, 501, 525 ], "score": 0.93, "content": "v ~ \\stackrel { - } { = } ~ [ T _ { 1 } ( \\stackrel { - } { x } ) , T _ { 2 } ( x ) . . T _ { 1 6 } ( x ) ]", "type": "inline_equation" }, { "bbox": [ 501, 512, 504, 525 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 523, 506, 537 ], "spans": [ { "bbox": [ 105, 523, 506, 537 ], "score": 1.0, "content": "While this does not result in realistic video, Dreamix can transform it into a high-quality edited", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 535, 372, 547 ], "spans": [ { "bbox": [ 106, 535, 372, 547 ], "score": 1.0, "content": "video. See Appendix D for details on the applied transformations.", "type": "text" } ], "index": 27 } ], "index": 20.5 }, { "type": "text", "bbox": [ 107, 551, 505, 651 ], "lines": [ { "bbox": [ 105, 551, 505, 564 ], "spans": [ { "bbox": [ 105, 551, 505, 564 ], "score": 1.0, "content": "Dreamix for subject-driven video generation. We propose to use Dreamix for text-conditioned", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 563, 505, 575 ], "spans": [ { "bbox": [ 106, 563, 505, 575 ], "score": 1.0, "content": "video generation given an image collection. Differently from existing methods, e.g., Dreambooth", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 573, 505, 587 ], "spans": [ { "bbox": [ 105, 573, 505, 587 ], "score": 1.0, "content": "Ruiz et al. (2022), it can add motion and not only change appearance. The input to our method", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 584, 505, 597 ], "spans": [ { "bbox": [ 105, 584, 505, 597 ], "score": 1.0, "content": "is a set of images, each containing the subject of interest. This can also use different frames from", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 595, 506, 609 ], "spans": [ { "bbox": [ 105, 595, 506, 609 ], "score": 1.0, "content": "the same video, as long as they show the same subject. Higher diversity of viewing angles and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 606, 505, 618 ], "spans": [ { "bbox": [ 105, 606, 505, 618 ], "score": 1.0, "content": "backgrounds is beneficial for the performance of the method. We then use the finetuning method", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 617, 506, 631 ], "spans": [ { "bbox": [ 105, 617, 380, 631 ], "score": 1.0, "content": "from Sec. 4.2, where we only use the masked attention finetuning", "type": "text" }, { "bbox": [ 381, 618, 411, 628 ], "score": 0.84, "content": "( \\alpha = 0", "type": "inline_equation" }, { "bbox": [ 411, 617, 506, 631 ], "score": 1.0, "content": "). After finetuning, we", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 628, 505, 641 ], "spans": [ { "bbox": [ 105, 628, 505, 641 ], "score": 1.0, "content": "use the text-to-video model without a conditioning video, but rather only using a text prompt (which", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 640, 227, 651 ], "spans": [ { "bbox": [ 106, 640, 211, 651 ], "score": 1.0, "content": "includes the special token", "type": "text" }, { "bbox": [ 211, 640, 221, 650 ], "score": 0.79, "content": "t ^ { * }", "type": "inline_equation" }, { "bbox": [ 221, 640, 227, 651 ], "score": 1.0, "content": ").", "type": "text" } ], "index": 36 } ], "index": 32 }, { "type": "title", "bbox": [ 107, 671, 200, 684 ], "lines": [ { "bbox": [ 105, 669, 201, 686 ], "spans": [ { "bbox": [ 105, 669, 201, 686 ], "score": 1.0, "content": "6 EXPERIMENTS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 104, 697, 506, 712 ], "spans": [ { "bbox": [ 104, 697, 506, 712 ], "score": 1.0, "content": "In this section, we establish that Dreamix is able to edit motion in real-world videos and images,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 709, 506, 721 ], "spans": [ { "bbox": [ 105, 709, 506, 721 ], "score": 1.0, "content": "a major improvement over the existing methods. To fully experience our results, please see the", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 721, 199, 732 ], "spans": [ { "bbox": [ 105, 721, 199, 732 ], "score": 1.0, "content": "supplementary videos.", "type": "text" } ], "index": 40 } ], "index": 39 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 107, 79, 503, 219 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 79, 503, 219 ], "group_id": 0, "lines": [ { "bbox": [ 107, 79, 503, 219 ], "spans": [ { "bbox": [ 107, 79, 503, 219 ], "score": 0.97, "type": "image", "image_path": "0f571953d5039d4088edb017708dbd10e6a2e17c5f758968ef5a7ae46ba2e726.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 79, 503, 125.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 107, 125.66666666666666, 503, 172.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 107, 172.33333333333331, 503, 218.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 224, 505, 302 ], "group_id": 0, "lines": [ { "bbox": [ 105, 224, 505, 237 ], "spans": [ { "bbox": [ 105, 224, 505, 237 ], "score": 1.0, "content": "Figure 5: Inference Overview: Our method supports multiple applications by converting the input", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 236, 505, 248 ], "spans": [ { "bbox": [ 105, 236, 505, 248 ], "score": 1.0, "content": "into a uniform video format (left). For image-to-video, the input image is duplicated and trans-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "formed using perspective transformations, synthesizing a coarse video with some camera motion.", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "For subject-driven video generation, the input is omitted - finetuning alone takes care of the fidelity.", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 267, 505, 281 ], "spans": [ { "bbox": [ 105, 267, 505, 281 ], "score": 1.0, "content": "This coarse video is then edited using our general “Dreamix Video Editor“ (right): we first corrupt", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 279, 506, 293 ], "spans": [ { "bbox": [ 105, 279, 506, 293 ], "score": 1.0, "content": "the video by downsampling followed by adding noise. We then apply the finetuned text-guided", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 290, 390, 303 ], "spans": [ { "bbox": [ 106, 290, 390, 303 ], "score": 1.0, "content": "VDM, which upscales the video to the final spatio-temporal resolution", "type": "text" } ], "index": 9 } ], "index": 6 } ], "index": 3.5 }, { "type": "title", "bbox": [ 108, 326, 271, 338 ], "lines": [ { "bbox": [ 105, 326, 273, 340 ], "spans": [ { "bbox": [ 105, 326, 273, 340 ], "score": 1.0, "content": "5 APPLICATIONS OF DREAMIX", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 108, 353, 505, 386 ], "lines": [ { "bbox": [ 106, 353, 505, 365 ], "spans": [ { "bbox": [ 106, 353, 505, 365 ], "score": 1.0, "content": "The method proposed in Sec. 4, can edit motion and appearance in real-world videos. In this section,", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 365, 505, 376 ], "spans": [ { "bbox": [ 106, 365, 505, 376 ], "score": 1.0, "content": "we propose a framework for using our Dreamix video editor for general, text-conditioned image-to-", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 374, 271, 388 ], "spans": [ { "bbox": [ 106, 374, 271, 388 ], "score": 1.0, "content": "video editing, see Fig. 5 for an overview.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 106, 353, 505, 388 ] }, { "type": "text", "bbox": [ 107, 392, 505, 546 ], "lines": [ { "bbox": [ 105, 391, 506, 405 ], "spans": [ { "bbox": [ 105, 391, 506, 405 ], "score": 1.0, "content": "Dreamix for Single Images. Provided our general video editing method, Dreamix, we now propose", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 402, 505, 417 ], "spans": [ { "bbox": [ 105, 402, 505, 417 ], "score": 1.0, "content": "a framework for image animation conditioned on a text prompt. The idea is to transform the image", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "or a set of images into a coarse, corrupted video and edit it using Dreamix. For example, given a", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 425, 505, 438 ], "spans": [ { "bbox": [ 106, 425, 161, 438 ], "score": 1.0, "content": "single image", "type": "text" }, { "bbox": [ 161, 427, 168, 435 ], "score": 0.68, "content": "x", "type": "inline_equation" }, { "bbox": [ 168, 425, 505, 438 ], "score": 1.0, "content": "as input, we can transform it to a video by replicating it 16 times to form a static", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 435, 506, 450 ], "spans": [ { "bbox": [ 105, 435, 131, 450 ], "score": 1.0, "content": "video", "type": "text" }, { "bbox": [ 132, 436, 198, 448 ], "score": 0.92, "content": "v = \\mathbf { \\bar { [ } } x , x , x . . . \\bar { x } ]", "type": "inline_equation" }, { "bbox": [ 198, 435, 506, 450 ], "score": 1.0, "content": ". We can then edit its appearance and motion using Dreamix conditioned on", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "a text prompt. Here, we do not wish to incorporate the motion of the input video (as it is static", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 457, 505, 471 ], "spans": [ { "bbox": [ 105, 457, 450, 471 ], "score": 1.0, "content": "and meaningless) and therefore use only the masked temporal attention finetuning", "type": "text" }, { "bbox": [ 450, 459, 482, 469 ], "score": 0.85, "content": "( \\alpha = 0", "type": "inline_equation" }, { "bbox": [ 482, 457, 505, 471 ], "score": 1.0, "content": "). To", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 469, 505, 482 ], "spans": [ { "bbox": [ 106, 469, 505, 482 ], "score": 1.0, "content": "create “cinematic” effects, we can further control the output video by simulating camera motion,", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 480, 506, 493 ], "spans": [ { "bbox": [ 106, 480, 506, 493 ], "score": 1.0, "content": "such as panning and zoom. We perform this by sampling a smooth sequence of 16 perspective", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 490, 505, 504 ], "spans": [ { "bbox": [ 105, 490, 172, 504 ], "score": 1.0, "content": "transformations", "type": "text" }, { "bbox": [ 172, 491, 218, 502 ], "score": 0.92, "content": "T _ { 1 } , T _ { 2 } . . T _ { 1 6 }", "type": "inline_equation" }, { "bbox": [ 219, 490, 505, 504 ], "score": 1.0, "content": "and apply each on the original image. When the perspective requires", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 502, 506, 515 ], "spans": [ { "bbox": [ 106, 502, 506, 515 ], "score": 1.0, "content": "pixels outside the input image, we simply outpaint them using reflection padding. We concatenate", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 512, 504, 525 ], "spans": [ { "bbox": [ 106, 512, 386, 525 ], "score": 1.0, "content": "the sequence of transformed images into a low quality input video", "type": "text" }, { "bbox": [ 386, 513, 501, 525 ], "score": 0.93, "content": "v ~ \\stackrel { - } { = } ~ [ T _ { 1 } ( \\stackrel { - } { x } ) , T _ { 2 } ( x ) . . T _ { 1 6 } ( x ) ]", "type": "inline_equation" }, { "bbox": [ 501, 512, 504, 525 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 523, 506, 537 ], "spans": [ { "bbox": [ 105, 523, 506, 537 ], "score": 1.0, "content": "While this does not result in realistic video, Dreamix can transform it into a high-quality edited", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 535, 372, 547 ], "spans": [ { "bbox": [ 106, 535, 372, 547 ], "score": 1.0, "content": "video. See Appendix D for details on the applied transformations.", "type": "text" } ], "index": 27 } ], "index": 20.5, "bbox_fs": [ 105, 391, 506, 547 ] }, { "type": "text", "bbox": [ 107, 551, 505, 651 ], "lines": [ { "bbox": [ 105, 551, 505, 564 ], "spans": [ { "bbox": [ 105, 551, 505, 564 ], "score": 1.0, "content": "Dreamix for subject-driven video generation. We propose to use Dreamix for text-conditioned", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 563, 505, 575 ], "spans": [ { "bbox": [ 106, 563, 505, 575 ], "score": 1.0, "content": "video generation given an image collection. Differently from existing methods, e.g., Dreambooth", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 573, 505, 587 ], "spans": [ { "bbox": [ 105, 573, 505, 587 ], "score": 1.0, "content": "Ruiz et al. (2022), it can add motion and not only change appearance. The input to our method", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 584, 505, 597 ], "spans": [ { "bbox": [ 105, 584, 505, 597 ], "score": 1.0, "content": "is a set of images, each containing the subject of interest. This can also use different frames from", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 595, 506, 609 ], "spans": [ { "bbox": [ 105, 595, 506, 609 ], "score": 1.0, "content": "the same video, as long as they show the same subject. Higher diversity of viewing angles and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 606, 505, 618 ], "spans": [ { "bbox": [ 105, 606, 505, 618 ], "score": 1.0, "content": "backgrounds is beneficial for the performance of the method. We then use the finetuning method", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 617, 506, 631 ], "spans": [ { "bbox": [ 105, 617, 380, 631 ], "score": 1.0, "content": "from Sec. 4.2, where we only use the masked attention finetuning", "type": "text" }, { "bbox": [ 381, 618, 411, 628 ], "score": 0.84, "content": "( \\alpha = 0", "type": "inline_equation" }, { "bbox": [ 411, 617, 506, 631 ], "score": 1.0, "content": "). After finetuning, we", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 628, 505, 641 ], "spans": [ { "bbox": [ 105, 628, 505, 641 ], "score": 1.0, "content": "use the text-to-video model without a conditioning video, but rather only using a text prompt (which", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 640, 227, 651 ], "spans": [ { "bbox": [ 106, 640, 211, 651 ], "score": 1.0, "content": "includes the special token", "type": "text" }, { "bbox": [ 211, 640, 221, 650 ], "score": 0.79, "content": "t ^ { * }", "type": "inline_equation" }, { "bbox": [ 221, 640, 227, 651 ], "score": 1.0, "content": ").", "type": "text" } ], "index": 36 } ], "index": 32, "bbox_fs": [ 105, 551, 506, 651 ] }, { "type": "title", "bbox": [ 107, 671, 200, 684 ], "lines": [ { "bbox": [ 105, 669, 201, 686 ], "spans": [ { "bbox": [ 105, 669, 201, 686 ], "score": 1.0, "content": "6 EXPERIMENTS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 104, 697, 506, 712 ], "spans": [ { "bbox": [ 104, 697, 506, 712 ], "score": 1.0, "content": "In this section, we establish that Dreamix is able to edit motion in real-world videos and images,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 709, 506, 721 ], "spans": [ { "bbox": [ 105, 709, 506, 721 ], "score": 1.0, "content": "a major improvement over the existing methods. To fully experience our results, please see the", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 721, 199, 732 ], "spans": [ { "bbox": [ 105, 721, 199, 732 ], "score": 1.0, "content": "supplementary videos.", "type": "text" } ], "index": 40 } ], "index": 39, "bbox_fs": [ 104, 697, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 182, 82, 502, 105 ], "lines": [ { "bbox": [ 182, 81, 504, 94 ], "spans": [ { "bbox": [ 182, 81, 504, 94 ], "score": 1.0, "content": "“The Merced river is overflowing, birds flying in the sky, camera is zooming out to reveal an", "type": "text" } ], "index": 0 }, { "bbox": [ 271, 93, 416, 107 ], "spans": [ { "bbox": [ 271, 93, 416, 107 ], "score": 1.0, "content": "American Buffalo bathing in the river”", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "image", "bbox": [ 109, 103, 501, 282 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 103, 501, 282 ], "group_id": 0, "lines": [ { "bbox": [ 109, 103, 501, 282 ], "spans": [ { "bbox": [ 109, 103, 501, 282 ], "score": 0.95, "type": "image", "image_path": "7147616b9362cf73d6ef0dea5dfc0c0df56272b6eff15bdc3c5c27e4eb26c23f.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 109, 103, 501, 162.66666666666666 ], "spans": [], "index": 2 }, { "bbox": [ 109, 162.66666666666666, 501, 222.33333333333331 ], "spans": [], "index": 3 }, { "bbox": [ 109, 222.33333333333331, 501, 282.0 ], "spans": [], "index": 4 } ] }, { "type": "image_caption", "bbox": [ 106, 295, 505, 340 ], "group_id": 0, "lines": [ { "bbox": [ 106, 296, 506, 308 ], "spans": [ { "bbox": [ 106, 296, 506, 308 ], "score": 1.0, "content": "“A bear walking”Figure 6:Input ImagesAdditional Image-to-Video Results: First row - the image is zoomed out to reveal a", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 307, 505, 319 ], "spans": [ { "bbox": [ 105, 307, 505, 319 ], "score": 1.0, "content": "bathing buffalo. Dreamix can also instill motion in a static image as in the second row where the", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "glass is gradually filled with coffee. Third row - animating a subject based on a small number of", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 328, 189, 343 ], "spans": [ { "bbox": [ 105, 328, 189, 343 ], "score": 1.0, "content": "independent images", "type": "text" } ], "index": 8 } ], "index": 6.5 } ], "index": 4.75 }, { "type": "table", "bbox": [ 196, 402, 415, 472 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 354, 504, 399 ], "group_id": 0, "lines": [ { "bbox": [ 105, 354, 506, 368 ], "spans": [ { "bbox": [ 105, 354, 506, 368 ], "score": 1.0, "content": "Table 1: User Study: Users rated editing results by quality, fidelity to the base video and alignment", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 366, 506, 379 ], "spans": [ { "bbox": [ 105, 366, 506, 379 ], "score": 1.0, "content": "with the text prompt. Based on visual inspection, we require an edit to score greater than 2.5 in", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "all dimensions to be successful and observe that Dreamix is the only method to achieve the desired", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 387, 145, 399 ], "spans": [ { "bbox": [ 105, 387, 145, 399 ], "score": 1.0, "content": "trade off", "type": "text" } ], "index": 12 } ], "index": 10.5 }, { "type": "table_body", "bbox": [ 196, 402, 415, 472 ], "group_id": 0, "lines": [ { "bbox": [ 196, 402, 415, 472 ], "spans": [ { "bbox": [ 196, 402, 415, 472 ], "score": 0.979, "html": "
MethodQualityFidelityAlignment Success
PnP2.16 ±1.13 3.78 ±0.993.39 ±1.3820%
TaVid1.99 ±0.92 3.29 ±1.212.69 ±1.5513%
Ours3.58±1.043.55 ±1.093.79 ±1.3376%
Uncond. 3.43 ±1.09 2.49±1.12 4.28 ±1.0245%
", "type": "table", "image_path": "3abc8575efd3aea279f68860a3f7f9e57f5460e0949166ad84fe140fdef6185f.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 196, 402, 415, 416.0 ], "spans": [], "index": 13 }, { "bbox": [ 196, 416.0, 415, 430.0 ], "spans": [], "index": 14 }, { "bbox": [ 196, 430.0, 415, 444.0 ], "spans": [], "index": 15 }, { "bbox": [ 196, 444.0, 415, 458.0 ], "spans": [], "index": 16 }, { "bbox": [ 196, 458.0, 415, 472.0 ], "spans": [], "index": 17 } ] } ], "index": 12.75 }, { "type": "title", "bbox": [ 107, 500, 232, 512 ], "lines": [ { "bbox": [ 106, 500, 232, 513 ], "spans": [ { "bbox": [ 106, 500, 232, 513 ], "score": 1.0, "content": "6.1 QUALITATIVE RESULTS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 522, 505, 600 ], "lines": [ { "bbox": [ 106, 522, 505, 536 ], "spans": [ { "bbox": [ 106, 522, 505, 536 ], "score": 1.0, "content": "Video Editing. In Fig. 1, we change the motion to dancing and the appearance from monkey to", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "bear while keeping the coarse attributes of the video fixed. Dreamix can also generate new motion", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 544, 505, 558 ], "spans": [ { "bbox": [ 105, 544, 505, 558 ], "score": 1.0, "content": "that does not necessarily align with the input video (puppy in Fig. 3, orangutan in supplementary", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 556, 505, 567 ], "spans": [ { "bbox": [ 106, 556, 505, 567 ], "score": 1.0, "content": "material (SM)), and can control camera movements (zoom-out example in the SM). Dreamix can", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 566, 505, 578 ], "spans": [ { "bbox": [ 105, 566, 505, 578 ], "score": 1.0, "content": "generate smooth visual modifications that align with the temporal information in the input video.", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 576, 505, 590 ], "spans": [ { "bbox": [ 106, 576, 505, 590 ], "score": 1.0, "content": "This includes adding effects (field and saxophone in the SM), adding or replacing objects (hat,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 587, 437, 600 ], "spans": [ { "bbox": [ 105, 587, 437, 600 ], "score": 1.0, "content": "skateboard, and robot in the SM), and changing the background (truck in the SM).", "type": "text" } ], "index": 25 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 604, 505, 660 ], "lines": [ { "bbox": [ 106, 605, 505, 617 ], "spans": [ { "bbox": [ 106, 605, 505, 617 ], "score": 1.0, "content": "Image-driven Videos. When the input is a single image, Dreamix can use its video prior to add new", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 615, 506, 629 ], "spans": [ { "bbox": [ 105, 615, 506, 629 ], "score": 1.0, "content": "moving objects (camel in SM), inject motion into the input (turtle in Fig. 2 and coffee in Fig. 6), or", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 626, 505, 639 ], "spans": [ { "bbox": [ 105, 626, 505, 639 ], "score": 1.0, "content": "new camera movements (buffalo in Fig. 6). Although Singer et al. (2022); Yu et al. (2022b) perform", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 638, 504, 650 ], "spans": [ { "bbox": [ 106, 638, 504, 650 ], "score": 1.0, "content": "image-driven animations, they can only add very simple motions (e.g. animating water or snowfall).", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 648, 502, 662 ], "spans": [ { "bbox": [ 105, 648, 502, 662 ], "score": 1.0, "content": "Our method is unique in adding large motions and moving objects into general, real-world images.", "type": "text" } ], "index": 30 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 666, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 506, 678 ], "spans": [ { "bbox": [ 106, 665, 506, 678 ], "score": 1.0, "content": "Subject-driven Video Generation. Dreamix can take an image collection showing the same subject", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 677, 505, 690 ], "spans": [ { "bbox": [ 106, 677, 505, 690 ], "score": 1.0, "content": "and generate new videos with this subject in motion. This is unique, as previous approaches could", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "only output still images. We demonstrate this on a range of subjects and actions including: the", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "weight-lifting toy fireman in Fig. 2, walking and drinking bear in Fig. 6 and SM. It can place the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 708, 506, 723 ], "spans": [ { "bbox": [ 105, 708, 506, 723 ], "score": 1.0, "content": "subjects in new surroundings, e.g., moving caterpillar on a leaf or even under a magnifying glass", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 720, 149, 733 ], "spans": [ { "bbox": [ 105, 720, 149, 733 ], "score": 1.0, "content": "(see SM).", "type": "text" } ], "index": 36 } ], "index": 33.5 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 182, 82, 502, 105 ], "lines": [ { "bbox": [ 182, 81, 504, 94 ], "spans": [ { "bbox": [ 182, 81, 504, 94 ], "score": 1.0, "content": "“The Merced river is overflowing, birds flying in the sky, camera is zooming out to reveal an", "type": "text" } ], "index": 0 }, { "bbox": [ 271, 93, 416, 107 ], "spans": [ { "bbox": [ 271, 93, 416, 107 ], "score": 1.0, "content": "American Buffalo bathing in the river”", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 182, 81, 504, 107 ] }, { "type": "image", "bbox": [ 109, 103, 501, 282 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 103, 501, 282 ], "group_id": 0, "lines": [ { "bbox": [ 109, 103, 501, 282 ], "spans": [ { "bbox": [ 109, 103, 501, 282 ], "score": 0.95, "type": "image", "image_path": "7147616b9362cf73d6ef0dea5dfc0c0df56272b6eff15bdc3c5c27e4eb26c23f.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 109, 103, 501, 162.66666666666666 ], "spans": [], "index": 2 }, { "bbox": [ 109, 162.66666666666666, 501, 222.33333333333331 ], "spans": [], "index": 3 }, { "bbox": [ 109, 222.33333333333331, 501, 282.0 ], "spans": [], "index": 4 } ] }, { "type": "image_caption", "bbox": [ 106, 295, 505, 340 ], "group_id": 0, "lines": [ { "bbox": [ 106, 296, 506, 308 ], "spans": [ { "bbox": [ 106, 296, 506, 308 ], "score": 1.0, "content": "“A bear walking”Figure 6:Input ImagesAdditional Image-to-Video Results: First row - the image is zoomed out to reveal a", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 307, 505, 319 ], "spans": [ { "bbox": [ 105, 307, 505, 319 ], "score": 1.0, "content": "bathing buffalo. Dreamix can also instill motion in a static image as in the second row where the", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "glass is gradually filled with coffee. Third row - animating a subject based on a small number of", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 328, 189, 343 ], "spans": [ { "bbox": [ 105, 328, 189, 343 ], "score": 1.0, "content": "independent images", "type": "text" } ], "index": 8 } ], "index": 6.5 } ], "index": 4.75 }, { "type": "table", "bbox": [ 196, 402, 415, 472 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 354, 504, 399 ], "group_id": 0, "lines": [ { "bbox": [ 105, 354, 506, 368 ], "spans": [ { "bbox": [ 105, 354, 506, 368 ], "score": 1.0, "content": "Table 1: User Study: Users rated editing results by quality, fidelity to the base video and alignment", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 366, 506, 379 ], "spans": [ { "bbox": [ 105, 366, 506, 379 ], "score": 1.0, "content": "with the text prompt. Based on visual inspection, we require an edit to score greater than 2.5 in", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "all dimensions to be successful and observe that Dreamix is the only method to achieve the desired", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 387, 145, 399 ], "spans": [ { "bbox": [ 105, 387, 145, 399 ], "score": 1.0, "content": "trade off", "type": "text" } ], "index": 12 } ], "index": 10.5 }, { "type": "table_body", "bbox": [ 196, 402, 415, 472 ], "group_id": 0, "lines": [ { "bbox": [ 196, 402, 415, 472 ], "spans": [ { "bbox": [ 196, 402, 415, 472 ], "score": 0.979, "html": "
MethodQualityFidelityAlignment Success
PnP2.16 ±1.13 3.78 ±0.993.39 ±1.3820%
TaVid1.99 ±0.92 3.29 ±1.212.69 ±1.5513%
Ours3.58±1.043.55 ±1.093.79 ±1.3376%
Uncond. 3.43 ±1.09 2.49±1.12 4.28 ±1.0245%
", "type": "table", "image_path": "3abc8575efd3aea279f68860a3f7f9e57f5460e0949166ad84fe140fdef6185f.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 196, 402, 415, 416.0 ], "spans": [], "index": 13 }, { "bbox": [ 196, 416.0, 415, 430.0 ], "spans": [], "index": 14 }, { "bbox": [ 196, 430.0, 415, 444.0 ], "spans": [], "index": 15 }, { "bbox": [ 196, 444.0, 415, 458.0 ], "spans": [], "index": 16 }, { "bbox": [ 196, 458.0, 415, 472.0 ], "spans": [], "index": 17 } ] } ], "index": 12.75 }, { "type": "title", "bbox": [ 107, 500, 232, 512 ], "lines": [ { "bbox": [ 106, 500, 232, 513 ], "spans": [ { "bbox": [ 106, 500, 232, 513 ], "score": 1.0, "content": "6.1 QUALITATIVE RESULTS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 522, 505, 600 ], "lines": [ { "bbox": [ 106, 522, 505, 536 ], "spans": [ { "bbox": [ 106, 522, 505, 536 ], "score": 1.0, "content": "Video Editing. In Fig. 1, we change the motion to dancing and the appearance from monkey to", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "bear while keeping the coarse attributes of the video fixed. Dreamix can also generate new motion", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 544, 505, 558 ], "spans": [ { "bbox": [ 105, 544, 505, 558 ], "score": 1.0, "content": "that does not necessarily align with the input video (puppy in Fig. 3, orangutan in supplementary", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 556, 505, 567 ], "spans": [ { "bbox": [ 106, 556, 505, 567 ], "score": 1.0, "content": "material (SM)), and can control camera movements (zoom-out example in the SM). Dreamix can", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 566, 505, 578 ], "spans": [ { "bbox": [ 105, 566, 505, 578 ], "score": 1.0, "content": "generate smooth visual modifications that align with the temporal information in the input video.", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 576, 505, 590 ], "spans": [ { "bbox": [ 106, 576, 505, 590 ], "score": 1.0, "content": "This includes adding effects (field and saxophone in the SM), adding or replacing objects (hat,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 587, 437, 600 ], "spans": [ { "bbox": [ 105, 587, 437, 600 ], "score": 1.0, "content": "skateboard, and robot in the SM), and changing the background (truck in the SM).", "type": "text" } ], "index": 25 } ], "index": 22, "bbox_fs": [ 105, 522, 505, 600 ] }, { "type": "text", "bbox": [ 107, 604, 505, 660 ], "lines": [ { "bbox": [ 106, 605, 505, 617 ], "spans": [ { "bbox": [ 106, 605, 505, 617 ], "score": 1.0, "content": "Image-driven Videos. When the input is a single image, Dreamix can use its video prior to add new", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 615, 506, 629 ], "spans": [ { "bbox": [ 105, 615, 506, 629 ], "score": 1.0, "content": "moving objects (camel in SM), inject motion into the input (turtle in Fig. 2 and coffee in Fig. 6), or", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 626, 505, 639 ], "spans": [ { "bbox": [ 105, 626, 505, 639 ], "score": 1.0, "content": "new camera movements (buffalo in Fig. 6). Although Singer et al. (2022); Yu et al. (2022b) perform", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 638, 504, 650 ], "spans": [ { "bbox": [ 106, 638, 504, 650 ], "score": 1.0, "content": "image-driven animations, they can only add very simple motions (e.g. animating water or snowfall).", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 648, 502, 662 ], "spans": [ { "bbox": [ 105, 648, 502, 662 ], "score": 1.0, "content": "Our method is unique in adding large motions and moving objects into general, real-world images.", "type": "text" } ], "index": 30 } ], "index": 28, "bbox_fs": [ 105, 605, 506, 662 ] }, { "type": "text", "bbox": [ 106, 666, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 506, 678 ], "spans": [ { "bbox": [ 106, 665, 506, 678 ], "score": 1.0, "content": "Subject-driven Video Generation. Dreamix can take an image collection showing the same subject", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 677, 505, 690 ], "spans": [ { "bbox": [ 106, 677, 505, 690 ], "score": 1.0, "content": "and generate new videos with this subject in motion. This is unique, as previous approaches could", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "only output still images. We demonstrate this on a range of subjects and actions including: the", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "weight-lifting toy fireman in Fig. 2, walking and drinking bear in Fig. 6 and SM. It can place the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 708, 506, 723 ], "spans": [ { "bbox": [ 105, 708, 506, 723 ], "score": 1.0, "content": "subjects in new surroundings, e.g., moving caterpillar on a leaf or even under a magnifying glass", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 720, 149, 733 ], "spans": [ { "bbox": [ 105, 720, 149, 733 ], "score": 1.0, "content": "(see SM).", "type": "text" } ], "index": 36 } ], "index": 33.5, "bbox_fs": [ 105, 665, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 195, 138, 416, 192 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 504, 134 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 93 ], "spans": [ { "bbox": [ 105, 79, 468, 93 ], "score": 1.0, "content": "Table 2: Baseline Comparisons: Our method achieves better temporal consistency than", "type": "text" }, { "bbox": [ 469, 81, 487, 91 ], "score": 0.53, "content": "\\mathrm { P n P }", "type": "inline_equation" }, { "bbox": [ 487, 79, 505, 93 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 505, 103 ], "spans": [ { "bbox": [ 106, 91, 505, 103 ], "score": 1.0, "content": "Tune-a-Video (TaVid). Moreover, Dreamix is successful at motion editing while other methods", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "cannot. This is reflected in the better quality (low LPIPS) and alignment (high CLIP Score). While", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 113, 505, 126 ], "spans": [ { "bbox": [ 106, 113, 505, 126 ], "score": 1.0, "content": "the unconditional method seems to outperform Dreamix, it has poor fidelity as it is not conditioned", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 124, 181, 136 ], "spans": [ { "bbox": [ 106, 124, 181, 136 ], "score": 1.0, "content": "on the input video", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 195, 138, 416, 192 ], "group_id": 0, "lines": [ { "bbox": [ 195, 138, 416, 192 ], "spans": [ { "bbox": [ 195, 138, 416, 192 ], "score": 0.96, "html": "
MetricPnP TaVid OursUncond.
LPIPS↓0.209 0.145 0.1120.101
CLIP Sc0re ↑ 0.304 0.303 0.3170.320
FidelitySee user study (Tab.1) for evaluation
", "type": "table", "image_path": "e439a9112c226964bc3308de0e531ff5e729f3e1cc3dbee762da2cee44890075.jpg" } ] } ], "index": 6.5, "virtual_lines": [ { "bbox": [ 195, 138, 416, 151.5 ], "spans": [], "index": 5 }, { "bbox": [ 195, 151.5, 416, 165.0 ], "spans": [], "index": 6 }, { "bbox": [ 195, 165.0, 416, 178.5 ], "spans": [], "index": 7 }, { "bbox": [ 195, 178.5, 416, 192.0 ], "spans": [], "index": 8 } ] } ], "index": 4.25 }, { "type": "image", "bbox": [ 116, 212, 493, 445 ], "blocks": [ { "type": "image_body", "bbox": [ 116, 212, 493, 445 ], "group_id": 0, "lines": [ { "bbox": [ 116, 212, 493, 445 ], "spans": [ { "bbox": [ 116, 212, 493, 445 ], "score": 0.823, "type": "image", "image_path": "f638a89d9ba5e86a64f394124c4fd6b20344e3149b14b42bb476de9b259437d2.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 116, 212, 493, 289.6666666666667 ], "spans": [], "index": 9 }, { "bbox": [ 116, 289.6666666666667, 493, 367.33333333333337 ], "spans": [], "index": 10 }, { "bbox": [ 116, 367.33333333333337, 493, 445.00000000000006 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 451, 505, 517 ], "group_id": 0, "lines": [ { "bbox": [ 105, 450, 506, 463 ], "spans": [ { "bbox": [ 105, 450, 506, 463 ], "score": 1.0, "content": "Figure 7: Comparison to Baseline Methods for Motion Edits: Although the quality and alignment", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 462, 504, 474 ], "spans": [ { "bbox": [ 106, 462, 504, 474 ], "score": 1.0, "content": "of unconditional generation are high, there is no resemblance to the original video (low fidelity).", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 472, 505, 485 ], "spans": [ { "bbox": [ 106, 472, 132, 485 ], "score": 1.0, "content": "While", "type": "text" }, { "bbox": [ 133, 473, 151, 483 ], "score": 0.29, "content": "\\mathrm { P n P }", "type": "inline_equation" }, { "bbox": [ 151, 472, 505, 485 ], "score": 1.0, "content": "and Tune-a-Video preserve the scene, they fail to edit the motion according to the prompt", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 483, 505, 496 ], "spans": [ { "bbox": [ 105, 483, 505, 496 ], "score": 1.0, "content": "(no waving) and suffer from poor temporal consistency (flickering). Our method is able to edit the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 494, 505, 507 ], "spans": [ { "bbox": [ 105, 494, 505, 507 ], "score": 1.0, "content": "motion according to the prompt while preserving the fidelity and generating a high quality video.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 505, 504, 518 ], "spans": [ { "bbox": [ 105, 505, 504, 518 ], "score": 1.0, "content": "Moreover, video-based methods (Uncond. and ours) exhibit motion blur, also present in real videos", "type": "text" } ], "index": 17 } ], "index": 14.5 } ], "index": 12.25 }, { "type": "title", "bbox": [ 107, 543, 240, 554 ], "lines": [ { "bbox": [ 106, 543, 241, 556 ], "spans": [ { "bbox": [ 106, 543, 241, 556 ], "score": 1.0, "content": "6.2 BASELINE COMPARISONS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 566, 504, 610 ], "lines": [ { "bbox": [ 105, 565, 506, 579 ], "spans": [ { "bbox": [ 105, 565, 506, 579 ], "score": 1.0, "content": "Baselines. We compare our method against three baselines: Unconditional. Directly mapping the", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "text prompt to a video, without conditioning on the input video using a model similar to Imagen-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 199, 601 ], "score": 1.0, "content": "Video. Plug-and-Play", "type": "text" }, { "bbox": [ 199, 588, 223, 599 ], "score": 0.32, "content": "( P n P )", "type": "inline_equation" }, { "bbox": [ 224, 587, 505, 601 ], "score": 1.0, "content": ". Applying PnPTumanyan et al. (2022) on each video frame indepen-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 599, 472, 612 ], "spans": [ { "bbox": [ 105, 599, 472, 612 ], "score": 1.0, "content": "dently. Tune-a-Video (TaVid). Finetuning Tune-a-VideoWu et al. (2022) on the input video.", "type": "text" } ], "index": 22 } ], "index": 20.5 }, { "type": "text", "bbox": [ 105, 615, 504, 638 ], "lines": [ { "bbox": [ 106, 615, 505, 628 ], "spans": [ { "bbox": [ 106, 615, 505, 628 ], "score": 1.0, "content": "Data. We created a dataset of 29 videos taken from YouTube-8M Abu-El-Haija et al. (2016) and", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 626, 371, 639 ], "spans": [ { "bbox": [ 106, 626, 371, 639 ], "score": 1.0, "content": "127 text prompts, spanning different categories (see Appendix E).", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "text", "bbox": [ 107, 643, 505, 732 ], "lines": [ { "bbox": [ 106, 643, 505, 655 ], "spans": [ { "bbox": [ 106, 643, 505, 655 ], "score": 1.0, "content": "Quantitative Comparison. We measure alignment by the frame-level CLIP Score Hessel et al.", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "(2021) and quality (stability) with LPIPS Zhang et al. (2018) between consecutive frames. As auto-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 666, 505, 677 ], "spans": [ { "bbox": [ 106, 666, 505, 677 ], "score": 1.0, "content": "matic metrics do not measure fidelity and are imperfectly aligned with human judgement, we also", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 677, 504, 689 ], "spans": [ { "bbox": [ 105, 677, 479, 689 ], "score": 1.0, "content": "conduct a user study. A panel of 20 evaluators rated each video/prompt pair on a scale of", "type": "text" }, { "bbox": [ 479, 677, 504, 687 ], "score": 0.49, "content": "1 - 5", "type": "inline_equation" } ], "index": 28 }, { "bbox": [ 105, 687, 505, 699 ], "spans": [ { "bbox": [ 105, 687, 505, 699 ], "score": 1.0, "content": "to evaluate its quality, fidelity and alignment. When visually inspecting the results we discover", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 699, 504, 710 ], "spans": [ { "bbox": [ 106, 699, 504, 710 ], "score": 1.0, "content": "that videos that received a score lower than 2.5 in any of the dimensions are usually clear failure", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 505, 722 ], "score": 1.0, "content": "cases. Therefore we also report the percentage of items where all dimensions are larger than 2.5 (i.e.", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 720, 423, 732 ], "spans": [ { "bbox": [ 106, 720, 423, 732 ], "score": 1.0, "content": "“Success“). See Appendix F.2 for additional details on the evaluation protocol.", "type": "text" } ], "index": 32 } ], "index": 28.5 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 300, 750, 309, 761 ], "spans": [ { "bbox": [ 300, 750, 309, 761 ], "score": 1.0, "content": "8", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 195, 138, 416, 192 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 504, 134 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 93 ], "spans": [ { "bbox": [ 105, 79, 468, 93 ], "score": 1.0, "content": "Table 2: Baseline Comparisons: Our method achieves better temporal consistency than", "type": "text" }, { "bbox": [ 469, 81, 487, 91 ], "score": 0.53, "content": "\\mathrm { P n P }", "type": "inline_equation" }, { "bbox": [ 487, 79, 505, 93 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 505, 103 ], "spans": [ { "bbox": [ 106, 91, 505, 103 ], "score": 1.0, "content": "Tune-a-Video (TaVid). Moreover, Dreamix is successful at motion editing while other methods", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "cannot. This is reflected in the better quality (low LPIPS) and alignment (high CLIP Score). While", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 113, 505, 126 ], "spans": [ { "bbox": [ 106, 113, 505, 126 ], "score": 1.0, "content": "the unconditional method seems to outperform Dreamix, it has poor fidelity as it is not conditioned", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 124, 181, 136 ], "spans": [ { "bbox": [ 106, 124, 181, 136 ], "score": 1.0, "content": "on the input video", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 195, 138, 416, 192 ], "group_id": 0, "lines": [ { "bbox": [ 195, 138, 416, 192 ], "spans": [ { "bbox": [ 195, 138, 416, 192 ], "score": 0.96, "html": "
MetricPnP TaVid OursUncond.
LPIPS↓0.209 0.145 0.1120.101
CLIP Sc0re ↑ 0.304 0.303 0.3170.320
FidelitySee user study (Tab.1) for evaluation
", "type": "table", "image_path": "e439a9112c226964bc3308de0e531ff5e729f3e1cc3dbee762da2cee44890075.jpg" } ] } ], "index": 6.5, "virtual_lines": [ { "bbox": [ 195, 138, 416, 151.5 ], "spans": [], "index": 5 }, { "bbox": [ 195, 151.5, 416, 165.0 ], "spans": [], "index": 6 }, { "bbox": [ 195, 165.0, 416, 178.5 ], "spans": [], "index": 7 }, { "bbox": [ 195, 178.5, 416, 192.0 ], "spans": [], "index": 8 } ] } ], "index": 4.25 }, { "type": "image", "bbox": [ 116, 212, 493, 445 ], "blocks": [ { "type": "image_body", "bbox": [ 116, 212, 493, 445 ], "group_id": 0, "lines": [ { "bbox": [ 116, 212, 493, 445 ], "spans": [ { "bbox": [ 116, 212, 493, 445 ], "score": 0.823, "type": "image", "image_path": "f638a89d9ba5e86a64f394124c4fd6b20344e3149b14b42bb476de9b259437d2.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 116, 212, 493, 289.6666666666667 ], "spans": [], "index": 9 }, { "bbox": [ 116, 289.6666666666667, 493, 367.33333333333337 ], "spans": [], "index": 10 }, { "bbox": [ 116, 367.33333333333337, 493, 445.00000000000006 ], "spans": [], "index": 11 } ] }, { "type": "image_caption", "bbox": [ 106, 451, 505, 517 ], "group_id": 0, "lines": [ { "bbox": [ 105, 450, 506, 463 ], "spans": [ { "bbox": [ 105, 450, 506, 463 ], "score": 1.0, "content": "Figure 7: Comparison to Baseline Methods for Motion Edits: Although the quality and alignment", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 462, 504, 474 ], "spans": [ { "bbox": [ 106, 462, 504, 474 ], "score": 1.0, "content": "of unconditional generation are high, there is no resemblance to the original video (low fidelity).", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 472, 505, 485 ], "spans": [ { "bbox": [ 106, 472, 132, 485 ], "score": 1.0, "content": "While", "type": "text" }, { "bbox": [ 133, 473, 151, 483 ], "score": 0.29, "content": "\\mathrm { P n P }", "type": "inline_equation" }, { "bbox": [ 151, 472, 505, 485 ], "score": 1.0, "content": "and Tune-a-Video preserve the scene, they fail to edit the motion according to the prompt", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 483, 505, 496 ], "spans": [ { "bbox": [ 105, 483, 505, 496 ], "score": 1.0, "content": "(no waving) and suffer from poor temporal consistency (flickering). Our method is able to edit the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 494, 505, 507 ], "spans": [ { "bbox": [ 105, 494, 505, 507 ], "score": 1.0, "content": "motion according to the prompt while preserving the fidelity and generating a high quality video.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 505, 504, 518 ], "spans": [ { "bbox": [ 105, 505, 504, 518 ], "score": 1.0, "content": "Moreover, video-based methods (Uncond. and ours) exhibit motion blur, also present in real videos", "type": "text" } ], "index": 17 } ], "index": 14.5 } ], "index": 12.25 }, { "type": "title", "bbox": [ 107, 543, 240, 554 ], "lines": [ { "bbox": [ 106, 543, 241, 556 ], "spans": [ { "bbox": [ 106, 543, 241, 556 ], "score": 1.0, "content": "6.2 BASELINE COMPARISONS", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 566, 504, 610 ], "lines": [ { "bbox": [ 105, 565, 506, 579 ], "spans": [ { "bbox": [ 105, 565, 506, 579 ], "score": 1.0, "content": "Baselines. We compare our method against three baselines: Unconditional. Directly mapping the", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "text prompt to a video, without conditioning on the input video using a model similar to Imagen-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 199, 601 ], "score": 1.0, "content": "Video. Plug-and-Play", "type": "text" }, { "bbox": [ 199, 588, 223, 599 ], "score": 0.32, "content": "( P n P )", "type": "inline_equation" }, { "bbox": [ 224, 587, 505, 601 ], "score": 1.0, "content": ". Applying PnPTumanyan et al. (2022) on each video frame indepen-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 599, 472, 612 ], "spans": [ { "bbox": [ 105, 599, 472, 612 ], "score": 1.0, "content": "dently. Tune-a-Video (TaVid). Finetuning Tune-a-VideoWu et al. (2022) on the input video.", "type": "text" } ], "index": 22 } ], "index": 20.5, "bbox_fs": [ 105, 565, 506, 612 ] }, { "type": "text", "bbox": [ 105, 615, 504, 638 ], "lines": [ { "bbox": [ 106, 615, 505, 628 ], "spans": [ { "bbox": [ 106, 615, 505, 628 ], "score": 1.0, "content": "Data. We created a dataset of 29 videos taken from YouTube-8M Abu-El-Haija et al. (2016) and", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 626, 371, 639 ], "spans": [ { "bbox": [ 106, 626, 371, 639 ], "score": 1.0, "content": "127 text prompts, spanning different categories (see Appendix E).", "type": "text" } ], "index": 24 } ], "index": 23.5, "bbox_fs": [ 106, 615, 505, 639 ] }, { "type": "text", "bbox": [ 107, 643, 505, 732 ], "lines": [ { "bbox": [ 106, 643, 505, 655 ], "spans": [ { "bbox": [ 106, 643, 505, 655 ], "score": 1.0, "content": "Quantitative Comparison. We measure alignment by the frame-level CLIP Score Hessel et al.", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "(2021) and quality (stability) with LPIPS Zhang et al. (2018) between consecutive frames. As auto-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 666, 505, 677 ], "spans": [ { "bbox": [ 106, 666, 505, 677 ], "score": 1.0, "content": "matic metrics do not measure fidelity and are imperfectly aligned with human judgement, we also", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 677, 504, 689 ], "spans": [ { "bbox": [ 105, 677, 479, 689 ], "score": 1.0, "content": "conduct a user study. A panel of 20 evaluators rated each video/prompt pair on a scale of", "type": "text" }, { "bbox": [ 479, 677, 504, 687 ], "score": 0.49, "content": "1 - 5", "type": "inline_equation" } ], "index": 28 }, { "bbox": [ 105, 687, 505, 699 ], "spans": [ { "bbox": [ 105, 687, 505, 699 ], "score": 1.0, "content": "to evaluate its quality, fidelity and alignment. When visually inspecting the results we discover", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 699, 504, 710 ], "spans": [ { "bbox": [ 106, 699, 504, 710 ], "score": 1.0, "content": "that videos that received a score lower than 2.5 in any of the dimensions are usually clear failure", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 505, 722 ], "score": 1.0, "content": "cases. Therefore we also report the percentage of items where all dimensions are larger than 2.5 (i.e.", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 720, 423, 732 ], "spans": [ { "bbox": [ 106, 720, 423, 732 ], "score": 1.0, "content": "“Success“). See Appendix F.2 for additional details on the evaluation protocol.", "type": "text" } ], "index": 32 } ], "index": 28.5, "bbox_fs": [ 105, 643, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 194, 144, 417, 219 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 80, 505, 136 ], "group_id": 0, "lines": [ { "bbox": [ 106, 81, 505, 92 ], "spans": [ { "bbox": [ 106, 81, 505, 92 ], "score": 1.0, "content": "Table 3: Ablation Study: Left: Users were asked to compare text-guided video edits of with (w/", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 505, 104 ], "spans": [ { "bbox": [ 105, 90, 505, 104 ], "score": 1.0, "content": "Ft) and without (w/o Ft) finetuning. “None“ indicates failure of both methods according to user.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 101, 505, 115 ], "spans": [ { "bbox": [ 105, 101, 505, 115 ], "score": 1.0, "content": "Apart from style-based edits, where high fidelity is not needed, finetuning significantly improves", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 111, 506, 127 ], "spans": [ { "bbox": [ 105, 111, 506, 127 ], "score": 1.0, "content": "the results. Right: Users were asked to compare video finetuning (Vid) with mixed video-image", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 124, 443, 136 ], "spans": [ { "bbox": [ 106, 124, 443, 136 ], "score": 1.0, "content": "finetuning (Mix). Mixed finetuning significantly improves the results for most cases", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 194, 144, 417, 219 ], "group_id": 0, "lines": [ { "bbox": [ 194, 144, 417, 219 ], "spans": [ { "bbox": [ 194, 144, 417, 219 ], "score": 0.977, "html": "
Tyiedifs w/o Ft. w/Ft. NoneVid Mix
Motion3617% 72%11%35% 65%
Object4436% 48%16%62% 38%
Background3219% 77%9%36% 64%
Style1567% 27%6%26% 74%
", "type": "table", "image_path": "9728b5e248d5f000cf2ee7bff5d5c693a4f5e6ea6b15d5d6680da7ba7ae54296.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 194, 144, 417, 159.0 ], "spans": [], "index": 5 }, { "bbox": [ 194, 159.0, 417, 174.0 ], "spans": [], "index": 6 }, { "bbox": [ 194, 174.0, 417, 189.0 ], "spans": [], "index": 7 }, { "bbox": [ 194, 189.0, 417, 204.0 ], "spans": [], "index": 8 }, { "bbox": [ 194, 204.0, 417, 219.0 ], "spans": [], "index": 9 } ] } ], "index": 4.5 }, { "type": "text", "bbox": [ 106, 230, 505, 296 ], "lines": [ { "bbox": [ 106, 230, 505, 243 ], "spans": [ { "bbox": [ 106, 230, 505, 243 ], "score": 1.0, "content": "The evaluation and user study are presented in Tab. 2 and Tab. 1. Image-based methods (PnP,", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 240, 505, 254 ], "spans": [ { "bbox": [ 105, 240, 505, 254 ], "score": 1.0, "content": "Tune-a-Video) exhibit impaired temporal consistency, resulting in low quality. Moreover, they are", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 251, 505, 265 ], "spans": [ { "bbox": [ 105, 251, 505, 265 ], "score": 1.0, "content": "unable to perform motion edits, resulting in poor alignment and high fidelity. Video-based methods", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "maintain temporal consistency while allowing motion editing. Although unconditional generation", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 273, 506, 286 ], "spans": [ { "bbox": [ 106, 273, 506, 286 ], "score": 1.0, "content": "outperforms our method in the automatic evaluations (Tab. 2), it has poor fidelity (Tab. 1) as it is not", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 285, 430, 297 ], "spans": [ { "bbox": [ 106, 285, 430, 297 ], "score": 1.0, "content": "conditioned on the input video. Overall, our method has the highest success rate.", "type": "text" } ], "index": 15 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 301, 505, 368 ], "lines": [ { "bbox": [ 106, 302, 505, 315 ], "spans": [ { "bbox": [ 106, 302, 505, 315 ], "score": 1.0, "content": "Qualitative Comparison. Figure 7 presents an example of motion editing by Dreamix compared", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 313, 505, 325 ], "spans": [ { "bbox": [ 105, 313, 505, 325 ], "score": 1.0, "content": "to the baselines. The text-to-video model achieves low fidelity edits as it is not conditioned on", "type": "text" } ], "index": 17 }, { "bbox": [ 104, 322, 505, 338 ], "spans": [ { "bbox": [ 104, 322, 505, 338 ], "score": 1.0, "content": "the original video. PnP preserves the scene but fails to perform the edit and lacks consistency", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 505, 347 ], "score": 1.0, "content": "between different frames. Tune-a-Video exhibits better temporal consistency but still fails to perform", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "the motion edit. Dreamix performs well on all three objectives, adding the desired motion while", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 357, 252, 369 ], "spans": [ { "bbox": [ 105, 357, 252, 369 ], "score": 1.0, "content": "preserving fidelity and high-quality.", "type": "text" } ], "index": 21 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 378, 209, 389 ], "lines": [ { "bbox": [ 106, 377, 210, 391 ], "spans": [ { "bbox": [ 106, 377, 210, 391 ], "score": 1.0, "content": "6.3 ABLATION STUDY", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 396, 505, 495 ], "lines": [ { "bbox": [ 106, 395, 505, 409 ], "spans": [ { "bbox": [ 106, 395, 505, 409 ], "score": 1.0, "content": "We ablate the use of finetuning and the mixed video-image finetuning by performing a user study", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 406, 505, 420 ], "spans": [ { "bbox": [ 105, 406, 505, 420 ], "score": 1.0, "content": "using the dataset described above. The ablation indeed supports the idea of using finetuning in cases", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 417, 505, 431 ], "spans": [ { "bbox": [ 105, 417, 505, 431 ], "score": 1.0, "content": "where high-editability is required. We can see that Motion changes require high-editability and are", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 429, 505, 442 ], "spans": [ { "bbox": [ 105, 429, 505, 442 ], "score": 1.0, "content": "thus improved by finetuning. Moreover, as the noising corrupts the video, preserving fine-details", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 438, 505, 454 ], "spans": [ { "bbox": [ 105, 438, 505, 454 ], "score": 1.0, "content": "in background, color or texture edits requires finetuning. In contrast, denoising without finetuning", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 451, 504, 463 ], "spans": [ { "bbox": [ 106, 451, 504, 463 ], "score": 1.0, "content": "worked well for style edits, where finetuning was often detrimental. This is expected as style edits", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 462, 506, 475 ], "spans": [ { "bbox": [ 105, 462, 506, 475 ], "score": 1.0, "content": "are often conflicted with high fidelity preservation (e.g. changing the texture of an object means", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 472, 506, 486 ], "spans": [ { "bbox": [ 105, 472, 506, 486 ], "score": 1.0, "content": "reducing fidelity). The ablation shows that in most cases mixed finetuning improves the results by a", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 484, 460, 497 ], "spans": [ { "bbox": [ 105, 484, 460, 497 ], "score": 1.0, "content": "wide margin. Results are presented in Tab. 3, a visual ablation is found in Appendix F.2.", "type": "text" } ], "index": 31 } ], "index": 27 }, { "type": "title", "bbox": [ 108, 508, 194, 521 ], "lines": [ { "bbox": [ 105, 507, 195, 524 ], "spans": [ { "bbox": [ 105, 507, 195, 524 ], "score": 1.0, "content": "7 LIMITATIONS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 108, 530, 488, 541 ], "lines": [ { "bbox": [ 107, 529, 489, 542 ], "spans": [ { "bbox": [ 107, 529, 489, 542 ], "score": 1.0, "content": "While Dreamix is the first diffusion-based video method that can edit motion, it has limitations.", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 106, 547, 504, 581 ], "lines": [ { "bbox": [ 106, 547, 505, 560 ], "spans": [ { "bbox": [ 106, 547, 505, 560 ], "score": 1.0, "content": "Computational Cost. VDMs are computationally expensive. Finetuning our model using 4 TPU", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 558, 505, 571 ], "spans": [ { "bbox": [ 106, 558, 505, 571 ], "score": 1.0, "content": "v4 accelerators requires around 30 minutes per video. Once finetuned, sampling takes roughly 2", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 568, 498, 582 ], "spans": [ { "bbox": [ 105, 568, 498, 582 ], "score": 1.0, "content": "minutes on similar hardware. Speeding it up will allow Dreamix to be used for more applications.", "type": "text" } ], "index": 36 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 586, 505, 663 ], "lines": [ { "bbox": [ 106, 586, 505, 599 ], "spans": [ { "bbox": [ 106, 586, 505, 599 ], "score": 1.0, "content": "Comparison to Image-based Methods. Dreamix uses VDMs while previous approaches used", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "image-level methods. As VDMs are nascent and have lower resolution than image DMs, this", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "presents an interesting trade-off. Dreamix has the ability to edit motion and has high temporal", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 620, 504, 631 ], "spans": [ { "bbox": [ 105, 620, 504, 631 ], "score": 1.0, "content": "consistency, while previous methods e.g., PnP and Tune-a-Video, can have higher spatial resolution.", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 630, 505, 643 ], "spans": [ { "bbox": [ 105, 630, 505, 643 ], "score": 1.0, "content": "Although Tune-a-Video can achieve high alignment for texture editing on videos with limited mo-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 641, 506, 654 ], "spans": [ { "bbox": [ 105, 641, 506, 654 ], "score": 1.0, "content": "tion, it suffers from poor temporal consistency (see SM). This highlights the importance of using a", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 651, 425, 666 ], "spans": [ { "bbox": [ 106, 651, 425, 666 ], "score": 1.0, "content": "VDM backbone that provides temporal consistency and enables motion editing.", "type": "text" } ], "index": 43 } ], "index": 40 }, { "type": "title", "bbox": [ 108, 676, 195, 689 ], "lines": [ { "bbox": [ 104, 674, 197, 693 ], "spans": [ { "bbox": [ 104, 674, 197, 693 ], "score": 1.0, "content": "8 CONCLUSION", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 107, 699, 505, 710 ], "spans": [ { "bbox": [ 107, 699, 505, 710 ], "score": 1.0, "content": "We presented the first diffusion-based method that can edit motion in real-world videos. Our method", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "can be applied to image animation and subject-driven video generation. Extensive experiments", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 720, 347, 732 ], "spans": [ { "bbox": [ 106, 720, 347, 732 ], "score": 1.0, "content": "demonstrated the unprecedented capabilities of our method.", "type": "text" } ], "index": 47 } ], "index": 46 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 25, 308, 38 ], "spans": [ { "bbox": [ 107, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 194, 144, 417, 219 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 80, 505, 136 ], "group_id": 0, "lines": [ { "bbox": [ 106, 81, 505, 92 ], "spans": [ { "bbox": [ 106, 81, 505, 92 ], "score": 1.0, "content": "Table 3: Ablation Study: Left: Users were asked to compare text-guided video edits of with (w/", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 505, 104 ], "spans": [ { "bbox": [ 105, 90, 505, 104 ], "score": 1.0, "content": "Ft) and without (w/o Ft) finetuning. “None“ indicates failure of both methods according to user.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 101, 505, 115 ], "spans": [ { "bbox": [ 105, 101, 505, 115 ], "score": 1.0, "content": "Apart from style-based edits, where high fidelity is not needed, finetuning significantly improves", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 111, 506, 127 ], "spans": [ { "bbox": [ 105, 111, 506, 127 ], "score": 1.0, "content": "the results. Right: Users were asked to compare video finetuning (Vid) with mixed video-image", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 124, 443, 136 ], "spans": [ { "bbox": [ 106, 124, 443, 136 ], "score": 1.0, "content": "finetuning (Mix). Mixed finetuning significantly improves the results for most cases", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table_body", "bbox": [ 194, 144, 417, 219 ], "group_id": 0, "lines": [ { "bbox": [ 194, 144, 417, 219 ], "spans": [ { "bbox": [ 194, 144, 417, 219 ], "score": 0.977, "html": "
Tyiedifs w/o Ft. w/Ft. NoneVid Mix
Motion3617% 72%11%35% 65%
Object4436% 48%16%62% 38%
Background3219% 77%9%36% 64%
Style1567% 27%6%26% 74%
", "type": "table", "image_path": "9728b5e248d5f000cf2ee7bff5d5c693a4f5e6ea6b15d5d6680da7ba7ae54296.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 194, 144, 417, 159.0 ], "spans": [], "index": 5 }, { "bbox": [ 194, 159.0, 417, 174.0 ], "spans": [], "index": 6 }, { "bbox": [ 194, 174.0, 417, 189.0 ], "spans": [], "index": 7 }, { "bbox": [ 194, 189.0, 417, 204.0 ], "spans": [], "index": 8 }, { "bbox": [ 194, 204.0, 417, 219.0 ], "spans": [], "index": 9 } ] } ], "index": 4.5 }, { "type": "text", "bbox": [ 106, 230, 505, 296 ], "lines": [ { "bbox": [ 106, 230, 505, 243 ], "spans": [ { "bbox": [ 106, 230, 505, 243 ], "score": 1.0, "content": "The evaluation and user study are presented in Tab. 2 and Tab. 1. Image-based methods (PnP,", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 240, 505, 254 ], "spans": [ { "bbox": [ 105, 240, 505, 254 ], "score": 1.0, "content": "Tune-a-Video) exhibit impaired temporal consistency, resulting in low quality. Moreover, they are", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 251, 505, 265 ], "spans": [ { "bbox": [ 105, 251, 505, 265 ], "score": 1.0, "content": "unable to perform motion edits, resulting in poor alignment and high fidelity. Video-based methods", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "maintain temporal consistency while allowing motion editing. Although unconditional generation", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 273, 506, 286 ], "spans": [ { "bbox": [ 106, 273, 506, 286 ], "score": 1.0, "content": "outperforms our method in the automatic evaluations (Tab. 2), it has poor fidelity (Tab. 1) as it is not", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 285, 430, 297 ], "spans": [ { "bbox": [ 106, 285, 430, 297 ], "score": 1.0, "content": "conditioned on the input video. Overall, our method has the highest success rate.", "type": "text" } ], "index": 15 } ], "index": 12.5, "bbox_fs": [ 105, 230, 506, 297 ] }, { "type": "text", "bbox": [ 107, 301, 505, 368 ], "lines": [ { "bbox": [ 106, 302, 505, 315 ], "spans": [ { "bbox": [ 106, 302, 505, 315 ], "score": 1.0, "content": "Qualitative Comparison. Figure 7 presents an example of motion editing by Dreamix compared", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 313, 505, 325 ], "spans": [ { "bbox": [ 105, 313, 505, 325 ], "score": 1.0, "content": "to the baselines. The text-to-video model achieves low fidelity edits as it is not conditioned on", "type": "text" } ], "index": 17 }, { "bbox": [ 104, 322, 505, 338 ], "spans": [ { "bbox": [ 104, 322, 505, 338 ], "score": 1.0, "content": "the original video. PnP preserves the scene but fails to perform the edit and lacks consistency", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 505, 347 ], "score": 1.0, "content": "between different frames. Tune-a-Video exhibits better temporal consistency but still fails to perform", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "the motion edit. Dreamix performs well on all three objectives, adding the desired motion while", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 357, 252, 369 ], "spans": [ { "bbox": [ 105, 357, 252, 369 ], "score": 1.0, "content": "preserving fidelity and high-quality.", "type": "text" } ], "index": 21 } ], "index": 18.5, "bbox_fs": [ 104, 302, 505, 369 ] }, { "type": "title", "bbox": [ 108, 378, 209, 389 ], "lines": [ { "bbox": [ 106, 377, 210, 391 ], "spans": [ { "bbox": [ 106, 377, 210, 391 ], "score": 1.0, "content": "6.3 ABLATION STUDY", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 396, 505, 495 ], "lines": [ { "bbox": [ 106, 395, 505, 409 ], "spans": [ { "bbox": [ 106, 395, 505, 409 ], "score": 1.0, "content": "We ablate the use of finetuning and the mixed video-image finetuning by performing a user study", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 406, 505, 420 ], "spans": [ { "bbox": [ 105, 406, 505, 420 ], "score": 1.0, "content": "using the dataset described above. The ablation indeed supports the idea of using finetuning in cases", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 417, 505, 431 ], "spans": [ { "bbox": [ 105, 417, 505, 431 ], "score": 1.0, "content": "where high-editability is required. We can see that Motion changes require high-editability and are", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 429, 505, 442 ], "spans": [ { "bbox": [ 105, 429, 505, 442 ], "score": 1.0, "content": "thus improved by finetuning. Moreover, as the noising corrupts the video, preserving fine-details", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 438, 505, 454 ], "spans": [ { "bbox": [ 105, 438, 505, 454 ], "score": 1.0, "content": "in background, color or texture edits requires finetuning. In contrast, denoising without finetuning", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 451, 504, 463 ], "spans": [ { "bbox": [ 106, 451, 504, 463 ], "score": 1.0, "content": "worked well for style edits, where finetuning was often detrimental. This is expected as style edits", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 462, 506, 475 ], "spans": [ { "bbox": [ 105, 462, 506, 475 ], "score": 1.0, "content": "are often conflicted with high fidelity preservation (e.g. changing the texture of an object means", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 472, 506, 486 ], "spans": [ { "bbox": [ 105, 472, 506, 486 ], "score": 1.0, "content": "reducing fidelity). The ablation shows that in most cases mixed finetuning improves the results by a", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 484, 460, 497 ], "spans": [ { "bbox": [ 105, 484, 460, 497 ], "score": 1.0, "content": "wide margin. Results are presented in Tab. 3, a visual ablation is found in Appendix F.2.", "type": "text" } ], "index": 31 } ], "index": 27, "bbox_fs": [ 105, 395, 506, 497 ] }, { "type": "title", "bbox": [ 108, 508, 194, 521 ], "lines": [ { "bbox": [ 105, 507, 195, 524 ], "spans": [ { "bbox": [ 105, 507, 195, 524 ], "score": 1.0, "content": "7 LIMITATIONS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 108, 530, 488, 541 ], "lines": [ { "bbox": [ 107, 529, 489, 542 ], "spans": [ { "bbox": [ 107, 529, 489, 542 ], "score": 1.0, "content": "While Dreamix is the first diffusion-based video method that can edit motion, it has limitations.", "type": "text" } ], "index": 33 } ], "index": 33, "bbox_fs": [ 107, 529, 489, 542 ] }, { "type": "text", "bbox": [ 106, 547, 504, 581 ], "lines": [ { "bbox": [ 106, 547, 505, 560 ], "spans": [ { "bbox": [ 106, 547, 505, 560 ], "score": 1.0, "content": "Computational Cost. VDMs are computationally expensive. Finetuning our model using 4 TPU", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 558, 505, 571 ], "spans": [ { "bbox": [ 106, 558, 505, 571 ], "score": 1.0, "content": "v4 accelerators requires around 30 minutes per video. Once finetuned, sampling takes roughly 2", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 568, 498, 582 ], "spans": [ { "bbox": [ 105, 568, 498, 582 ], "score": 1.0, "content": "minutes on similar hardware. Speeding it up will allow Dreamix to be used for more applications.", "type": "text" } ], "index": 36 } ], "index": 35, "bbox_fs": [ 105, 547, 505, 582 ] }, { "type": "text", "bbox": [ 107, 586, 505, 663 ], "lines": [ { "bbox": [ 106, 586, 505, 599 ], "spans": [ { "bbox": [ 106, 586, 505, 599 ], "score": 1.0, "content": "Comparison to Image-based Methods. Dreamix uses VDMs while previous approaches used", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "image-level methods. As VDMs are nascent and have lower resolution than image DMs, this", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "presents an interesting trade-off. Dreamix has the ability to edit motion and has high temporal", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 620, 504, 631 ], "spans": [ { "bbox": [ 105, 620, 504, 631 ], "score": 1.0, "content": "consistency, while previous methods e.g., PnP and Tune-a-Video, can have higher spatial resolution.", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 630, 505, 643 ], "spans": [ { "bbox": [ 105, 630, 505, 643 ], "score": 1.0, "content": "Although Tune-a-Video can achieve high alignment for texture editing on videos with limited mo-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 641, 506, 654 ], "spans": [ { "bbox": [ 105, 641, 506, 654 ], "score": 1.0, "content": "tion, it suffers from poor temporal consistency (see SM). This highlights the importance of using a", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 651, 425, 666 ], "spans": [ { "bbox": [ 106, 651, 425, 666 ], "score": 1.0, "content": "VDM backbone that provides temporal consistency and enables motion editing.", "type": "text" } ], "index": 43 } ], "index": 40, "bbox_fs": [ 105, 586, 506, 666 ] }, { "type": "title", "bbox": [ 108, 676, 195, 689 ], "lines": [ { "bbox": [ 104, 674, 197, 693 ], "spans": [ { "bbox": [ 104, 674, 197, 693 ], "score": 1.0, "content": "8 CONCLUSION", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 107, 699, 505, 710 ], "spans": [ { "bbox": [ 107, 699, 505, 710 ], "score": 1.0, "content": "We presented the first diffusion-based method that can edit motion in real-world videos. Our method", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "can be applied to image animation and subject-driven video generation. Extensive experiments", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 720, 347, 732 ], "spans": [ { "bbox": [ 106, 720, 347, 732 ], "score": 1.0, "content": "demonstrated the unprecedented capabilities of our method.", "type": "text" } ], "index": 47 } ], "index": 46, "bbox_fs": [ 106, 699, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 100, 68, 507, 734 ], "lines": [ { "bbox": [ 106, 80, 176, 94 ], "spans": [ { "bbox": [ 106, 80, 176, 94 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 99, 506, 113 ], "spans": [ { "bbox": [ 106, 99, 506, 113 ], "score": 1.0, "content": "Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan", "type": "text" } ], "index": 1 }, { "bbox": [ 114, 109, 507, 126 ], "spans": [ { "bbox": [ 114, 109, 351, 126 ], "score": 1.0, "content": "Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-", "type": "text" }, { "bbox": [ 351, 111, 365, 122 ], "score": 0.32, "content": ". 8 \\mathrm { m }", "type": "inline_equation" }, { "bbox": [ 366, 109, 507, 126 ], "score": 1.0, "content": ": A large-scale video classification", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 122, 352, 136 ], "spans": [ { "bbox": [ 115, 122, 352, 136 ], "score": 1.0, "content": "benchmark. arXiv preprint arXiv:1609.08675, 2016. 8, 15", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 141, 506, 155 ], "spans": [ { "bbox": [ 106, 141, 506, 155 ], "score": 1.0, "content": "Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint", "type": "text" } ], "index": 4 }, { "bbox": [ 114, 152, 233, 166 ], "spans": [ { "bbox": [ 114, 152, 233, 166 ], "score": 1.0, "content": "arXiv:2206.02779, 2022a. 3", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 171, 506, 185 ], "spans": [ { "bbox": [ 106, 171, 506, 185 ], "score": 1.0, "content": "Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani", "type": "text" } ], "index": 6 }, { "bbox": [ 114, 181, 506, 197 ], "spans": [ { "bbox": [ 114, 181, 506, 197 ], "score": 1.0, "content": "Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 193, 352, 208 ], "spans": [ { "bbox": [ 114, 193, 352, 208 ], "score": 1.0, "content": "generation. arXiv preprint arXiv:2211.14305, 2022b. 1, 2", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 213, 506, 227 ], "spans": [ { "bbox": [ 106, 213, 506, 227 ], "score": 1.0, "content": "Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of", "type": "text" } ], "index": 9 }, { "bbox": [ 114, 224, 506, 238 ], "spans": [ { "bbox": [ 114, 224, 506, 238 ], "score": 1.0, "content": "natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 10 }, { "bbox": [ 114, 235, 284, 248 ], "spans": [ { "bbox": [ 114, 235, 284, 248 ], "score": 1.0, "content": "Recognition, pp. 18208–18218, 2022c. 3", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 254, 505, 268 ], "spans": [ { "bbox": [ 106, 254, 505, 268 ], "score": 1.0, "content": "Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 265, 504, 279 ], "spans": [ { "bbox": [ 115, 265, 504, 279 ], "score": 1.0, "content": "driven layered image and video editing. In Computer Vision–ECCV 2022: 17th European Con-", "type": "text" } ], "index": 13 }, { "bbox": [ 112, 274, 506, 291 ], "spans": [ { "bbox": [ 112, 274, 506, 291 ], "score": 1.0, "content": "ference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pp. 707–723. Springer,", "type": "text" } ], "index": 14 }, { "bbox": [ 114, 286, 151, 300 ], "spans": [ { "bbox": [ 114, 286, 151, 300 ], "score": 1.0, "content": "2022. 3", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 307, 504, 319 ], "spans": [ { "bbox": [ 106, 307, 504, 319 ], "score": 1.0, "content": "David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and An-", "type": "text" } ], "index": 16 }, { "bbox": [ 114, 317, 506, 331 ], "spans": [ { "bbox": [ 114, 317, 506, 331 ], "score": 1.0, "content": "tonio Torralba. Semantic photo manipulation with a generative image prior. arXiv preprint", "type": "text" } ], "index": 17 }, { "bbox": [ 114, 329, 230, 342 ], "spans": [ { "bbox": [ 114, 329, 230, 342 ], "score": 1.0, "content": "arXiv:2005.07727, 2020. 3", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 348, 505, 362 ], "spans": [ { "bbox": [ 106, 348, 505, 362 ], "score": 1.0, "content": "Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 360, 381, 372 ], "spans": [ { "bbox": [ 116, 360, 381, 372 ], "score": 1.0, "content": "editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 3", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 378, 506, 392 ], "spans": [ { "bbox": [ 104, 378, 506, 392 ], "score": 1.0, "content": "Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative", "type": "text" } ], "index": 21 }, { "bbox": [ 114, 388, 506, 404 ], "spans": [ { "bbox": [ 114, 388, 506, 404 ], "score": 1.0, "content": "image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 400, 278, 414 ], "spans": [ { "bbox": [ 114, 400, 278, 414 ], "score": 1.0, "content": "Recognition, pp. 11315–11325, 2022. 1", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 419, 506, 433 ], "spans": [ { "bbox": [ 106, 419, 506, 433 ], "score": 1.0, "content": "Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 430, 505, 445 ], "spans": [ { "bbox": [ 116, 430, 505, 445 ], "score": 1.0, "content": "Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image gen-", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 442, 464, 454 ], "spans": [ { "bbox": [ 114, 442, 464, 454 ], "score": 1.0, "content": "eration via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 461, 506, 475 ], "spans": [ { "bbox": [ 106, 461, 506, 475 ], "score": 1.0, "content": "Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating con-", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 470, 506, 487 ], "spans": [ { "bbox": [ 114, 470, 506, 487 ], "score": 1.0, "content": "ditional diffusion models for inverse problems through stochastic contraction. In Proceedings of", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 483, 504, 497 ], "spans": [ { "bbox": [ 115, 483, 504, 497 ], "score": 1.0, "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12413–12422, 2022.", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 495, 124, 505 ], "spans": [ { "bbox": [ 116, 495, 124, 505 ], "score": 1.0, "content": "2", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 512, 506, 527 ], "spans": [ { "bbox": [ 105, 512, 506, 527 ], "score": 1.0, "content": "Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 525, 386, 537 ], "spans": [ { "bbox": [ 116, 525, 386, 537 ], "score": 1.0, "content": "in Neural Information Processing Systems, 34:8780–8794, 2021. 2", "type": "text" } ], "index": 32 }, { "bbox": [ 104, 541, 505, 559 ], "spans": [ { "bbox": [ 104, 541, 505, 559 ], "score": 1.0, "content": "Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 554, 480, 569 ], "spans": [ { "bbox": [ 114, 554, 480, 569 ], "score": 1.0, "content": "guided domain adaptation of image generators. arXiv preprint arXiv:2108.00946, 2021. 3", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 574, 506, 587 ], "spans": [ { "bbox": [ 106, 574, 506, 587 ], "score": 1.0, "content": "Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel", "type": "text" } ], "index": 35 }, { "bbox": [ 114, 583, 506, 600 ], "spans": [ { "bbox": [ 114, 583, 506, 600 ], "score": 1.0, "content": "Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 596, 331, 610 ], "spans": [ { "bbox": [ 115, 596, 331, 610 ], "score": 1.0, "content": "inversion. arXiv preprint arXiv:2208.01618, 2022. 3", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 616, 504, 628 ], "spans": [ { "bbox": [ 106, 616, 504, 628 ], "score": 1.0, "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 626, 505, 640 ], "spans": [ { "bbox": [ 115, 626, 505, 640 ], "score": 1.0, "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 637, 248, 651 ], "spans": [ { "bbox": [ 114, 637, 248, 651 ], "score": 1.0, "content": "ACM, 63(11):139–144, 2020. 2", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 657, 505, 670 ], "spans": [ { "bbox": [ 106, 657, 505, 670 ], "score": 1.0, "content": "Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or.", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 667, 506, 683 ], "spans": [ { "bbox": [ 114, 667, 506, 683 ], "score": 1.0, "content": "Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626,", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 678, 172, 693 ], "spans": [ { "bbox": [ 115, 678, 172, 693 ], "score": 1.0, "content": "2022. 1, 2, 3", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 709, 505, 723 ], "spans": [ { "bbox": [ 114, 709, 505, 723 ], "score": 1.0, "content": "reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.", "type": "text" } ], "index": 45 }, { "bbox": [ 115, 720, 140, 733 ], "spans": [ { "bbox": [ 115, 720, 140, 733 ], "score": 1.0, "content": "8, 17", "type": "text" } ], "index": 46 } ], "index": 23 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "10", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 100, 68, 507, 734 ], "lines": [ { "bbox": [ 106, 80, 176, 94 ], "spans": [ { "bbox": [ 106, 80, 176, 94 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 0, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 106, 99, 506, 113 ], "spans": [ { "bbox": [ 106, 99, 506, 113 ], "score": 1.0, "content": "Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan", "type": "text" } ], "index": 1, "is_list_start_line": true }, { "bbox": [ 114, 109, 507, 126 ], "spans": [ { "bbox": [ 114, 109, 351, 126 ], "score": 1.0, "content": "Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-", "type": "text" }, { "bbox": [ 351, 111, 365, 122 ], "score": 0.32, "content": ". 8 \\mathrm { m }", "type": "inline_equation" }, { "bbox": [ 366, 109, 507, 126 ], "score": 1.0, "content": ": A large-scale video classification", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 122, 352, 136 ], "spans": [ { "bbox": [ 115, 122, 352, 136 ], "score": 1.0, "content": "benchmark. arXiv preprint arXiv:1609.08675, 2016. 8, 15", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 106, 141, 506, 155 ], "spans": [ { "bbox": [ 106, 141, 506, 155 ], "score": 1.0, "content": "Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint", "type": "text" } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 114, 152, 233, 166 ], "spans": [ { "bbox": [ 114, 152, 233, 166 ], "score": 1.0, "content": "arXiv:2206.02779, 2022a. 3", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 106, 171, 506, 185 ], "spans": [ { "bbox": [ 106, 171, 506, 185 ], "score": 1.0, "content": "Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 114, 181, 506, 197 ], "spans": [ { "bbox": [ 114, 181, 506, 197 ], "score": 1.0, "content": "Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 193, 352, 208 ], "spans": [ { "bbox": [ 114, 193, 352, 208 ], "score": 1.0, "content": "generation. arXiv preprint arXiv:2211.14305, 2022b. 1, 2", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 106, 213, 506, 227 ], "spans": [ { "bbox": [ 106, 213, 506, 227 ], "score": 1.0, "content": "Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 114, 224, 506, 238 ], "spans": [ { "bbox": [ 114, 224, 506, 238 ], "score": 1.0, "content": "natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 10 }, { "bbox": [ 114, 235, 284, 248 ], "spans": [ { "bbox": [ 114, 235, 284, 248 ], "score": 1.0, "content": "Recognition, pp. 18208–18218, 2022c. 3", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 106, 254, 505, 268 ], "spans": [ { "bbox": [ 106, 254, 505, 268 ], "score": 1.0, "content": "Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 115, 265, 504, 279 ], "spans": [ { "bbox": [ 115, 265, 504, 279 ], "score": 1.0, "content": "driven layered image and video editing. In Computer Vision–ECCV 2022: 17th European Con-", "type": "text" } ], "index": 13 }, { "bbox": [ 112, 274, 506, 291 ], "spans": [ { "bbox": [ 112, 274, 506, 291 ], "score": 1.0, "content": "ference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pp. 707–723. Springer,", "type": "text" } ], "index": 14 }, { "bbox": [ 114, 286, 151, 300 ], "spans": [ { "bbox": [ 114, 286, 151, 300 ], "score": 1.0, "content": "2022. 3", "type": "text" } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 106, 307, 504, 319 ], "spans": [ { "bbox": [ 106, 307, 504, 319 ], "score": 1.0, "content": "David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and An-", "type": "text" } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 114, 317, 506, 331 ], "spans": [ { "bbox": [ 114, 317, 506, 331 ], "score": 1.0, "content": "tonio Torralba. Semantic photo manipulation with a generative image prior. arXiv preprint", "type": "text" } ], "index": 17 }, { "bbox": [ 114, 329, 230, 342 ], "spans": [ { "bbox": [ 114, 329, 230, 342 ], "score": 1.0, "content": "arXiv:2005.07727, 2020. 3", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 106, 348, 505, 362 ], "spans": [ { "bbox": [ 106, 348, 505, 362 ], "score": 1.0, "content": "Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image", "type": "text" } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 116, 360, 381, 372 ], "spans": [ { "bbox": [ 116, 360, 381, 372 ], "score": 1.0, "content": "editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 3", "type": "text" } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 104, 378, 506, 392 ], "spans": [ { "bbox": [ 104, 378, 506, 392 ], "score": 1.0, "content": "Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative", "type": "text" } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 114, 388, 506, 404 ], "spans": [ { "bbox": [ 114, 388, 506, 404 ], "score": 1.0, "content": "image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 400, 278, 414 ], "spans": [ { "bbox": [ 114, 400, 278, 414 ], "score": 1.0, "content": "Recognition, pp. 11315–11325, 2022. 1", "type": "text" } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 106, 419, 506, 433 ], "spans": [ { "bbox": [ 106, 419, 506, 433 ], "score": 1.0, "content": "Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 116, 430, 505, 445 ], "spans": [ { "bbox": [ 116, 430, 505, 445 ], "score": 1.0, "content": "Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image gen-", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 442, 464, 454 ], "spans": [ { "bbox": [ 114, 442, 464, 454 ], "score": 1.0, "content": "eration via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1", "type": "text" } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 106, 461, 506, 475 ], "spans": [ { "bbox": [ 106, 461, 506, 475 ], "score": 1.0, "content": "Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating con-", "type": "text" } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 114, 470, 506, 487 ], "spans": [ { "bbox": [ 114, 470, 506, 487 ], "score": 1.0, "content": "ditional diffusion models for inverse problems through stochastic contraction. In Proceedings of", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 483, 504, 497 ], "spans": [ { "bbox": [ 115, 483, 504, 497 ], "score": 1.0, "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12413–12422, 2022.", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 495, 124, 505 ], "spans": [ { "bbox": [ 116, 495, 124, 505 ], "score": 1.0, "content": "2", "type": "text" } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 105, 512, 506, 527 ], "spans": [ { "bbox": [ 105, 512, 506, 527 ], "score": 1.0, "content": "Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 116, 525, 386, 537 ], "spans": [ { "bbox": [ 116, 525, 386, 537 ], "score": 1.0, "content": "in Neural Information Processing Systems, 34:8780–8794, 2021. 2", "type": "text" } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 104, 541, 505, 559 ], "spans": [ { "bbox": [ 104, 541, 505, 559 ], "score": 1.0, "content": "Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-", "type": "text" } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 114, 554, 480, 569 ], "spans": [ { "bbox": [ 114, 554, 480, 569 ], "score": 1.0, "content": "guided domain adaptation of image generators. arXiv preprint arXiv:2108.00946, 2021. 3", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 106, 574, 506, 587 ], "spans": [ { "bbox": [ 106, 574, 506, 587 ], "score": 1.0, "content": "Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 114, 583, 506, 600 ], "spans": [ { "bbox": [ 114, 583, 506, 600 ], "score": 1.0, "content": "Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 596, 331, 610 ], "spans": [ { "bbox": [ 115, 596, 331, 610 ], "score": 1.0, "content": "inversion. arXiv preprint arXiv:2208.01618, 2022. 3", "type": "text" } ], "index": 37, "is_list_end_line": true }, { "bbox": [ 106, 616, 504, 628 ], "spans": [ { "bbox": [ 106, 616, 504, 628 ], "score": 1.0, "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", "type": "text" } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 115, 626, 505, 640 ], "spans": [ { "bbox": [ 115, 626, 505, 640 ], "score": 1.0, "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 637, 248, 651 ], "spans": [ { "bbox": [ 114, 637, 248, 651 ], "score": 1.0, "content": "ACM, 63(11):139–144, 2020. 2", "type": "text" } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 106, 657, 505, 670 ], "spans": [ { "bbox": [ 106, 657, 505, 670 ], "score": 1.0, "content": "Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or.", "type": "text" } ], "index": 41, "is_list_start_line": true }, { "bbox": [ 114, 667, 506, 683 ], "spans": [ { "bbox": [ 114, 667, 506, 683 ], "score": 1.0, "content": "Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626,", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 678, 172, 693 ], "spans": [ { "bbox": [ 115, 678, 172, 693 ], "score": 1.0, "content": "2022. 1, 2, 3", "type": "text" } ], "index": 43, "is_list_end_line": true }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A", "type": "text" } ], "index": 44, "is_list_start_line": true }, { "bbox": [ 114, 709, 505, 723 ], "spans": [ { "bbox": [ 114, 709, 505, 723 ], "score": 1.0, "content": "reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.", "type": "text" } ], "index": 45 }, { "bbox": [ 115, 720, 140, 733 ], "spans": [ { "bbox": [ 115, 720, 140, 733 ], "score": 1.0, "content": "8, 17", "type": "text" } ], "index": 46, "is_list_end_line": true } ], "index": 23, "bbox_fs": [ 104, 80, 507, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 81, 505, 95 ], "spans": [ { "bbox": [ 105, 81, 505, 95 ], "score": 1.0, "content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 387, 106 ], "spans": [ { "bbox": [ 115, 93, 387, 106 ], "score": 1.0, "content": "Neural Information Processing Systems, 33:6840–6851, 2020. 1, 2", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 100, 96, 506, 739 ], "lines": [ { "bbox": [ 104, 111, 506, 127 ], "spans": [ { "bbox": [ 104, 111, 506, 127 ], "score": 1.0, "content": "Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 124, 506, 137 ], "spans": [ { "bbox": [ 115, 124, 506, 137 ], "score": 1.0, "content": "Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 135, 506, 150 ], "spans": [ { "bbox": [ 115, 135, 506, 150 ], "score": 1.0, "content": "video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022a. 1, 2, 4, 5, 14,", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 146, 129, 158 ], "spans": [ { "bbox": [ 117, 146, 129, 158 ], "score": 1.0, "content": "15", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 166, 504, 180 ], "spans": [ { "bbox": [ 105, 166, 504, 180 ], "score": 1.0, "content": "Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Sali-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 176, 505, 192 ], "spans": [ { "bbox": [ 115, 176, 505, 192 ], "score": 1.0, "content": "mans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 188, 181, 202 ], "spans": [ { "bbox": [ 116, 188, 181, 202 ], "score": 1.0, "content": "47–1, 2022b. 2", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 208, 505, 222 ], "spans": [ { "bbox": [ 105, 208, 505, 222 ], "score": 1.0, "content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 219, 437, 232 ], "spans": [ { "bbox": [ 116, 219, 437, 232 ], "score": 1.0, "content": "Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022c. 1, 2, 5", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 239, 505, 252 ], "spans": [ { "bbox": [ 105, 239, 505, 252 ], "score": 1.0, "content": "Eliahu Horwitz and Yedid Hoshen. Conffusion: Confidence intervals for diffusion models. arXiv", "type": "text" } ], "index": 11 }, { "bbox": [ 112, 249, 265, 264 ], "spans": [ { "bbox": [ 112, 249, 265, 264 ], "score": 1.0, "content": "preprint arXiv:2211.09795, 2022. 2", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 270, 504, 283 ], "spans": [ { "bbox": [ 105, 270, 504, 283 ], "score": 1.0, "content": "Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score match-¨", "type": "text" } ], "index": 13 }, { "bbox": [ 114, 280, 353, 295 ], "spans": [ { "bbox": [ 114, 280, 353, 295 ], "score": 1.0, "content": "ing. Journal of Machine Learning Research, 6(4), 2005. 2", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 301, 504, 315 ], "spans": [ { "bbox": [ 105, 301, 504, 315 ], "score": 1.0, "content": "Ondˇrej Jamriska, ˇ Sˇ arka Sochorov ´ a, Ond ´ ˇrej Texler, Michal Luka´c, Jakub Fi ˇ ser, Jingwan Lu, Eliˇ", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 313, 505, 326 ], "spans": [ { "bbox": [ 116, 313, 505, 326 ], "score": 1.0, "content": "Shechtman, and Daniel Sykora. Stylizing video by example. ´ ACM Transactions on Graphics, 38", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 324, 168, 338 ], "spans": [ { "bbox": [ 116, 324, 168, 338 ], "score": 1.0, "content": "(4), 2019. 3", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 342, 505, 358 ], "spans": [ { "bbox": [ 105, 342, 505, 358 ], "score": 1.0, "content": "Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-", "type": "text" } ], "index": 18 }, { "bbox": [ 114, 354, 402, 369 ], "spans": [ { "bbox": [ 114, 354, 402, 369 ], "score": 1.0, "content": "based generative models. arXiv preprint arXiv:2206.00364, 2022. 2, 3", "type": "text" } ], "index": 19 }, { "bbox": [ 107, 374, 505, 388 ], "spans": [ { "bbox": [ 107, 374, 505, 388 ], "score": 1.0, "content": "Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video", "type": "text" } ], "index": 20 }, { "bbox": [ 116, 385, 394, 399 ], "spans": [ { "bbox": [ 116, 385, 394, 399 ], "score": 1.0, "content": "editing. ACM Transactions on Graphics (TOG), 40(6):1–12, 2021. 3", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 405, 505, 419 ], "spans": [ { "bbox": [ 105, 405, 505, 419 ], "score": 1.0, "content": "Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri,", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 417, 505, 430 ], "spans": [ { "bbox": [ 115, 417, 505, 430 ], "score": 1.0, "content": "and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint", "type": "text" } ], "index": 23 }, { "bbox": [ 116, 427, 239, 441 ], "spans": [ { "bbox": [ 116, 427, 239, 441 ], "score": 1.0, "content": "arXiv:2210.09276, 2022. 1, 3", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 447, 504, 460 ], "spans": [ { "bbox": [ 105, 447, 504, 460 ], "score": 1.0, "content": "Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao.", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 458, 504, 472 ], "spans": [ { "bbox": [ 116, 458, 504, 472 ], "score": 1.0, "content": "DeepFaceVideoEditing: Sketch-based deep editing of face videos. ACM Transactions on Graph-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 469, 252, 482 ], "spans": [ { "bbox": [ 115, 469, 252, 482 ], "score": 1.0, "content": "ics, 41(4):167:1–167:16, 2022. 3", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 488, 505, 504 ], "spans": [ { "bbox": [ 105, 488, 505, 504 ], "score": 1.0, "content": "Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool.", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 499, 506, 515 ], "spans": [ { "bbox": [ 115, 499, 506, 515 ], "score": 1.0, "content": "Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 511, 505, 526 ], "spans": [ { "bbox": [ 114, 511, 505, 526 ], "score": 1.0, "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11461–11471, 2022.", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 522, 126, 534 ], "spans": [ { "bbox": [ 115, 522, 126, 534 ], "score": 1.0, "content": "2", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 541, 504, 555 ], "spans": [ { "bbox": [ 105, 541, 504, 555 ], "score": 1.0, "content": "Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Im-", "type": "text" } ], "index": 32 }, { "bbox": [ 114, 552, 505, 568 ], "spans": [ { "bbox": [ 114, 552, 505, 568 ], "score": 1.0, "content": "age synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073,", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 562, 171, 578 ], "spans": [ { "bbox": [ 115, 562, 171, 578 ], "score": 1.0, "content": "2021. 2, 3, 4", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 583, 505, 596 ], "spans": [ { "bbox": [ 105, 583, 505, 596 ], "score": 1.0, "content": "Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for", "type": "text" } ], "index": 35 }, { "bbox": [ 116, 595, 493, 609 ], "spans": [ { "bbox": [ 116, 595, 493, 609 ], "score": 1.0, "content": "editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 614, 506, 628 ], "spans": [ { "bbox": [ 105, 614, 506, 628 ], "score": 1.0, "content": "Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and ˜", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 626, 506, 639 ], "spans": [ { "bbox": [ 116, 626, 506, 639 ], "score": 1.0, "content": "Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 636, 229, 650 ], "spans": [ { "bbox": [ 115, 636, 229, 650 ], "score": 1.0, "content": "arXiv:2203.09494, 2022. 3", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 656, 504, 669 ], "spans": [ { "bbox": [ 105, 656, 504, 669 ], "score": 1.0, "content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew,", "type": "text" } ], "index": 40 }, { "bbox": [ 114, 667, 506, 682 ], "spans": [ { "bbox": [ 114, 667, 506, 682 ], "score": 1.0, "content": "Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 678, 408, 691 ], "spans": [ { "bbox": [ 114, 678, 408, 691 ], "score": 1.0, "content": "text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 708, 506, 724 ], "spans": [ { "bbox": [ 115, 708, 506, 724 ], "score": 1.0, "content": "spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 721, 315, 734 ], "spans": [ { "bbox": [ 115, 721, 315, 734 ], "score": 1.0, "content": "and pattern recognition, pp. 2337–2346, 2019. 3", "type": "text" } ], "index": 45 } ], "index": 23.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 81, 505, 95 ], "spans": [ { "bbox": [ 105, 81, 505, 95 ], "score": 1.0, "content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 387, 106 ], "spans": [ { "bbox": [ 115, 93, 387, 106 ], "score": 1.0, "content": "Neural Information Processing Systems, 33:6840–6851, 2020. 1, 2", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 81, 505, 106 ] }, { "type": "list", "bbox": [ 100, 96, 506, 739 ], "lines": [ { "bbox": [ 104, 111, 506, 127 ], "spans": [ { "bbox": [ 104, 111, 506, 127 ], "score": 1.0, "content": "Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 115, 124, 506, 137 ], "spans": [ { "bbox": [ 115, 124, 506, 137 ], "score": 1.0, "content": "Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 135, 506, 150 ], "spans": [ { "bbox": [ 115, 135, 506, 150 ], "score": 1.0, "content": "video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022a. 1, 2, 4, 5, 14,", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 146, 129, 158 ], "spans": [ { "bbox": [ 117, 146, 129, 158 ], "score": 1.0, "content": "15", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 105, 166, 504, 180 ], "spans": [ { "bbox": [ 105, 166, 504, 180 ], "score": 1.0, "content": "Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Sali-", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 115, 176, 505, 192 ], "spans": [ { "bbox": [ 115, 176, 505, 192 ], "score": 1.0, "content": "mans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 188, 181, 202 ], "spans": [ { "bbox": [ 116, 188, 181, 202 ], "score": 1.0, "content": "47–1, 2022b. 2", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 105, 208, 505, 222 ], "spans": [ { "bbox": [ 105, 208, 505, 222 ], "score": 1.0, "content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 116, 219, 437, 232 ], "spans": [ { "bbox": [ 116, 219, 437, 232 ], "score": 1.0, "content": "Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022c. 1, 2, 5", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 105, 239, 505, 252 ], "spans": [ { "bbox": [ 105, 239, 505, 252 ], "score": 1.0, "content": "Eliahu Horwitz and Yedid Hoshen. Conffusion: Confidence intervals for diffusion models. arXiv", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 112, 249, 265, 264 ], "spans": [ { "bbox": [ 112, 249, 265, 264 ], "score": 1.0, "content": "preprint arXiv:2211.09795, 2022. 2", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 105, 270, 504, 283 ], "spans": [ { "bbox": [ 105, 270, 504, 283 ], "score": 1.0, "content": "Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score match-¨", "type": "text" } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 114, 280, 353, 295 ], "spans": [ { "bbox": [ 114, 280, 353, 295 ], "score": 1.0, "content": "ing. Journal of Machine Learning Research, 6(4), 2005. 2", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 105, 301, 504, 315 ], "spans": [ { "bbox": [ 105, 301, 504, 315 ], "score": 1.0, "content": "Ondˇrej Jamriska, ˇ Sˇ arka Sochorov ´ a, Ond ´ ˇrej Texler, Michal Luka´c, Jakub Fi ˇ ser, Jingwan Lu, Eliˇ", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 116, 313, 505, 326 ], "spans": [ { "bbox": [ 116, 313, 505, 326 ], "score": 1.0, "content": "Shechtman, and Daniel Sykora. Stylizing video by example. ´ ACM Transactions on Graphics, 38", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 324, 168, 338 ], "spans": [ { "bbox": [ 116, 324, 168, 338 ], "score": 1.0, "content": "(4), 2019. 3", "type": "text" } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 105, 342, 505, 358 ], "spans": [ { "bbox": [ 105, 342, 505, 358 ], "score": 1.0, "content": "Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-", "type": "text" } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 114, 354, 402, 369 ], "spans": [ { "bbox": [ 114, 354, 402, 369 ], "score": 1.0, "content": "based generative models. arXiv preprint arXiv:2206.00364, 2022. 2, 3", "type": "text" } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 107, 374, 505, 388 ], "spans": [ { "bbox": [ 107, 374, 505, 388 ], "score": 1.0, "content": "Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video", "type": "text" } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 116, 385, 394, 399 ], "spans": [ { "bbox": [ 116, 385, 394, 399 ], "score": 1.0, "content": "editing. ACM Transactions on Graphics (TOG), 40(6):1–12, 2021. 3", "type": "text" } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 105, 405, 505, 419 ], "spans": [ { "bbox": [ 105, 405, 505, 419 ], "score": 1.0, "content": "Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri,", "type": "text" } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 115, 417, 505, 430 ], "spans": [ { "bbox": [ 115, 417, 505, 430 ], "score": 1.0, "content": "and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint", "type": "text" } ], "index": 23 }, { "bbox": [ 116, 427, 239, 441 ], "spans": [ { "bbox": [ 116, 427, 239, 441 ], "score": 1.0, "content": "arXiv:2210.09276, 2022. 1, 3", "type": "text" } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 105, 447, 504, 460 ], "spans": [ { "bbox": [ 105, 447, 504, 460 ], "score": 1.0, "content": "Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao.", "type": "text" } ], "index": 25, "is_list_start_line": true }, { "bbox": [ 116, 458, 504, 472 ], "spans": [ { "bbox": [ 116, 458, 504, 472 ], "score": 1.0, "content": "DeepFaceVideoEditing: Sketch-based deep editing of face videos. ACM Transactions on Graph-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 469, 252, 482 ], "spans": [ { "bbox": [ 115, 469, 252, 482 ], "score": 1.0, "content": "ics, 41(4):167:1–167:16, 2022. 3", "type": "text" } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 105, 488, 505, 504 ], "spans": [ { "bbox": [ 105, 488, 505, 504 ], "score": 1.0, "content": "Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool.", "type": "text" } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 115, 499, 506, 515 ], "spans": [ { "bbox": [ 115, 499, 506, 515 ], "score": 1.0, "content": "Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 511, 505, 526 ], "spans": [ { "bbox": [ 114, 511, 505, 526 ], "score": 1.0, "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11461–11471, 2022.", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 522, 126, 534 ], "spans": [ { "bbox": [ 115, 522, 126, 534 ], "score": 1.0, "content": "2", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 105, 541, 504, 555 ], "spans": [ { "bbox": [ 105, 541, 504, 555 ], "score": 1.0, "content": "Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Im-", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 114, 552, 505, 568 ], "spans": [ { "bbox": [ 114, 552, 505, 568 ], "score": 1.0, "content": "age synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073,", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 562, 171, 578 ], "spans": [ { "bbox": [ 115, 562, 171, 578 ], "score": 1.0, "content": "2021. 2, 3, 4", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 105, 583, 505, 596 ], "spans": [ { "bbox": [ 105, 583, 505, 596 ], "score": 1.0, "content": "Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 116, 595, 493, 609 ], "spans": [ { "bbox": [ 116, 595, 493, 609 ], "score": 1.0, "content": "editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 614, 506, 628 ], "spans": [ { "bbox": [ 105, 614, 506, 628 ], "score": 1.0, "content": "Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and ˜", "type": "text" } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 116, 626, 506, 639 ], "spans": [ { "bbox": [ 116, 626, 506, 639 ], "score": 1.0, "content": "Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 636, 229, 650 ], "spans": [ { "bbox": [ 115, 636, 229, 650 ], "score": 1.0, "content": "arXiv:2203.09494, 2022. 3", "type": "text" } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 105, 656, 504, 669 ], "spans": [ { "bbox": [ 105, 656, 504, 669 ], "score": 1.0, "content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew,", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 114, 667, 506, 682 ], "spans": [ { "bbox": [ 114, 667, 506, 682 ], "score": 1.0, "content": "Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 678, 408, 691 ], "spans": [ { "bbox": [ 114, 678, 408, 691 ], "score": 1.0, "content": "text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", "type": "text" } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with", "type": "text" } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 115, 708, 506, 724 ], "spans": [ { "bbox": [ 115, 708, 506, 724 ], "score": 1.0, "content": "spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 721, 315, 734 ], "spans": [ { "bbox": [ 115, 721, 315, 734 ], "score": 1.0, "content": "and pattern recognition, pp. 2337–2346, 2019. 3", "type": "text" } ], "index": 45, "is_list_end_line": true }, { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Confer-", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 114, 104, 321, 117 ], "spans": [ { "bbox": [ 114, 104, 321, 117 ], "score": 1.0, "content": "ence on Computer Vision, pp. 2085–2094, 2021. 3", "type": "text", "cross_page": true } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 106, 126, 504, 137 ], "spans": [ { "bbox": [ 106, 126, 504, 137 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text", "cross_page": true } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 115, 136, 505, 150 ], "spans": [ { "bbox": [ 115, 136, 505, 150 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 114, 145, 505, 162 ], "spans": [ { "bbox": [ 114, 145, 505, 162 ], "score": 1.0, "content": "models from natural language supervision. In International Conference on Machine Learning,", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 114, 157, 263, 172 ], "spans": [ { "bbox": [ 114, 157, 263, 172 ], "score": 1.0, "content": "pp. 8748–8763. PMLR, 2021. 1, 17", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 105, 178, 505, 192 ], "spans": [ { "bbox": [ 105, 178, 505, 192 ], "score": 1.0, "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 115, 189, 505, 203 ], "spans": [ { "bbox": [ 115, 189, 505, 203 ], "score": 1.0, "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 114, 201, 461, 213 ], "spans": [ { "bbox": [ 114, 201, 461, 213 ], "score": 1.0, "content": "transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020. 14", "type": "text", "cross_page": true } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 105, 220, 505, 234 ], "spans": [ { "bbox": [ 105, 220, 505, 234 ], "score": 1.0, "content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-", "type": "text", "cross_page": true } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 115, 232, 487, 246 ], "spans": [ { "bbox": [ 115, 232, 487, 246 ], "score": 1.0, "content": "conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 104, 250, 506, 267 ], "spans": [ { "bbox": [ 104, 250, 506, 267 ], "score": 1.0, "content": "Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 115, 263, 453, 277 ], "spans": [ { "bbox": [ 115, 263, 453, 277 ], "score": 1.0, "content": "editing of real images. ACM Transactions on Graphics (TOG), 42(1):1–13, 2022. 3", "type": "text", "cross_page": true } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 105, 284, 505, 298 ], "spans": [ { "bbox": [ 105, 284, 505, 298 ], "score": 1.0, "content": "Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High- ¨", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 114, 294, 505, 309 ], "spans": [ { "bbox": [ 114, 294, 505, 309 ], "score": 1.0, "content": "resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Con-", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 113, 305, 451, 320 ], "spans": [ { "bbox": [ 113, 305, 451, 320 ], "score": 1.0, "content": "ference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022. 1, 2", "type": "text", "cross_page": true } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 105, 326, 505, 340 ], "spans": [ { "bbox": [ 105, 326, 505, 340 ], "score": 1.0, "content": "Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman.", "type": "text", "cross_page": true } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 115, 338, 505, 352 ], "spans": [ { "bbox": [ 115, 338, 505, 352 ], "score": 1.0, "content": "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2,", "type": "text", "cross_page": true } ], "index": 18 }, { "bbox": [ 114, 348, 147, 363 ], "spans": [ { "bbox": [ 114, 348, 147, 363 ], "score": 1.0, "content": "3, 5, 6", "type": "text", "cross_page": true } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 105, 369, 506, 382 ], "spans": [ { "bbox": [ 105, 369, 506, 382 ], "score": 1.0, "content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David", "type": "text", "cross_page": true } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 114, 379, 507, 394 ], "spans": [ { "bbox": [ 114, 379, 507, 394 ], "score": 1.0, "content": "Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH", "type": "text", "cross_page": true } ], "index": 21 }, { "bbox": [ 116, 391, 321, 405 ], "spans": [ { "bbox": [ 116, 391, 321, 405 ], "score": 1.0, "content": "2022 Conference Proceedings, pp. 1–10, 2022a. 2", "type": "text", "cross_page": true } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 106, 411, 505, 425 ], "spans": [ { "bbox": [ 106, 411, 505, 425 ], "score": 1.0, "content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam-", "type": "text", "cross_page": true } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 115, 423, 505, 436 ], "spans": [ { "bbox": [ 115, 423, 505, 436 ], "score": 1.0, "content": "yar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al.", "type": "text", "cross_page": true } ], "index": 24 }, { "bbox": [ 114, 432, 506, 449 ], "spans": [ { "bbox": [ 114, 432, 506, 449 ], "score": 1.0, "content": "Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint", "type": "text", "cross_page": true } ], "index": 25 }, { "bbox": [ 114, 444, 245, 457 ], "spans": [ { "bbox": [ 114, 444, 245, 457 ], "score": 1.0, "content": "arXiv:2205.11487, 2022b. 1, 2", "type": "text", "cross_page": true } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 106, 465, 505, 477 ], "spans": [ { "bbox": [ 106, 465, 505, 477 ], "score": 1.0, "content": "Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad", "type": "text", "cross_page": true } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 115, 476, 506, 489 ], "spans": [ { "bbox": [ 115, 476, 506, 489 ], "score": 1.0, "content": "Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis", "type": "text", "cross_page": true } ], "index": 28 }, { "bbox": [ 116, 488, 263, 500 ], "spans": [ { "bbox": [ 116, 488, 263, 500 ], "score": 1.0, "content": "and Machine Intelligence, 2022c. 2", "type": "text", "cross_page": true } ], "index": 29, "is_list_end_line": true }, { "bbox": [ 104, 506, 505, 522 ], "spans": [ { "bbox": [ 104, 506, 505, 522 ], "score": 1.0, "content": "Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry", "type": "text", "cross_page": true } ], "index": 30, "is_list_start_line": true }, { "bbox": [ 116, 519, 506, 533 ], "spans": [ { "bbox": [ 116, 519, 506, 533 ], "score": 1.0, "content": "Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video", "type": "text", "cross_page": true } ], "index": 31 }, { "bbox": [ 114, 528, 331, 544 ], "spans": [ { "bbox": [ 114, 528, 331, 544 ], "score": 1.0, "content": "data. arXiv preprint arXiv:2209.14792, 2022. 1, 2, 7", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 105, 550, 506, 564 ], "spans": [ { "bbox": [ 105, 550, 506, 564 ], "score": 1.0, "content": "Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 114, 560, 506, 576 ], "spans": [ { "bbox": [ 114, 560, 506, 576 ], "score": 1.0, "content": "generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 114, 571, 449, 587 ], "spans": [ { "bbox": [ 114, 571, 449, 587 ], "score": 1.0, "content": "Conference on Computer Vision and Pattern Recognition, pp. 3626–3636, 2022. 3", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 105, 592, 505, 606 ], "spans": [ { "bbox": [ 105, 592, 505, 606 ], "score": 1.0, "content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 114, 603, 505, 618 ], "spans": [ { "bbox": [ 114, 603, 505, 618 ], "score": 1.0, "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learn-", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 114, 614, 267, 628 ], "spans": [ { "bbox": [ 114, 614, 267, 628 ], "score": 1.0, "content": "ing, pp. 2256–2265. PMLR, 2015. 2", "type": "text", "cross_page": true } ], "index": 38, "is_list_end_line": true }, { "bbox": [ 104, 635, 505, 649 ], "spans": [ { "bbox": [ 104, 635, 505, 649 ], "score": 1.0, "content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv", "type": "text", "cross_page": true } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 114, 646, 265, 659 ], "spans": [ { "bbox": [ 114, 646, 265, 659 ], "score": 1.0, "content": "preprint arXiv:2010.02502, 2020. 2", "type": "text", "cross_page": true } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 105, 667, 505, 681 ], "spans": [ { "bbox": [ 105, 667, 505, 681 ], "score": 1.0, "content": "Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for", "type": "text", "cross_page": true } ], "index": 41, "is_list_start_line": true }, { "bbox": [ 114, 676, 468, 692 ], "spans": [ { "bbox": [ 114, 676, 468, 692 ], "score": 1.0, "content": "text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 1, 3, 8", "type": "text", "cross_page": true } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time:", "type": "text", "cross_page": true } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 114, 707, 506, 725 ], "spans": [ { "bbox": [ 114, 707, 506, 725 ], "score": 1.0, "content": "Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pp. 1–9,", "type": "text", "cross_page": true } ], "index": 44 }, { "bbox": [ 116, 720, 151, 732 ], "spans": [ { "bbox": [ 116, 720, 151, 732 ], "score": 1.0, "content": "2022. 3", "type": "text", "cross_page": true } ], "index": 45, "is_list_end_line": true }, { "bbox": [ 106, 83, 504, 96 ], "spans": [ { "bbox": [ 106, 83, 504, 96 ], "score": 1.0, "content": "Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image edit-", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 93, 505, 108 ], "spans": [ { "bbox": [ 115, 93, 505, 108 ], "score": 1.0, "content": "ing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477,", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 114, 102, 172, 118 ], "spans": [ { "bbox": [ 114, 102, 172, 118 ], "score": 1.0, "content": "2022. 1, 2, 3", "type": "text", "cross_page": true } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 105, 122, 506, 138 ], "spans": [ { "bbox": [ 105, 122, 506, 138 ], "score": 1.0, "content": "Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang,", "type": "text", "cross_page": true } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 115, 135, 505, 147 ], "spans": [ { "bbox": [ 115, 135, 505, 147 ], "score": 1.0, "content": "Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 115, 145, 505, 159 ], "spans": [ { "bbox": [ 115, 145, 505, 159 ], "score": 1.0, "content": "length video generation from open domain textual description. arXiv preprint arXiv:2210.02399,", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 115, 156, 151, 169 ], "spans": [ { "bbox": [ 115, 156, 151, 169 ], "score": 1.0, "content": "2022. 3", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 104, 173, 505, 190 ], "spans": [ { "bbox": [ 104, 173, 505, 190 ], "score": 1.0, "content": "Pascal Vincent. A connection between score matching and denoising autoencoders. Neural compu-", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 115, 186, 254, 199 ], "spans": [ { "bbox": [ 115, 186, 254, 199 ], "score": 1.0, "content": "tation, 23(7):1661–1674, 2011. 2", "type": "text", "cross_page": true } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 106, 204, 506, 218 ], "spans": [ { "bbox": [ 106, 204, 506, 218 ], "score": 1.0, "content": "Yael Vinker, Eliahu Horwitz, Nir Zabari, and Yedid Hoshen. Image shape manipulation from a", "type": "text", "cross_page": true } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 115, 215, 506, 230 ], "spans": [ { "bbox": [ 115, 215, 506, 230 ], "score": 1.0, "content": "single augmented training sample. In Proceedings of the IEEE/CVF International Conference on", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 115, 226, 365, 240 ], "spans": [ { "bbox": [ 115, 226, 365, 240 ], "score": 1.0, "content": "Computer Vision (ICCV), pp. 13769–13778, October 2021. 3", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 106, 246, 504, 259 ], "spans": [ { "bbox": [ 106, 246, 504, 259 ], "score": 1.0, "content": "Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion mod-", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 115, 256, 306, 271 ], "spans": [ { "bbox": [ 115, 256, 306, 271 ], "score": 1.0, "content": "els. arXiv preprint arXiv:2211.13752, 2022. 3", "type": "text", "cross_page": true } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 106, 276, 504, 289 ], "spans": [ { "bbox": [ 106, 276, 504, 289 ], "score": 1.0, "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catan-", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 115, 286, 422, 300 ], "spans": [ { "bbox": [ 115, 286, 422, 300 ], "score": 1.0, "content": "zaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018a. 3", "type": "text", "cross_page": true } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 106, 306, 505, 319 ], "spans": [ { "bbox": [ 106, 306, 505, 319 ], "score": 1.0, "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-", "type": "text", "cross_page": true } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 115, 316, 506, 330 ], "spans": [ { "bbox": [ 115, 316, 506, 330 ], "score": 1.0, "content": "resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 115, 327, 484, 341 ], "spans": [ { "bbox": [ 115, 327, 484, 341 ], "score": 1.0, "content": "the IEEE conference on computer vision and pattern recognition, pp. 8798–8807, 2018b. 3", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 347, 505, 359 ], "spans": [ { "bbox": [ 105, 347, 505, 359 ], "score": 1.0, "content": "Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan,", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 358, 505, 370 ], "spans": [ { "bbox": [ 115, 358, 505, 370 ], "score": 1.0, "content": "Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models", "type": "text", "cross_page": true } ], "index": 20 }, { "bbox": [ 115, 368, 415, 383 ], "spans": [ { "bbox": [ 115, 368, 415, 383 ], "score": 1.0, "content": "for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 3, 8", "type": "text", "cross_page": true } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 105, 385, 507, 402 ], "spans": [ { "bbox": [ 105, 385, 507, 402 ], "score": 1.0, "content": "Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In", "type": "text", "cross_page": true } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 117, 398, 504, 411 ], "spans": [ { "bbox": [ 117, 398, 504, 411 ], "score": 1.0, "content": "Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022,", "type": "text", "cross_page": true } ], "index": 23 }, { "bbox": [ 115, 409, 336, 423 ], "spans": [ { "bbox": [ 115, 409, 336, 423 ], "score": 1.0, "content": "Proceedings, Part XV, pp. 357–374. Springer, 2022. 3", "type": "text", "cross_page": true } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 105, 427, 506, 442 ], "spans": [ { "bbox": [ 105, 427, 506, 442 ], "score": 1.0, "content": "Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A latent transformer for disentangled", "type": "text", "cross_page": true } ], "index": 25, "is_list_start_line": true }, { "bbox": [ 115, 439, 507, 453 ], "spans": [ { "bbox": [ 115, 439, 507, 453 ], "score": 1.0, "content": "face editing in images and videos. In Proceedings of the IEEE/CVF international conference on", "type": "text", "cross_page": true } ], "index": 26 }, { "bbox": [ 115, 450, 295, 463 ], "spans": [ { "bbox": [ 115, 450, 295, 463 ], "score": 1.0, "content": "computer vision, pp. 13789–13798, 2021. 3", "type": "text", "cross_page": true } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 104, 468, 506, 483 ], "spans": [ { "bbox": [ 104, 468, 506, 483 ], "score": 1.0, "content": "Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan,", "type": "text", "cross_page": true } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 115, 479, 506, 494 ], "spans": [ { "bbox": [ 115, 479, 506, 494 ], "score": 1.0, "content": "Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 116, 492, 414, 504 ], "spans": [ { "bbox": [ 116, 492, 414, 504 ], "score": 1.0, "content": "rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022a. 1", "type": "text", "cross_page": true } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 105, 509, 506, 524 ], "spans": [ { "bbox": [ 105, 509, 506, 524 ], "score": 1.0, "content": "Lijun Yu, Yong Cheng, Kihyuk Sohn, Jose Lezama, Han Zhang, Huiwen Chang, Alexander G ´", "type": "text", "cross_page": true } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 116, 521, 504, 534 ], "spans": [ { "bbox": [ 116, 521, 504, 534 ], "score": 1.0, "content": "Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video", "type": "text", "cross_page": true } ], "index": 32 }, { "bbox": [ 114, 531, 365, 545 ], "spans": [ { "bbox": [ 114, 531, 365, 545 ], "score": 1.0, "content": "transformer. arXiv preprint arXiv:2212.05199, 2022b. 1, 3, 7", "type": "text", "cross_page": true } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 106, 551, 504, 564 ], "spans": [ { "bbox": [ 106, 551, 504, 564 ], "score": 1.0, "content": "Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable", "type": "text", "cross_page": true } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 115, 561, 505, 576 ], "spans": [ { "bbox": [ 115, 561, 505, 576 ], "score": 1.0, "content": "effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on", "type": "text", "cross_page": true } ], "index": 35 }, { "bbox": [ 115, 572, 386, 587 ], "spans": [ { "bbox": [ 115, 572, 386, 587 ], "score": 1.0, "content": "computer vision and pattern recognition, pp. 586–595, 2018. 8, 17", "type": "text", "cross_page": true } ], "index": 36, "is_list_end_line": true } ], "index": 23.5, "bbox_fs": [ 104, 111, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 104, 70, 506, 735 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Confer-", "type": "text" } ], "index": 1 }, { "bbox": [ 114, 104, 321, 117 ], "spans": [ { "bbox": [ 114, 104, 321, 117 ], "score": 1.0, "content": "ence on Computer Vision, pp. 2085–2094, 2021. 3", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 126, 504, 137 ], "spans": [ { "bbox": [ 106, 126, 504, 137 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 136, 505, 150 ], "spans": [ { "bbox": [ 115, 136, 505, 150 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text" } ], "index": 4 }, { "bbox": [ 114, 145, 505, 162 ], "spans": [ { "bbox": [ 114, 145, 505, 162 ], "score": 1.0, "content": "models from natural language supervision. In International Conference on Machine Learning,", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 157, 263, 172 ], "spans": [ { "bbox": [ 114, 157, 263, 172 ], "score": 1.0, "content": "pp. 8748–8763. PMLR, 2021. 1, 17", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 178, 505, 192 ], "spans": [ { "bbox": [ 105, 178, 505, 192 ], "score": 1.0, "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 189, 505, 203 ], "spans": [ { "bbox": [ 115, 189, 505, 203 ], "score": 1.0, "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", "type": "text" } ], "index": 8 }, { "bbox": [ 114, 201, 461, 213 ], "spans": [ { "bbox": [ 114, 201, 461, 213 ], "score": 1.0, "content": "transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020. 14", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 220, 505, 234 ], "spans": [ { "bbox": [ 105, 220, 505, 234 ], "score": 1.0, "content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 232, 487, 246 ], "spans": [ { "bbox": [ 115, 232, 487, 246 ], "score": 1.0, "content": "conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 250, 506, 267 ], "spans": [ { "bbox": [ 104, 250, 506, 267 ], "score": 1.0, "content": "Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 263, 453, 277 ], "spans": [ { "bbox": [ 115, 263, 453, 277 ], "score": 1.0, "content": "editing of real images. ACM Transactions on Graphics (TOG), 42(1):1–13, 2022. 3", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 284, 505, 298 ], "spans": [ { "bbox": [ 105, 284, 505, 298 ], "score": 1.0, "content": "Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High- ¨", "type": "text" } ], "index": 14 }, { "bbox": [ 114, 294, 505, 309 ], "spans": [ { "bbox": [ 114, 294, 505, 309 ], "score": 1.0, "content": "resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Con-", "type": "text" } ], "index": 15 }, { "bbox": [ 113, 305, 451, 320 ], "spans": [ { "bbox": [ 113, 305, 451, 320 ], "score": 1.0, "content": "ference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022. 1, 2", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 326, 505, 340 ], "spans": [ { "bbox": [ 105, 326, 505, 340 ], "score": 1.0, "content": "Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman.", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 338, 505, 352 ], "spans": [ { "bbox": [ 115, 338, 505, 352 ], "score": 1.0, "content": "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2,", "type": "text" } ], "index": 18 }, { "bbox": [ 114, 348, 147, 363 ], "spans": [ { "bbox": [ 114, 348, 147, 363 ], "score": 1.0, "content": "3, 5, 6", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 369, 506, 382 ], "spans": [ { "bbox": [ 105, 369, 506, 382 ], "score": 1.0, "content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 379, 507, 394 ], "spans": [ { "bbox": [ 114, 379, 507, 394 ], "score": 1.0, "content": "Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 391, 321, 405 ], "spans": [ { "bbox": [ 116, 391, 321, 405 ], "score": 1.0, "content": "2022 Conference Proceedings, pp. 1–10, 2022a. 2", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 411, 505, 425 ], "spans": [ { "bbox": [ 106, 411, 505, 425 ], "score": 1.0, "content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 423, 505, 436 ], "spans": [ { "bbox": [ 115, 423, 505, 436 ], "score": 1.0, "content": "yar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al.", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 432, 506, 449 ], "spans": [ { "bbox": [ 114, 432, 506, 449 ], "score": 1.0, "content": "Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 444, 245, 457 ], "spans": [ { "bbox": [ 114, 444, 245, 457 ], "score": 1.0, "content": "arXiv:2205.11487, 2022b. 1, 2", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 465, 505, 477 ], "spans": [ { "bbox": [ 106, 465, 505, 477 ], "score": 1.0, "content": "Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 476, 506, 489 ], "spans": [ { "bbox": [ 115, 476, 506, 489 ], "score": 1.0, "content": "Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 488, 263, 500 ], "spans": [ { "bbox": [ 116, 488, 263, 500 ], "score": 1.0, "content": "and Machine Intelligence, 2022c. 2", "type": "text" } ], "index": 29 }, { "bbox": [ 104, 506, 505, 522 ], "spans": [ { "bbox": [ 104, 506, 505, 522 ], "score": 1.0, "content": "Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 519, 506, 533 ], "spans": [ { "bbox": [ 116, 519, 506, 533 ], "score": 1.0, "content": "Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video", "type": "text" } ], "index": 31 }, { "bbox": [ 114, 528, 331, 544 ], "spans": [ { "bbox": [ 114, 528, 331, 544 ], "score": 1.0, "content": "data. arXiv preprint arXiv:2209.14792, 2022. 1, 2, 7", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 550, 506, 564 ], "spans": [ { "bbox": [ 105, 550, 506, 564 ], "score": 1.0, "content": "Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 560, 506, 576 ], "spans": [ { "bbox": [ 114, 560, 506, 576 ], "score": 1.0, "content": "generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF", "type": "text" } ], "index": 34 }, { "bbox": [ 114, 571, 449, 587 ], "spans": [ { "bbox": [ 114, 571, 449, 587 ], "score": 1.0, "content": "Conference on Computer Vision and Pattern Recognition, pp. 3626–3636, 2022. 3", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 592, 505, 606 ], "spans": [ { "bbox": [ 105, 592, 505, 606 ], "score": 1.0, "content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", "type": "text" } ], "index": 36 }, { "bbox": [ 114, 603, 505, 618 ], "spans": [ { "bbox": [ 114, 603, 505, 618 ], "score": 1.0, "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learn-", "type": "text" } ], "index": 37 }, { "bbox": [ 114, 614, 267, 628 ], "spans": [ { "bbox": [ 114, 614, 267, 628 ], "score": 1.0, "content": "ing, pp. 2256–2265. PMLR, 2015. 2", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 635, 505, 649 ], "spans": [ { "bbox": [ 104, 635, 505, 649 ], "score": 1.0, "content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 646, 265, 659 ], "spans": [ { "bbox": [ 114, 646, 265, 659 ], "score": 1.0, "content": "preprint arXiv:2010.02502, 2020. 2", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 667, 505, 681 ], "spans": [ { "bbox": [ 105, 667, 505, 681 ], "score": 1.0, "content": "Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 676, 468, 692 ], "spans": [ { "bbox": [ 114, 676, 468, 692 ], "score": 1.0, "content": "text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 1, 3, 8", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time:", "type": "text" } ], "index": 43 }, { "bbox": [ 114, 707, 506, 725 ], "spans": [ { "bbox": [ 114, 707, 506, 725 ], "score": 1.0, "content": "Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pp. 1–9,", "type": "text" } ], "index": 44 }, { "bbox": [ 116, 720, 151, 732 ], "spans": [ { "bbox": [ 116, 720, 151, 732 ], "score": 1.0, "content": "2022. 3", "type": "text" } ], "index": 45 } ], "index": 22.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 104, 70, 506, 735 ], "lines": [], "index": 22.5, "bbox_fs": [ 104, 82, 507, 732 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 102, 70, 506, 588 ], "lines": [ { "bbox": [ 106, 83, 504, 96 ], "spans": [ { "bbox": [ 106, 83, 504, 96 ], "score": 1.0, "content": "Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image edit-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 505, 108 ], "spans": [ { "bbox": [ 115, 93, 505, 108 ], "score": 1.0, "content": "ing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477,", "type": "text" } ], "index": 1 }, { "bbox": [ 114, 102, 172, 118 ], "spans": [ { "bbox": [ 114, 102, 172, 118 ], "score": 1.0, "content": "2022. 1, 2, 3", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 122, 506, 138 ], "spans": [ { "bbox": [ 105, 122, 506, 138 ], "score": 1.0, "content": "Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang,", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 135, 505, 147 ], "spans": [ { "bbox": [ 115, 135, 505, 147 ], "score": 1.0, "content": "Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 145, 505, 159 ], "spans": [ { "bbox": [ 115, 145, 505, 159 ], "score": 1.0, "content": "length video generation from open domain textual description. arXiv preprint arXiv:2210.02399,", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 156, 151, 169 ], "spans": [ { "bbox": [ 115, 156, 151, 169 ], "score": 1.0, "content": "2022. 3", "type": "text" } ], "index": 6 }, { "bbox": [ 104, 173, 505, 190 ], "spans": [ { "bbox": [ 104, 173, 505, 190 ], "score": 1.0, "content": "Pascal Vincent. A connection between score matching and denoising autoencoders. Neural compu-", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 186, 254, 199 ], "spans": [ { "bbox": [ 115, 186, 254, 199 ], "score": 1.0, "content": "tation, 23(7):1661–1674, 2011. 2", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 204, 506, 218 ], "spans": [ { "bbox": [ 106, 204, 506, 218 ], "score": 1.0, "content": "Yael Vinker, Eliahu Horwitz, Nir Zabari, and Yedid Hoshen. Image shape manipulation from a", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 215, 506, 230 ], "spans": [ { "bbox": [ 115, 215, 506, 230 ], "score": 1.0, "content": "single augmented training sample. In Proceedings of the IEEE/CVF International Conference on", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 226, 365, 240 ], "spans": [ { "bbox": [ 115, 226, 365, 240 ], "score": 1.0, "content": "Computer Vision (ICCV), pp. 13769–13778, October 2021. 3", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 246, 504, 259 ], "spans": [ { "bbox": [ 106, 246, 504, 259 ], "score": 1.0, "content": "Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion mod-", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 256, 306, 271 ], "spans": [ { "bbox": [ 115, 256, 306, 271 ], "score": 1.0, "content": "els. arXiv preprint arXiv:2211.13752, 2022. 3", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 276, 504, 289 ], "spans": [ { "bbox": [ 106, 276, 504, 289 ], "score": 1.0, "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catan-", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 286, 422, 300 ], "spans": [ { "bbox": [ 115, 286, 422, 300 ], "score": 1.0, "content": "zaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018a. 3", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 306, 505, 319 ], "spans": [ { "bbox": [ 106, 306, 505, 319 ], "score": 1.0, "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 316, 506, 330 ], "spans": [ { "bbox": [ 115, 316, 506, 330 ], "score": 1.0, "content": "resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 327, 484, 341 ], "spans": [ { "bbox": [ 115, 327, 484, 341 ], "score": 1.0, "content": "the IEEE conference on computer vision and pattern recognition, pp. 8798–8807, 2018b. 3", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 347, 505, 359 ], "spans": [ { "bbox": [ 105, 347, 505, 359 ], "score": 1.0, "content": "Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 358, 505, 370 ], "spans": [ { "bbox": [ 115, 358, 505, 370 ], "score": 1.0, "content": "Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 368, 415, 383 ], "spans": [ { "bbox": [ 115, 368, 415, 383 ], "score": 1.0, "content": "for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 3, 8", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 385, 507, 402 ], "spans": [ { "bbox": [ 105, 385, 507, 402 ], "score": 1.0, "content": "Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In", "type": "text" } ], "index": 22 }, { "bbox": [ 117, 398, 504, 411 ], "spans": [ { "bbox": [ 117, 398, 504, 411 ], "score": 1.0, "content": "Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022,", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 409, 336, 423 ], "spans": [ { "bbox": [ 115, 409, 336, 423 ], "score": 1.0, "content": "Proceedings, Part XV, pp. 357–374. Springer, 2022. 3", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 427, 506, 442 ], "spans": [ { "bbox": [ 105, 427, 506, 442 ], "score": 1.0, "content": "Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A latent transformer for disentangled", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 439, 507, 453 ], "spans": [ { "bbox": [ 115, 439, 507, 453 ], "score": 1.0, "content": "face editing in images and videos. In Proceedings of the IEEE/CVF international conference on", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 450, 295, 463 ], "spans": [ { "bbox": [ 115, 450, 295, 463 ], "score": 1.0, "content": "computer vision, pp. 13789–13798, 2021. 3", "type": "text" } ], "index": 27 }, { "bbox": [ 104, 468, 506, 483 ], "spans": [ { "bbox": [ 104, 468, 506, 483 ], "score": 1.0, "content": "Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan,", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 479, 506, 494 ], "spans": [ { "bbox": [ 115, 479, 506, 494 ], "score": 1.0, "content": "Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 492, 414, 504 ], "spans": [ { "bbox": [ 116, 492, 414, 504 ], "score": 1.0, "content": "rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022a. 1", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 509, 506, 524 ], "spans": [ { "bbox": [ 105, 509, 506, 524 ], "score": 1.0, "content": "Lijun Yu, Yong Cheng, Kihyuk Sohn, Jose Lezama, Han Zhang, Huiwen Chang, Alexander G ´", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 521, 504, 534 ], "spans": [ { "bbox": [ 116, 521, 504, 534 ], "score": 1.0, "content": "Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video", "type": "text" } ], "index": 32 }, { "bbox": [ 114, 531, 365, 545 ], "spans": [ { "bbox": [ 114, 531, 365, 545 ], "score": 1.0, "content": "transformer. arXiv preprint arXiv:2212.05199, 2022b. 1, 3, 7", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 551, 504, 564 ], "spans": [ { "bbox": [ 106, 551, 504, 564 ], "score": 1.0, "content": "Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 561, 505, 576 ], "spans": [ { "bbox": [ 115, 561, 505, 576 ], "score": 1.0, "content": "effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 572, 386, 587 ], "spans": [ { "bbox": [ 115, 572, 386, 587 ], "score": 1.0, "content": "computer vision and pattern recognition, pp. 586–595, 2018. 8, 17", "type": "text" } ], "index": 36 } ], "index": 18 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 102, 70, 506, 588 ], "lines": [], "index": 18, "bbox_fs": [ 104, 83, 507, 587 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 228, 94 ], "lines": [ { "bbox": [ 106, 79, 229, 96 ], "spans": [ { "bbox": [ 106, 79, 229, 96 ], "score": 1.0, "content": "A ATTACHED VIDEOS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 106, 105, 505, 128 ], "lines": [ { "bbox": [ 106, 106, 505, 118 ], "spans": [ { "bbox": [ 106, 106, 505, 118 ], "score": 1.0, "content": "In addition to this appendix, we include a number of videos, we highly encourage the reviewer to", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 116, 257, 129 ], "spans": [ { "bbox": [ 106, 116, 257, 129 ], "score": 1.0, "content": "view them. The included videos are:", "type": "text" } ], "index": 2 } ], "index": 1.5 }, { "type": "text", "bbox": [ 129, 136, 505, 249 ], "lines": [ { "bbox": [ 129, 136, 504, 149 ], "spans": [ { "bbox": [ 129, 136, 504, 149 ], "score": 1.0, "content": "1. “1 dreamix overview video.mp4“ - An overview video of our method with audio narration.", "type": "text" } ], "index": 3 }, { "bbox": [ 128, 150, 505, 164 ], "spans": [ { "bbox": [ 128, 150, 505, 164 ], "score": 1.0, "content": "2. “2 dreamix video editing examples.mp4“ - A number of video editing examples generated", "type": "text" } ], "index": 4 }, { "bbox": [ 142, 163, 249, 174 ], "spans": [ { "bbox": [ 142, 163, 249, 174 ], "score": 1.0, "content": "by our method (Dreamix).", "type": "text" } ], "index": 5 }, { "bbox": [ 129, 176, 504, 189 ], "spans": [ { "bbox": [ 129, 176, 504, 189 ], "score": 1.0, "content": "3. “3 dreamix image2video examples.mp4“ - A number of image-to-video examples gener-", "type": "text" } ], "index": 6 }, { "bbox": [ 141, 188, 268, 200 ], "spans": [ { "bbox": [ 141, 188, 268, 200 ], "score": 1.0, "content": "ated by our method (Dreamix).", "type": "text" } ], "index": 7 }, { "bbox": [ 128, 201, 505, 215 ], "spans": [ { "bbox": [ 128, 201, 505, 215 ], "score": 1.0, "content": "4. “4 dreamix subject driven video generation examples.mp4“ - A number of subject-driven", "type": "text" } ], "index": 8 }, { "bbox": [ 142, 214, 400, 226 ], "spans": [ { "bbox": [ 142, 214, 400, 226 ], "score": 1.0, "content": "video generation examples generated by our method (Dreamix).", "type": "text" } ], "index": 9 }, { "bbox": [ 128, 226, 506, 241 ], "spans": [ { "bbox": [ 128, 226, 506, 241 ], "score": 1.0, "content": "5. “5 dreamix baseline comparisons.mp4“ - A number of videos comparing our method", "type": "text" } ], "index": 10 }, { "bbox": [ 142, 239, 276, 250 ], "spans": [ { "bbox": [ 142, 239, 276, 250 ], "score": 1.0, "content": "(Dreamix) to the other baselines.", "type": "text" } ], "index": 11 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 258, 504, 281 ], "lines": [ { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "Note: to match the conference requirement of maximum 100MB for the supplementary, all", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 269, 492, 282 ], "spans": [ { "bbox": [ 105, 269, 492, 282 ], "score": 1.0, "content": "the videos are compressed, the uncompressed versions will be released in the final revision.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "title", "bbox": [ 108, 297, 210, 309 ], "lines": [ { "bbox": [ 105, 295, 212, 311 ], "spans": [ { "bbox": [ 105, 295, 212, 311 ], "score": 1.0, "content": "B SOCIAL IMPACT", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 321, 505, 420 ], "lines": [ { "bbox": [ 106, 322, 504, 333 ], "spans": [ { "bbox": [ 106, 322, 504, 333 ], "score": 1.0, "content": "Our primary aim in this work is to advance research on tools to enable users to animate their personal", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 330, 505, 347 ], "spans": [ { "bbox": [ 104, 330, 505, 347 ], "score": 1.0, "content": "content. While the development of end-user applications is out of the scope of this work, we recog-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 343, 505, 356 ], "spans": [ { "bbox": [ 105, 343, 505, 356 ], "score": 1.0, "content": "nize both the opportunities and risks that may follow from our contributions. As discussed above,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 354, 505, 367 ], "spans": [ { "bbox": [ 105, 354, 505, 367 ], "score": 1.0, "content": "we anticipate multiple possible applications for this work that have the potential to augment and", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 365, 505, 378 ], "spans": [ { "bbox": [ 105, 365, 505, 378 ], "score": 1.0, "content": "extend creative practices. The personalized component of our approach brings particular promise", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 376, 504, 388 ], "spans": [ { "bbox": [ 106, 376, 504, 388 ], "score": 1.0, "content": "as it will enable users to better align content with their intent, despite potential biases present in", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 386, 506, 400 ], "spans": [ { "bbox": [ 104, 386, 506, 400 ], "score": 1.0, "content": "general VDMs. On the other hand, our method carries similar risks as other highly capable media", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 398, 506, 411 ], "spans": [ { "bbox": [ 105, 398, 506, 411 ], "score": 1.0, "content": "generation approaches. Malicious parties may try to use edited videos to mis-lead viewers or to", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 409, 472, 423 ], "spans": [ { "bbox": [ 106, 409, 472, 423 ], "score": 1.0, "content": "engage in targeted harassment. Future research must continue investigating these concerns.", "type": "text" } ], "index": 23 } ], "index": 19 }, { "type": "title", "bbox": [ 108, 436, 268, 448 ], "lines": [ { "bbox": [ 105, 435, 270, 451 ], "spans": [ { "bbox": [ 105, 435, 270, 451 ], "score": 1.0, "content": "C IMPLEMENTATION DETAILS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 108, 460, 202, 471 ], "lines": [ { "bbox": [ 106, 460, 203, 473 ], "spans": [ { "bbox": [ 106, 460, 203, 473 ], "score": 1.0, "content": "C.1 ARCHITECTURE", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 480, 504, 504 ], "lines": [ { "bbox": [ 106, 480, 504, 493 ], "spans": [ { "bbox": [ 106, 480, 504, 493 ], "score": 1.0, "content": "All of our experiments were preformed on a VDM that is similar to Imagen-Video Ho et al. (2022a),", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 492, 446, 505 ], "spans": [ { "bbox": [ 105, 492, 446, 505 ], "score": 1.0, "content": "a pertrained cascaded text-to-video diffusion model, with the following components:", "type": "text" } ], "index": 27 } ], "index": 26.5 }, { "type": "text", "bbox": [ 129, 512, 505, 629 ], "lines": [ { "bbox": [ 130, 512, 505, 524 ], "spans": [ { "bbox": [ 130, 512, 505, 524 ], "score": 1.0, "content": "1. A T5-XXLRaffel et al. (2020) text encoder, that computes embeddings from the textual", "type": "text" } ], "index": 28 }, { "bbox": [ 141, 523, 447, 536 ], "spans": [ { "bbox": [ 141, 523, 447, 536 ], "score": 1.0, "content": "prompt. This embeddings are then used as conditioning by all other models.", "type": "text" } ], "index": 29 }, { "bbox": [ 128, 536, 504, 550 ], "spans": [ { "bbox": [ 128, 536, 433, 550 ], "score": 1.0, "content": "2. A base video diffusion model, conditioned on text. It generates videos at", "type": "text" }, { "bbox": [ 434, 537, 504, 548 ], "score": 0.9, "content": "1 6 \\times 2 4 \\times 4 0 \\times 3", "type": "inline_equation" } ], "index": 30 }, { "bbox": [ 141, 549, 379, 561 ], "spans": [ { "bbox": [ 141, 549, 217, 561 ], "score": 1.0, "content": "resolution (frames", "type": "text" }, { "bbox": [ 218, 549, 228, 559 ], "score": 0.8, "content": "X", "type": "inline_equation" }, { "bbox": [ 228, 549, 256, 561 ], "score": 1.0, "content": "height", "type": "text" }, { "bbox": [ 257, 549, 267, 559 ], "score": 0.8, "content": "X", "type": "inline_equation" }, { "bbox": [ 267, 549, 293, 561 ], "score": 1.0, "content": "width", "type": "text" }, { "bbox": [ 293, 549, 304, 559 ], "score": 0.79, "content": "X", "type": "inline_equation" }, { "bbox": [ 304, 549, 379, 561 ], "score": 1.0, "content": "channels) at 3 fps.", "type": "text" } ], "index": 31 }, { "bbox": [ 128, 562, 506, 576 ], "spans": [ { "bbox": [ 128, 562, 506, 576 ], "score": 1.0, "content": "3. 6 super-resolution video diffusion models, each conditioned on the text and the output", "type": "text" } ], "index": 32 }, { "bbox": [ 142, 574, 505, 586 ], "spans": [ { "bbox": [ 142, 574, 505, 586 ], "score": 1.0, "content": "video of the previous model. Each model is either spatial (SSR), i.e. upscales resolution,", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 585, 506, 597 ], "spans": [ { "bbox": [ 141, 585, 506, 597 ], "score": 1.0, "content": "or temporal (TSR), i.e. fills in intermediate frames between the input frames. The order of", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 596, 505, 608 ], "spans": [ { "bbox": [ 141, 596, 268, 608 ], "score": 1.0, "content": "super resolution models is TSR", "type": "text" }, { "bbox": [ 269, 596, 286, 607 ], "score": 0.4, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 286, 596, 310, 608 ], "score": 1.0, "content": ", SSR", "type": "text" }, { "bbox": [ 311, 596, 327, 607 ], "score": 0.35, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 327, 596, 390, 608 ], "score": 1.0, "content": ", SSR(4x), TSR", "type": "text" }, { "bbox": [ 390, 596, 406, 607 ], "score": 0.3, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 406, 596, 429, 608 ], "score": 1.0, "content": ", TSR", "type": "text" }, { "bbox": [ 429, 596, 445, 607 ], "score": 0.27, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 446, 596, 505, 608 ], "score": 1.0, "content": ", and SSR(4x).", "type": "text" } ], "index": 35 }, { "bbox": [ 142, 606, 505, 620 ], "spans": [ { "bbox": [ 142, 606, 505, 620 ], "score": 1.0, "content": "The multiplier in the parenthesis for output frames (for TSR), and for output pixels in height", "type": "text" } ], "index": 36 }, { "bbox": [ 142, 618, 471, 630 ], "spans": [ { "bbox": [ 142, 618, 338, 630 ], "score": 1.0, "content": "and width (for SSR). The final output video is in", "type": "text" }, { "bbox": [ 338, 618, 430, 628 ], "score": 0.91, "content": "1 2 8 \\times 7 6 8 \\times 1 2 8 0 \\times 3", "type": "inline_equation" }, { "bbox": [ 431, 618, 471, 630 ], "score": 1.0, "content": "at 24 fps.", "type": "text" } ], "index": 37 } ], "index": 32.5 }, { "type": "text", "bbox": [ 109, 638, 503, 671 ], "lines": [ { "bbox": [ 106, 637, 505, 651 ], "spans": [ { "bbox": [ 106, 637, 505, 651 ], "score": 1.0, "content": "Note that the diffusion models are pretrained on both videos and images, with frozen temporal", "type": "text" } ], "index": 38 }, { "bbox": [ 107, 649, 504, 660 ], "spans": [ { "bbox": [ 107, 649, 504, 660 ], "score": 1.0, "content": "attention and convolution for the latter. Our mixed finetuning approach treats video frames as if they", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 660, 162, 674 ], "spans": [ { "bbox": [ 106, 660, 162, 674 ], "score": 1.0, "content": "were images.", "type": "text" } ], "index": 40 } ], "index": 39 }, { "type": "text", "bbox": [ 108, 677, 504, 732 ], "lines": [ { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "Distillation. For some of these models, we use a distilled version to allow for faster sampling times.", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 687, 505, 699 ], "spans": [ { "bbox": [ 106, 687, 505, 699 ], "score": 1.0, "content": "The base model is a distilled model with 64 sampling steps. The first two SSR models are non-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "distilled models with 128 sampling steps (due to finetuning considerations, see below). All other SR", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "models use 8 sampling steps. All models use classifier-free-guidance weight of 1.0 (meaning that", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 721, 257, 733 ], "spans": [ { "bbox": [ 106, 721, 257, 733 ], "score": 1.0, "content": "classifier free guidance is turned off).", "type": "text" } ], "index": 45 } ], "index": 43 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 228, 94 ], "lines": [ { "bbox": [ 106, 79, 229, 96 ], "spans": [ { "bbox": [ 106, 79, 229, 96 ], "score": 1.0, "content": "A ATTACHED VIDEOS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 106, 105, 505, 128 ], "lines": [ { "bbox": [ 106, 106, 505, 118 ], "spans": [ { "bbox": [ 106, 106, 505, 118 ], "score": 1.0, "content": "In addition to this appendix, we include a number of videos, we highly encourage the reviewer to", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 116, 257, 129 ], "spans": [ { "bbox": [ 106, 116, 257, 129 ], "score": 1.0, "content": "view them. The included videos are:", "type": "text" } ], "index": 2 } ], "index": 1.5, "bbox_fs": [ 106, 106, 505, 129 ] }, { "type": "list", "bbox": [ 129, 136, 505, 249 ], "lines": [ { "bbox": [ 129, 136, 504, 149 ], "spans": [ { "bbox": [ 129, 136, 504, 149 ], "score": 1.0, "content": "1. “1 dreamix overview video.mp4“ - An overview video of our method with audio narration.", "type": "text" } ], "index": 3, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 128, 150, 505, 164 ], "spans": [ { "bbox": [ 128, 150, 505, 164 ], "score": 1.0, "content": "2. “2 dreamix video editing examples.mp4“ - A number of video editing examples generated", "type": "text" } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 142, 163, 249, 174 ], "spans": [ { "bbox": [ 142, 163, 249, 174 ], "score": 1.0, "content": "by our method (Dreamix).", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 129, 176, 504, 189 ], "spans": [ { "bbox": [ 129, 176, 504, 189 ], "score": 1.0, "content": "3. “3 dreamix image2video examples.mp4“ - A number of image-to-video examples gener-", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 141, 188, 268, 200 ], "spans": [ { "bbox": [ 141, 188, 268, 200 ], "score": 1.0, "content": "ated by our method (Dreamix).", "type": "text" } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 128, 201, 505, 215 ], "spans": [ { "bbox": [ 128, 201, 505, 215 ], "score": 1.0, "content": "4. “4 dreamix subject driven video generation examples.mp4“ - A number of subject-driven", "type": "text" } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 142, 214, 400, 226 ], "spans": [ { "bbox": [ 142, 214, 400, 226 ], "score": 1.0, "content": "video generation examples generated by our method (Dreamix).", "type": "text" } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 128, 226, 506, 241 ], "spans": [ { "bbox": [ 128, 226, 506, 241 ], "score": 1.0, "content": "5. “5 dreamix baseline comparisons.mp4“ - A number of videos comparing our method", "type": "text" } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 142, 239, 276, 250 ], "spans": [ { "bbox": [ 142, 239, 276, 250 ], "score": 1.0, "content": "(Dreamix) to the other baselines.", "type": "text" } ], "index": 11, "is_list_end_line": true } ], "index": 7, "bbox_fs": [ 128, 136, 506, 250 ] }, { "type": "text", "bbox": [ 107, 258, 504, 281 ], "lines": [ { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "Note: to match the conference requirement of maximum 100MB for the supplementary, all", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 269, 492, 282 ], "spans": [ { "bbox": [ 105, 269, 492, 282 ], "score": 1.0, "content": "the videos are compressed, the uncompressed versions will be released in the final revision.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 105, 258, 505, 282 ] }, { "type": "title", "bbox": [ 108, 297, 210, 309 ], "lines": [ { "bbox": [ 105, 295, 212, 311 ], "spans": [ { "bbox": [ 105, 295, 212, 311 ], "score": 1.0, "content": "B SOCIAL IMPACT", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 321, 505, 420 ], "lines": [ { "bbox": [ 106, 322, 504, 333 ], "spans": [ { "bbox": [ 106, 322, 504, 333 ], "score": 1.0, "content": "Our primary aim in this work is to advance research on tools to enable users to animate their personal", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 330, 505, 347 ], "spans": [ { "bbox": [ 104, 330, 505, 347 ], "score": 1.0, "content": "content. While the development of end-user applications is out of the scope of this work, we recog-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 343, 505, 356 ], "spans": [ { "bbox": [ 105, 343, 505, 356 ], "score": 1.0, "content": "nize both the opportunities and risks that may follow from our contributions. As discussed above,", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 354, 505, 367 ], "spans": [ { "bbox": [ 105, 354, 505, 367 ], "score": 1.0, "content": "we anticipate multiple possible applications for this work that have the potential to augment and", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 365, 505, 378 ], "spans": [ { "bbox": [ 105, 365, 505, 378 ], "score": 1.0, "content": "extend creative practices. The personalized component of our approach brings particular promise", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 376, 504, 388 ], "spans": [ { "bbox": [ 106, 376, 504, 388 ], "score": 1.0, "content": "as it will enable users to better align content with their intent, despite potential biases present in", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 386, 506, 400 ], "spans": [ { "bbox": [ 104, 386, 506, 400 ], "score": 1.0, "content": "general VDMs. On the other hand, our method carries similar risks as other highly capable media", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 398, 506, 411 ], "spans": [ { "bbox": [ 105, 398, 506, 411 ], "score": 1.0, "content": "generation approaches. Malicious parties may try to use edited videos to mis-lead viewers or to", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 409, 472, 423 ], "spans": [ { "bbox": [ 106, 409, 472, 423 ], "score": 1.0, "content": "engage in targeted harassment. Future research must continue investigating these concerns.", "type": "text" } ], "index": 23 } ], "index": 19, "bbox_fs": [ 104, 322, 506, 423 ] }, { "type": "title", "bbox": [ 108, 436, 268, 448 ], "lines": [ { "bbox": [ 105, 435, 270, 451 ], "spans": [ { "bbox": [ 105, 435, 270, 451 ], "score": 1.0, "content": "C IMPLEMENTATION DETAILS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 108, 460, 202, 471 ], "lines": [ { "bbox": [ 106, 460, 203, 473 ], "spans": [ { "bbox": [ 106, 460, 203, 473 ], "score": 1.0, "content": "C.1 ARCHITECTURE", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 480, 504, 504 ], "lines": [ { "bbox": [ 106, 480, 504, 493 ], "spans": [ { "bbox": [ 106, 480, 504, 493 ], "score": 1.0, "content": "All of our experiments were preformed on a VDM that is similar to Imagen-Video Ho et al. (2022a),", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 492, 446, 505 ], "spans": [ { "bbox": [ 105, 492, 446, 505 ], "score": 1.0, "content": "a pertrained cascaded text-to-video diffusion model, with the following components:", "type": "text" } ], "index": 27 } ], "index": 26.5, "bbox_fs": [ 105, 480, 504, 505 ] }, { "type": "list", "bbox": [ 129, 512, 505, 629 ], "lines": [ { "bbox": [ 130, 512, 505, 524 ], "spans": [ { "bbox": [ 130, 512, 505, 524 ], "score": 1.0, "content": "1. A T5-XXLRaffel et al. (2020) text encoder, that computes embeddings from the textual", "type": "text" } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 141, 523, 447, 536 ], "spans": [ { "bbox": [ 141, 523, 447, 536 ], "score": 1.0, "content": "prompt. This embeddings are then used as conditioning by all other models.", "type": "text" } ], "index": 29, "is_list_end_line": true }, { "bbox": [ 128, 536, 504, 550 ], "spans": [ { "bbox": [ 128, 536, 433, 550 ], "score": 1.0, "content": "2. A base video diffusion model, conditioned on text. It generates videos at", "type": "text" }, { "bbox": [ 434, 537, 504, 548 ], "score": 0.9, "content": "1 6 \\times 2 4 \\times 4 0 \\times 3", "type": "inline_equation" } ], "index": 30, "is_list_start_line": true }, { "bbox": [ 141, 549, 379, 561 ], "spans": [ { "bbox": [ 141, 549, 217, 561 ], "score": 1.0, "content": "resolution (frames", "type": "text" }, { "bbox": [ 218, 549, 228, 559 ], "score": 0.8, "content": "X", "type": "inline_equation" }, { "bbox": [ 228, 549, 256, 561 ], "score": 1.0, "content": "height", "type": "text" }, { "bbox": [ 257, 549, 267, 559 ], "score": 0.8, "content": "X", "type": "inline_equation" }, { "bbox": [ 267, 549, 293, 561 ], "score": 1.0, "content": "width", "type": "text" }, { "bbox": [ 293, 549, 304, 559 ], "score": 0.79, "content": "X", "type": "inline_equation" }, { "bbox": [ 304, 549, 379, 561 ], "score": 1.0, "content": "channels) at 3 fps.", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 128, 562, 506, 576 ], "spans": [ { "bbox": [ 128, 562, 506, 576 ], "score": 1.0, "content": "3. 6 super-resolution video diffusion models, each conditioned on the text and the output", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 142, 574, 505, 586 ], "spans": [ { "bbox": [ 142, 574, 505, 586 ], "score": 1.0, "content": "video of the previous model. Each model is either spatial (SSR), i.e. upscales resolution,", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 585, 506, 597 ], "spans": [ { "bbox": [ 141, 585, 506, 597 ], "score": 1.0, "content": "or temporal (TSR), i.e. fills in intermediate frames between the input frames. The order of", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 596, 505, 608 ], "spans": [ { "bbox": [ 141, 596, 268, 608 ], "score": 1.0, "content": "super resolution models is TSR", "type": "text" }, { "bbox": [ 269, 596, 286, 607 ], "score": 0.4, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 286, 596, 310, 608 ], "score": 1.0, "content": ", SSR", "type": "text" }, { "bbox": [ 311, 596, 327, 607 ], "score": 0.35, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 327, 596, 390, 608 ], "score": 1.0, "content": ", SSR(4x), TSR", "type": "text" }, { "bbox": [ 390, 596, 406, 607 ], "score": 0.3, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 406, 596, 429, 608 ], "score": 1.0, "content": ", TSR", "type": "text" }, { "bbox": [ 429, 596, 445, 607 ], "score": 0.27, "content": "( 2 \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 446, 596, 505, 608 ], "score": 1.0, "content": ", and SSR(4x).", "type": "text" } ], "index": 35 }, { "bbox": [ 142, 606, 505, 620 ], "spans": [ { "bbox": [ 142, 606, 505, 620 ], "score": 1.0, "content": "The multiplier in the parenthesis for output frames (for TSR), and for output pixels in height", "type": "text" } ], "index": 36 }, { "bbox": [ 142, 618, 471, 630 ], "spans": [ { "bbox": [ 142, 618, 338, 630 ], "score": 1.0, "content": "and width (for SSR). The final output video is in", "type": "text" }, { "bbox": [ 338, 618, 430, 628 ], "score": 0.91, "content": "1 2 8 \\times 7 6 8 \\times 1 2 8 0 \\times 3", "type": "inline_equation" }, { "bbox": [ 431, 618, 471, 630 ], "score": 1.0, "content": "at 24 fps.", "type": "text" } ], "index": 37, "is_list_end_line": true } ], "index": 32.5, "bbox_fs": [ 128, 512, 506, 630 ] }, { "type": "text", "bbox": [ 109, 638, 503, 671 ], "lines": [ { "bbox": [ 106, 637, 505, 651 ], "spans": [ { "bbox": [ 106, 637, 505, 651 ], "score": 1.0, "content": "Note that the diffusion models are pretrained on both videos and images, with frozen temporal", "type": "text" } ], "index": 38 }, { "bbox": [ 107, 649, 504, 660 ], "spans": [ { "bbox": [ 107, 649, 504, 660 ], "score": 1.0, "content": "attention and convolution for the latter. Our mixed finetuning approach treats video frames as if they", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 660, 162, 674 ], "spans": [ { "bbox": [ 106, 660, 162, 674 ], "score": 1.0, "content": "were images.", "type": "text" } ], "index": 40 } ], "index": 39, "bbox_fs": [ 106, 637, 505, 674 ] }, { "type": "text", "bbox": [ 108, 677, 504, 732 ], "lines": [ { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "Distillation. For some of these models, we use a distilled version to allow for faster sampling times.", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 687, 505, 699 ], "spans": [ { "bbox": [ 106, 687, 505, 699 ], "score": 1.0, "content": "The base model is a distilled model with 64 sampling steps. The first two SSR models are non-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "distilled models with 128 sampling steps (due to finetuning considerations, see below). All other SR", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "models use 8 sampling steps. All models use classifier-free-guidance weight of 1.0 (meaning that", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 721, 257, 733 ], "spans": [ { "bbox": [ 106, 721, 257, 733 ], "score": 1.0, "content": "classifier free guidance is turned off).", "type": "text" } ], "index": 45 } ], "index": 43, "bbox_fs": [ 105, 677, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 82, 188, 93 ], "lines": [ { "bbox": [ 106, 82, 189, 95 ], "spans": [ { "bbox": [ 106, 82, 189, 95 ], "score": 1.0, "content": "C.2 FINETUNING", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 103, 505, 159 ], "lines": [ { "bbox": [ 106, 103, 506, 116 ], "spans": [ { "bbox": [ 106, 103, 506, 116 ], "score": 1.0, "content": "To reduce finetuning time, we only finetune the base model and the first 2 SSR models. In our", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 114, 505, 127 ], "spans": [ { "bbox": [ 105, 114, 505, 127 ], "score": 1.0, "content": "experiments, finetuning the first 2 SSR models using the distilled models (with 8 sampling steps)", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 125, 506, 138 ], "spans": [ { "bbox": [ 105, 125, 506, 138 ], "score": 1.0, "content": "did not yield good quality. We therefore use the non-distilled versions of these models for all exper-", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 136, 505, 149 ], "spans": [ { "bbox": [ 106, 136, 505, 149 ], "score": 1.0, "content": "iments (including non-finetuned experiments). When using “Mixed Video-Image Finetuning“ we", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 146, 505, 161 ], "spans": [ { "bbox": [ 104, 146, 122, 161 ], "score": 1.0, "content": "use", "type": "text" }, { "bbox": [ 122, 147, 161, 158 ], "score": 0.88, "content": "\\alpha = 0 . 3 5", "type": "inline_equation" }, { "bbox": [ 161, 146, 468, 161 ], "score": 1.0, "content": ", and finetune for 300 steps. For all our experiments we use a learning rate of", "type": "text" }, { "bbox": [ 468, 147, 501, 158 ], "score": 0.9, "content": "6 \\cdot \\mathrm { { 1 0 ^ { - 6 } } }", "type": "inline_equation" }, { "bbox": [ 501, 146, 505, 161 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 5 } ], "index": 3 }, { "type": "title", "bbox": [ 107, 173, 179, 184 ], "lines": [ { "bbox": [ 106, 172, 181, 186 ], "spans": [ { "bbox": [ 106, 172, 181, 186 ], "score": 1.0, "content": "C.3 SAMPLING", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 194, 505, 238 ], "lines": [ { "bbox": [ 106, 194, 505, 205 ], "spans": [ { "bbox": [ 106, 194, 505, 205 ], "score": 1.0, "content": "We use a DDIM sampler with stochastic noise correction, following Ho et al. (2022a). For the last", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 205, 505, 217 ], "spans": [ { "bbox": [ 105, 205, 505, 217 ], "score": 1.0, "content": "highest resolution SSR, for capacity reasons, we use the model to sample a sub-chunks of 32 frames", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 216, 505, 228 ], "spans": [ { "bbox": [ 105, 216, 505, 228 ], "score": 1.0, "content": "of the input lower resolution videos, and then we concatenate all the outputs together back to 128", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 226, 164, 239 ], "spans": [ { "bbox": [ 105, 226, 164, 239 ], "score": 1.0, "content": "frame videos.", "type": "text" } ], "index": 10 } ], "index": 8.5 }, { "type": "title", "bbox": [ 107, 255, 327, 268 ], "lines": [ { "bbox": [ 105, 254, 329, 270 ], "spans": [ { "bbox": [ 105, 254, 329, 270 ], "score": 1.0, "content": "D IMAGE-TO-VIDEO TRANSFORMATIONS", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 280, 504, 336 ], "lines": [ { "bbox": [ 106, 280, 505, 294 ], "spans": [ { "bbox": [ 106, 280, 505, 294 ], "score": 1.0, "content": "We only use perspective transformations to create “cinematic” effects, e.g., panning, zooming, and", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 292, 505, 304 ], "spans": [ { "bbox": [ 106, 292, 505, 304 ], "score": 1.0, "content": "camera shake. In our supplementary, we included Image-to-Video examples with different per-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 302, 506, 316 ], "spans": [ { "bbox": [ 105, 302, 506, 316 ], "score": 1.0, "content": "spective transformations applied to them. We detail these transformations in Tab. 4. Some of the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 313, 506, 326 ], "spans": [ { "bbox": [ 105, 313, 506, 326 ], "score": 1.0, "content": "examples did not use the perspective transformations at all. Also, ensuring the smoothness of the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "transformed sequence is unnecessary as this is fixed by the diffusion and super-resolution processes.", "type": "text" } ], "index": 16 } ], "index": 14 }, { "type": "table", "bbox": [ 200, 371, 410, 503 ], "blocks": [ { "type": "table_caption", "bbox": [ 229, 347, 381, 359 ], "group_id": 0, "lines": [ { "bbox": [ 227, 346, 383, 360 ], "spans": [ { "bbox": [ 227, 346, 383, 360 ], "score": 1.0, "content": "Table 4: Perspective Transformations", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table_body", "bbox": [ 200, 371, 410, 503 ], "group_id": 0, "lines": [ { "bbox": [ 200, 371, 410, 503 ], "spans": [ { "bbox": [ 200, 371, 410, 503 ], "score": 0.98, "html": "
VideoTimestamp TransformationEffect
Plant00:00TranslatePan
Turtle00:11Rand. translateShake
Coffee00:22TranslatePan
Camel00:33NoneNone
Volcano00:43Rand. translateShake
Bear00:54PerspectivePan
Penguins01:05NoneNone
Unicorn01:15ScaleZoom out
Buffalo01:26ScaleZoom out
Bigfoot01:37TranslatePan
", "type": "table", "image_path": "6c8044d77c84b11f2c4896d127b702016082a3364a9e9d67cefda83a58f19eb9.jpg" } ] } ], "index": 22.5, "virtual_lines": [ { "bbox": [ 200, 371, 410, 384.2 ], "spans": [], "index": 18 }, { "bbox": [ 200, 384.2, 410, 397.4 ], "spans": [], "index": 19 }, { "bbox": [ 200, 397.4, 410, 410.59999999999997 ], "spans": [], "index": 20 }, { "bbox": [ 200, 410.59999999999997, 410, 423.79999999999995 ], "spans": [], "index": 21 }, { "bbox": [ 200, 423.79999999999995, 410, 436.99999999999994 ], "spans": [], "index": 22 }, { "bbox": [ 200, 436.99999999999994, 410, 450.19999999999993 ], "spans": [], "index": 23 }, { "bbox": [ 200, 450.19999999999993, 410, 463.3999999999999 ], "spans": [], "index": 24 }, { "bbox": [ 200, 463.3999999999999, 410, 476.5999999999999 ], "spans": [], "index": 25 }, { "bbox": [ 200, 476.5999999999999, 410, 489.7999999999999 ], "spans": [], "index": 26 }, { "bbox": [ 200, 489.7999999999999, 410, 502.9999999999999 ], "spans": [], "index": 27 } ] } ], "index": 19.75 }, { "type": "title", "bbox": [ 107, 531, 243, 544 ], "lines": [ { "bbox": [ 105, 531, 244, 546 ], "spans": [ { "bbox": [ 105, 531, 244, 546 ], "score": 1.0, "content": "E EVALUATION DATASET", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 556, 505, 667 ], "lines": [ { "bbox": [ 105, 556, 505, 569 ], "spans": [ { "bbox": [ 105, 556, 505, 569 ], "score": 1.0, "content": "In all evaluations described in the paper, we used a dataset of 29 videos with 127 edit prompts.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 568, 505, 579 ], "spans": [ { "bbox": [ 106, 568, 505, 579 ], "score": 1.0, "content": "The dataset videos were selected from YouTube-8M Abu-El-Haija et al. (2016) and show animals,", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 578, 506, 592 ], "spans": [ { "bbox": [ 105, 578, 506, 592 ], "score": 1.0, "content": "people performing actions, vehicles, and other objects. The edit prompt categories are motion,", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 588, 506, 604 ], "spans": [ { "bbox": [ 105, 588, 506, 604 ], "score": 1.0, "content": "object, background, and style. In the motion category the prompts perform motion editing (e.g.", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 601, 506, 614 ], "spans": [ { "bbox": [ 106, 601, 506, 614 ], "score": 1.0, "content": "adding motion with the prompt “An orangutan next to a pond waving both arms in the air”), the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 610, 506, 626 ], "spans": [ { "bbox": [ 105, 610, 506, 626 ], "score": 1.0, "content": "object category performs object level edits (e.g. adding a party hat with the prompt “A puppy", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 621, 506, 636 ], "spans": [ { "bbox": [ 105, 621, 506, 636 ], "score": 1.0, "content": "walking with a party hat”), the background category performs edits of the background (e.g. adding", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 633, 505, 646 ], "spans": [ { "bbox": [ 105, 633, 505, 646 ], "score": 1.0, "content": "a river with the prompt “A blue pickup truck crossing a deep river”), and the style category performs", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 645, 505, 658 ], "spans": [ { "bbox": [ 105, 645, 505, 658 ], "score": 1.0, "content": "style-transfer like edits (e.g. changing the style to cartoon with the prompt “A cartoon of a man", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 656, 201, 668 ], "spans": [ { "bbox": [ 105, 656, 201, 668 ], "score": 1.0, "content": "playing a saxophone”).", "type": "text" } ], "index": 38 } ], "index": 33.5 }, { "type": "title", "bbox": [ 108, 684, 282, 697 ], "lines": [ { "bbox": [ 105, 683, 283, 698 ], "spans": [ { "bbox": [ 105, 683, 283, 698 ], "score": 1.0, "content": "F HUMAN EVALUATION DETAILS", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "We performed human evaluations for the baseline comparison and the ablation analysis. The eval-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 720, 505, 732 ], "spans": [ { "bbox": [ 105, 720, 505, 732 ], "score": 1.0, "content": "uations were conducted by a panel of 20 human raters using the dataset described in ??. The video", "type": "text" } ], "index": 41 } ], "index": 40.5 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 82, 188, 93 ], "lines": [ { "bbox": [ 106, 82, 189, 95 ], "spans": [ { "bbox": [ 106, 82, 189, 95 ], "score": 1.0, "content": "C.2 FINETUNING", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 103, 505, 159 ], "lines": [ { "bbox": [ 106, 103, 506, 116 ], "spans": [ { "bbox": [ 106, 103, 506, 116 ], "score": 1.0, "content": "To reduce finetuning time, we only finetune the base model and the first 2 SSR models. In our", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 114, 505, 127 ], "spans": [ { "bbox": [ 105, 114, 505, 127 ], "score": 1.0, "content": "experiments, finetuning the first 2 SSR models using the distilled models (with 8 sampling steps)", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 125, 506, 138 ], "spans": [ { "bbox": [ 105, 125, 506, 138 ], "score": 1.0, "content": "did not yield good quality. We therefore use the non-distilled versions of these models for all exper-", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 136, 505, 149 ], "spans": [ { "bbox": [ 106, 136, 505, 149 ], "score": 1.0, "content": "iments (including non-finetuned experiments). When using “Mixed Video-Image Finetuning“ we", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 146, 505, 161 ], "spans": [ { "bbox": [ 104, 146, 122, 161 ], "score": 1.0, "content": "use", "type": "text" }, { "bbox": [ 122, 147, 161, 158 ], "score": 0.88, "content": "\\alpha = 0 . 3 5", "type": "inline_equation" }, { "bbox": [ 161, 146, 468, 161 ], "score": 1.0, "content": ", and finetune for 300 steps. For all our experiments we use a learning rate of", "type": "text" }, { "bbox": [ 468, 147, 501, 158 ], "score": 0.9, "content": "6 \\cdot \\mathrm { { 1 0 ^ { - 6 } } }", "type": "inline_equation" }, { "bbox": [ 501, 146, 505, 161 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 5 } ], "index": 3, "bbox_fs": [ 104, 103, 506, 161 ] }, { "type": "title", "bbox": [ 107, 173, 179, 184 ], "lines": [ { "bbox": [ 106, 172, 181, 186 ], "spans": [ { "bbox": [ 106, 172, 181, 186 ], "score": 1.0, "content": "C.3 SAMPLING", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 194, 505, 238 ], "lines": [ { "bbox": [ 106, 194, 505, 205 ], "spans": [ { "bbox": [ 106, 194, 505, 205 ], "score": 1.0, "content": "We use a DDIM sampler with stochastic noise correction, following Ho et al. (2022a). For the last", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 205, 505, 217 ], "spans": [ { "bbox": [ 105, 205, 505, 217 ], "score": 1.0, "content": "highest resolution SSR, for capacity reasons, we use the model to sample a sub-chunks of 32 frames", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 216, 505, 228 ], "spans": [ { "bbox": [ 105, 216, 505, 228 ], "score": 1.0, "content": "of the input lower resolution videos, and then we concatenate all the outputs together back to 128", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 226, 164, 239 ], "spans": [ { "bbox": [ 105, 226, 164, 239 ], "score": 1.0, "content": "frame videos.", "type": "text" } ], "index": 10 } ], "index": 8.5, "bbox_fs": [ 105, 194, 505, 239 ] }, { "type": "title", "bbox": [ 107, 255, 327, 268 ], "lines": [ { "bbox": [ 105, 254, 329, 270 ], "spans": [ { "bbox": [ 105, 254, 329, 270 ], "score": 1.0, "content": "D IMAGE-TO-VIDEO TRANSFORMATIONS", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 280, 504, 336 ], "lines": [ { "bbox": [ 106, 280, 505, 294 ], "spans": [ { "bbox": [ 106, 280, 505, 294 ], "score": 1.0, "content": "We only use perspective transformations to create “cinematic” effects, e.g., panning, zooming, and", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 292, 505, 304 ], "spans": [ { "bbox": [ 106, 292, 505, 304 ], "score": 1.0, "content": "camera shake. In our supplementary, we included Image-to-Video examples with different per-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 302, 506, 316 ], "spans": [ { "bbox": [ 105, 302, 506, 316 ], "score": 1.0, "content": "spective transformations applied to them. We detail these transformations in Tab. 4. Some of the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 313, 506, 326 ], "spans": [ { "bbox": [ 105, 313, 506, 326 ], "score": 1.0, "content": "examples did not use the perspective transformations at all. Also, ensuring the smoothness of the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "transformed sequence is unnecessary as this is fixed by the diffusion and super-resolution processes.", "type": "text" } ], "index": 16 } ], "index": 14, "bbox_fs": [ 105, 280, 506, 338 ] }, { "type": "table", "bbox": [ 200, 371, 410, 503 ], "blocks": [ { "type": "table_caption", "bbox": [ 229, 347, 381, 359 ], "group_id": 0, "lines": [ { "bbox": [ 227, 346, 383, 360 ], "spans": [ { "bbox": [ 227, 346, 383, 360 ], "score": 1.0, "content": "Table 4: Perspective Transformations", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table_body", "bbox": [ 200, 371, 410, 503 ], "group_id": 0, "lines": [ { "bbox": [ 200, 371, 410, 503 ], "spans": [ { "bbox": [ 200, 371, 410, 503 ], "score": 0.98, "html": "
VideoTimestamp TransformationEffect
Plant00:00TranslatePan
Turtle00:11Rand. translateShake
Coffee00:22TranslatePan
Camel00:33NoneNone
Volcano00:43Rand. translateShake
Bear00:54PerspectivePan
Penguins01:05NoneNone
Unicorn01:15ScaleZoom out
Buffalo01:26ScaleZoom out
Bigfoot01:37TranslatePan
", "type": "table", "image_path": "6c8044d77c84b11f2c4896d127b702016082a3364a9e9d67cefda83a58f19eb9.jpg" } ] } ], "index": 22.5, "virtual_lines": [ { "bbox": [ 200, 371, 410, 384.2 ], "spans": [], "index": 18 }, { "bbox": [ 200, 384.2, 410, 397.4 ], "spans": [], "index": 19 }, { "bbox": [ 200, 397.4, 410, 410.59999999999997 ], "spans": [], "index": 20 }, { "bbox": [ 200, 410.59999999999997, 410, 423.79999999999995 ], "spans": [], "index": 21 }, { "bbox": [ 200, 423.79999999999995, 410, 436.99999999999994 ], "spans": [], "index": 22 }, { "bbox": [ 200, 436.99999999999994, 410, 450.19999999999993 ], "spans": [], "index": 23 }, { "bbox": [ 200, 450.19999999999993, 410, 463.3999999999999 ], "spans": [], "index": 24 }, { "bbox": [ 200, 463.3999999999999, 410, 476.5999999999999 ], "spans": [], "index": 25 }, { "bbox": [ 200, 476.5999999999999, 410, 489.7999999999999 ], "spans": [], "index": 26 }, { "bbox": [ 200, 489.7999999999999, 410, 502.9999999999999 ], "spans": [], "index": 27 } ] } ], "index": 19.75 }, { "type": "title", "bbox": [ 107, 531, 243, 544 ], "lines": [ { "bbox": [ 105, 531, 244, 546 ], "spans": [ { "bbox": [ 105, 531, 244, 546 ], "score": 1.0, "content": "E EVALUATION DATASET", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 556, 505, 667 ], "lines": [ { "bbox": [ 105, 556, 505, 569 ], "spans": [ { "bbox": [ 105, 556, 505, 569 ], "score": 1.0, "content": "In all evaluations described in the paper, we used a dataset of 29 videos with 127 edit prompts.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 568, 505, 579 ], "spans": [ { "bbox": [ 106, 568, 505, 579 ], "score": 1.0, "content": "The dataset videos were selected from YouTube-8M Abu-El-Haija et al. (2016) and show animals,", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 578, 506, 592 ], "spans": [ { "bbox": [ 105, 578, 506, 592 ], "score": 1.0, "content": "people performing actions, vehicles, and other objects. The edit prompt categories are motion,", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 588, 506, 604 ], "spans": [ { "bbox": [ 105, 588, 506, 604 ], "score": 1.0, "content": "object, background, and style. In the motion category the prompts perform motion editing (e.g.", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 601, 506, 614 ], "spans": [ { "bbox": [ 106, 601, 506, 614 ], "score": 1.0, "content": "adding motion with the prompt “An orangutan next to a pond waving both arms in the air”), the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 610, 506, 626 ], "spans": [ { "bbox": [ 105, 610, 506, 626 ], "score": 1.0, "content": "object category performs object level edits (e.g. adding a party hat with the prompt “A puppy", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 621, 506, 636 ], "spans": [ { "bbox": [ 105, 621, 506, 636 ], "score": 1.0, "content": "walking with a party hat”), the background category performs edits of the background (e.g. adding", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 633, 505, 646 ], "spans": [ { "bbox": [ 105, 633, 505, 646 ], "score": 1.0, "content": "a river with the prompt “A blue pickup truck crossing a deep river”), and the style category performs", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 645, 505, 658 ], "spans": [ { "bbox": [ 105, 645, 505, 658 ], "score": 1.0, "content": "style-transfer like edits (e.g. changing the style to cartoon with the prompt “A cartoon of a man", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 656, 201, 668 ], "spans": [ { "bbox": [ 105, 656, 201, 668 ], "score": 1.0, "content": "playing a saxophone”).", "type": "text" } ], "index": 38 } ], "index": 33.5, "bbox_fs": [ 105, 556, 506, 668 ] }, { "type": "title", "bbox": [ 108, 684, 282, 697 ], "lines": [ { "bbox": [ 105, 683, 283, 698 ], "spans": [ { "bbox": [ 105, 683, 283, 698 ], "score": 1.0, "content": "F HUMAN EVALUATION DETAILS", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "We performed human evaluations for the baseline comparison and the ablation analysis. The eval-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 720, 505, 732 ], "spans": [ { "bbox": [ 105, 720, 505, 732 ], "score": 1.0, "content": "uations were conducted by a panel of 20 human raters using the dataset described in ??. The video", "type": "text" } ], "index": 41 } ], "index": 40.5, "bbox_fs": [ 105, 709, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 106, 83, 505, 94 ], "spans": [ { "bbox": [ 106, 83, 233, 94 ], "score": 1.0, "content": "resolution shown to raters was", "type": "text" }, { "bbox": [ 234, 83, 278, 93 ], "score": 0.9, "content": "3 5 0 \\times 2 0 0", "type": "inline_equation" }, { "bbox": [ 278, 83, 505, 94 ], "score": 1.0, "content": ", except for tune-a-video where we used a resolution of", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 93, 399, 107 ], "spans": [ { "bbox": [ 106, 93, 150, 105 ], "score": 0.88, "content": "2 0 0 \\times 2 0 0", "type": "inline_equation" }, { "bbox": [ 151, 93, 399, 107 ], "score": 1.0, "content": "(because we observed it performs better with square outputs).", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "title", "bbox": [ 107, 121, 209, 133 ], "lines": [ { "bbox": [ 105, 120, 210, 135 ], "spans": [ { "bbox": [ 105, 120, 210, 135 ], "score": 1.0, "content": "F.1 ABLATION STUDY", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 103, 143, 505, 165 ], "lines": [ { "bbox": [ 104, 141, 505, 156 ], "spans": [ { "bbox": [ 104, 141, 505, 156 ], "score": 1.0, "content": "In the ablation study the raters were asked to select the best edited video out of 12 hyperparameter", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 154, 165, 165 ], "spans": [ { "bbox": [ 106, 154, 165, 165 ], "score": 1.0, "content": "combinations:", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 133, 175, 477, 225 ], "lines": [ { "bbox": [ 133, 176, 307, 188 ], "spans": [ { "bbox": [ 133, 176, 220, 188 ], "score": 1.0, "content": "• no finetuning, with", "type": "text" }, { "bbox": [ 221, 176, 307, 188 ], "score": 0.51, "content": "s \\in { 0 . 4 , 0 . 7 , 0 . 8 , 0 . 8 5 }", "type": "inline_equation" } ], "index": 5 }, { "bbox": [ 134, 194, 375, 206 ], "spans": [ { "bbox": [ 134, 194, 281, 206 ], "score": 1.0, "content": "• video finetuning for 64 steps, with", "type": "text" }, { "bbox": [ 282, 194, 375, 206 ], "score": 0.65, "content": "s \\in { 0 . 8 , 0 . 9 , 0 . 9 5 , 0 . 9 8 }", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 133, 210, 477, 225 ], "spans": [ { "bbox": [ 133, 210, 236, 225 ], "score": 1.0, "content": "• mixed finetuning, with", "type": "text" }, { "bbox": [ 236, 211, 477, 225 ], "score": 0.58, "content": "( f t _ { s t e p s } , s ) \\in ( 1 5 0 , 0 . 9 8 ) , ( 2 0 0 , 0 . 9 8 ) , ( 2 0 0 , 1 . 0 ) , ( 3 0 0 , 1 . 0 )", "type": "inline_equation" } ], "index": 7 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 234, 505, 257 ], "lines": [ { "bbox": [ 105, 233, 505, 247 ], "spans": [ { "bbox": [ 105, 233, 505, 247 ], "score": 1.0, "content": "A visual example of the tradeoff between the amount of noise and the amount of finetuning is shown", "type": "text" } ], "index": 8 }, { "bbox": [ 104, 243, 147, 260 ], "spans": [ { "bbox": [ 104, 243, 147, 260 ], "score": 1.0, "content": "in Fig. 8.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "image", "bbox": [ 108, 273, 497, 515 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 273, 497, 515 ], "group_id": 0, "lines": [ { "bbox": [ 108, 273, 497, 515 ], "spans": [ { "bbox": [ 108, 273, 497, 515 ], "score": 0.976, "type": "image", "image_path": "f61d12177e88fc45bec3637d3db694589034642670dfd46ed02f40958e594f9d.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 108, 273, 497, 353.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 108, 353.6666666666667, 497, 434.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 108, 434.33333333333337, 497, 515.0 ], "spans": [], "index": 12 } ] }, { "type": "image_caption", "bbox": [ 106, 526, 506, 592 ], "group_id": 0, "lines": [ { "bbox": [ 106, 527, 505, 538 ], "spans": [ { "bbox": [ 106, 527, 505, 538 ], "score": 1.0, "content": "Figure 8: Noise-Finetuning Tradeoffs: We compare the effect of noise magnitude and number of", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 505, 550 ], "score": 1.0, "content": "finetuning iterations on edited videos. The original frame is on the bottom left, the rest were gener-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 548, 505, 560 ], "spans": [ { "bbox": [ 105, 548, 505, 560 ], "score": 1.0, "content": "ated by different parameters for the prompt ”An orangutan with orange hair bathing in a bathroom”.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 559, 505, 572 ], "spans": [ { "bbox": [ 106, 559, 505, 572 ], "score": 1.0, "content": "We can observe that higher noise allows for larger edits but reduces fidelity. More finetuning it-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 569, 506, 585 ], "spans": [ { "bbox": [ 105, 569, 506, 585 ], "score": 1.0, "content": "erations improve fidelity at higher noises. The best results are obtained for high noise and a large", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 581, 233, 594 ], "spans": [ { "bbox": [ 105, 581, 233, 594 ], "score": 1.0, "content": "number of finetuning iterations", "type": "text" } ], "index": 18 } ], "index": 15.5 } ], "index": 13.25 }, { "type": "title", "bbox": [ 107, 619, 235, 630 ], "lines": [ { "bbox": [ 105, 619, 236, 632 ], "spans": [ { "bbox": [ 105, 619, 236, 632 ], "score": 1.0, "content": "F.2 BASELINE COMPARISON", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 640, 504, 674 ], "lines": [ { "bbox": [ 105, 640, 505, 654 ], "spans": [ { "bbox": [ 105, 640, 505, 654 ], "score": 1.0, "content": "In the baseline comparisons described in the paper, raters evaluated videos on quality, fidelity and", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 651, 506, 665 ], "spans": [ { "bbox": [ 105, 651, 506, 665 ], "score": 1.0, "content": "alignment. The raters saw the original video alongside an edited video and answered the following", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 663, 227, 674 ], "spans": [ { "bbox": [ 105, 663, 199, 674 ], "score": 1.0, "content": "questions on a scale of", "type": "text" }, { "bbox": [ 199, 663, 222, 673 ], "score": 0.6, "content": "1 - 5", "type": "inline_equation" }, { "bbox": [ 223, 663, 227, 674 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 22 } ], "index": 21 }, { "type": "text", "bbox": [ 128, 684, 468, 732 ], "lines": [ { "bbox": [ 130, 684, 418, 697 ], "spans": [ { "bbox": [ 130, 684, 418, 697 ], "score": 1.0, "content": "1. “Rate the overall visual quality and smoothness of the edited video.”", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 702, 457, 716 ], "spans": [ { "bbox": [ 129, 702, 457, 716 ], "score": 1.0, "content": "2. “How well does the edited video match the textual edit description provided?”", "type": "text" } ], "index": 24 }, { "bbox": [ 128, 719, 469, 734 ], "spans": [ { "bbox": [ 128, 719, 469, 734 ], "score": 1.0, "content": "3. “How well does the edited video preserve unedited details of the original video?”", "type": "text" } ], "index": 25 } ], "index": 24 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 26, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "16", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 106, 83, 505, 94 ], "spans": [ { "bbox": [ 106, 83, 233, 94 ], "score": 1.0, "content": "resolution shown to raters was", "type": "text" }, { "bbox": [ 234, 83, 278, 93 ], "score": 0.9, "content": "3 5 0 \\times 2 0 0", "type": "inline_equation" }, { "bbox": [ 278, 83, 505, 94 ], "score": 1.0, "content": ", except for tune-a-video where we used a resolution of", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 93, 399, 107 ], "spans": [ { "bbox": [ 106, 93, 150, 105 ], "score": 0.88, "content": "2 0 0 \\times 2 0 0", "type": "inline_equation" }, { "bbox": [ 151, 93, 399, 107 ], "score": 1.0, "content": "(because we observed it performs better with square outputs).", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 106, 83, 505, 107 ] }, { "type": "title", "bbox": [ 107, 121, 209, 133 ], "lines": [ { "bbox": [ 105, 120, 210, 135 ], "spans": [ { "bbox": [ 105, 120, 210, 135 ], "score": 1.0, "content": "F.1 ABLATION STUDY", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 103, 143, 505, 165 ], "lines": [ { "bbox": [ 104, 141, 505, 156 ], "spans": [ { "bbox": [ 104, 141, 505, 156 ], "score": 1.0, "content": "In the ablation study the raters were asked to select the best edited video out of 12 hyperparameter", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 154, 165, 165 ], "spans": [ { "bbox": [ 106, 154, 165, 165 ], "score": 1.0, "content": "combinations:", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 104, 141, 505, 165 ] }, { "type": "list", "bbox": [ 133, 175, 477, 225 ], "lines": [ { "bbox": [ 133, 176, 307, 188 ], "spans": [ { "bbox": [ 133, 176, 220, 188 ], "score": 1.0, "content": "• no finetuning, with", "type": "text" }, { "bbox": [ 221, 176, 307, 188 ], "score": 0.51, "content": "s \\in { 0 . 4 , 0 . 7 , 0 . 8 , 0 . 8 5 }", "type": "inline_equation" } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 134, 194, 375, 206 ], "spans": [ { "bbox": [ 134, 194, 281, 206 ], "score": 1.0, "content": "• video finetuning for 64 steps, with", "type": "text" }, { "bbox": [ 282, 194, 375, 206 ], "score": 0.65, "content": "s \\in { 0 . 8 , 0 . 9 , 0 . 9 5 , 0 . 9 8 }", "type": "inline_equation" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 133, 210, 477, 225 ], "spans": [ { "bbox": [ 133, 210, 236, 225 ], "score": 1.0, "content": "• mixed finetuning, with", "type": "text" }, { "bbox": [ 236, 211, 477, 225 ], "score": 0.58, "content": "( f t _ { s t e p s } , s ) \\in ( 1 5 0 , 0 . 9 8 ) , ( 2 0 0 , 0 . 9 8 ) , ( 2 0 0 , 1 . 0 ) , ( 3 0 0 , 1 . 0 )", "type": "inline_equation" } ], "index": 7, "is_list_start_line": true } ], "index": 6, "bbox_fs": [ 133, 176, 477, 225 ] }, { "type": "text", "bbox": [ 107, 234, 505, 257 ], "lines": [ { "bbox": [ 105, 233, 505, 247 ], "spans": [ { "bbox": [ 105, 233, 505, 247 ], "score": 1.0, "content": "A visual example of the tradeoff between the amount of noise and the amount of finetuning is shown", "type": "text" } ], "index": 8 }, { "bbox": [ 104, 243, 147, 260 ], "spans": [ { "bbox": [ 104, 243, 147, 260 ], "score": 1.0, "content": "in Fig. 8.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 104, 233, 505, 260 ] }, { "type": "image", "bbox": [ 108, 273, 497, 515 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 273, 497, 515 ], "group_id": 0, "lines": [ { "bbox": [ 108, 273, 497, 515 ], "spans": [ { "bbox": [ 108, 273, 497, 515 ], "score": 0.976, "type": "image", "image_path": "f61d12177e88fc45bec3637d3db694589034642670dfd46ed02f40958e594f9d.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 108, 273, 497, 353.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 108, 353.6666666666667, 497, 434.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 108, 434.33333333333337, 497, 515.0 ], "spans": [], "index": 12 } ] }, { "type": "image_caption", "bbox": [ 106, 526, 506, 592 ], "group_id": 0, "lines": [ { "bbox": [ 106, 527, 505, 538 ], "spans": [ { "bbox": [ 106, 527, 505, 538 ], "score": 1.0, "content": "Figure 8: Noise-Finetuning Tradeoffs: We compare the effect of noise magnitude and number of", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 505, 550 ], "score": 1.0, "content": "finetuning iterations on edited videos. The original frame is on the bottom left, the rest were gener-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 548, 505, 560 ], "spans": [ { "bbox": [ 105, 548, 505, 560 ], "score": 1.0, "content": "ated by different parameters for the prompt ”An orangutan with orange hair bathing in a bathroom”.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 559, 505, 572 ], "spans": [ { "bbox": [ 106, 559, 505, 572 ], "score": 1.0, "content": "We can observe that higher noise allows for larger edits but reduces fidelity. More finetuning it-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 569, 506, 585 ], "spans": [ { "bbox": [ 105, 569, 506, 585 ], "score": 1.0, "content": "erations improve fidelity at higher noises. The best results are obtained for high noise and a large", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 581, 233, 594 ], "spans": [ { "bbox": [ 105, 581, 233, 594 ], "score": 1.0, "content": "number of finetuning iterations", "type": "text" } ], "index": 18 } ], "index": 15.5 } ], "index": 13.25 }, { "type": "title", "bbox": [ 107, 619, 235, 630 ], "lines": [ { "bbox": [ 105, 619, 236, 632 ], "spans": [ { "bbox": [ 105, 619, 236, 632 ], "score": 1.0, "content": "F.2 BASELINE COMPARISON", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 640, 504, 674 ], "lines": [ { "bbox": [ 105, 640, 505, 654 ], "spans": [ { "bbox": [ 105, 640, 505, 654 ], "score": 1.0, "content": "In the baseline comparisons described in the paper, raters evaluated videos on quality, fidelity and", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 651, 506, 665 ], "spans": [ { "bbox": [ 105, 651, 506, 665 ], "score": 1.0, "content": "alignment. The raters saw the original video alongside an edited video and answered the following", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 663, 227, 674 ], "spans": [ { "bbox": [ 105, 663, 199, 674 ], "score": 1.0, "content": "questions on a scale of", "type": "text" }, { "bbox": [ 199, 663, 222, 673 ], "score": 0.6, "content": "1 - 5", "type": "inline_equation" }, { "bbox": [ 223, 663, 227, 674 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 22 } ], "index": 21, "bbox_fs": [ 105, 640, 506, 674 ] }, { "type": "index", "bbox": [ 128, 684, 468, 732 ], "lines": [ { "bbox": [ 130, 684, 418, 697 ], "spans": [ { "bbox": [ 130, 684, 418, 697 ], "score": 1.0, "content": "1. “Rate the overall visual quality and smoothness of the edited video.”", "type": "text" } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 129, 702, 457, 716 ], "spans": [ { "bbox": [ 129, 702, 457, 716 ], "score": 1.0, "content": "2. “How well does the edited video match the textual edit description provided?”", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 128, 719, 469, 734 ], "spans": [ { "bbox": [ 128, 719, 469, 734 ], "score": 1.0, "content": "3. “How well does the edited video preserve unedited details of the original video?”", "type": "text" } ], "index": 25, "is_list_start_line": true } ], "index": 24, "bbox_fs": [ 128, 684, 469, 734 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 82, 223, 93 ], "lines": [ { "bbox": [ 105, 81, 225, 96 ], "spans": [ { "bbox": [ 105, 81, 225, 96 ], "score": 1.0, "content": "F.3 DIRECT COMPARISON", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 102, 505, 147 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "We also conducted a direct comparisons between the different editing methods. In this comparison", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 506, 127 ], "spans": [ { "bbox": [ 105, 113, 506, 127 ], "score": 1.0, "content": "raters saw the videos simultaneously and selected the best edit. We conducted the comparison once", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 125, 505, 137 ], "spans": [ { "bbox": [ 105, 125, 505, 137 ], "score": 1.0, "content": "with a fixed set of hyperparameters for Dreamix, and once more showing a single Dreamix video", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 136, 388, 148 ], "spans": [ { "bbox": [ 106, 136, 388, 148 ], "score": 1.0, "content": "chosen among 12 hyperparameters sets. Results can be seen in Tab. 5.", "type": "text" } ], "index": 4 } ], "index": 2.5 }, { "type": "table", "bbox": [ 198, 215, 411, 286 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 157, 505, 203 ], "group_id": 0, "lines": [ { "bbox": [ 105, 156, 506, 170 ], "spans": [ { "bbox": [ 105, 156, 506, 170 ], "score": 1.0, "content": "Table 5: Direct comparison of editing methods: Users were shown editing results of different", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 168, 505, 180 ], "spans": [ { "bbox": [ 106, 168, 505, 180 ], "score": 1.0, "content": "editing methods were asked to pick the best one. In the ”Multiple HP” column, the Dreamix video", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 179, 506, 192 ], "spans": [ { "bbox": [ 105, 179, 506, 192 ], "score": 1.0, "content": "was chosen from a set of 12 hyperparameters. We show the number of times each method got a", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 191, 232, 203 ], "spans": [ { "bbox": [ 106, 191, 232, 203 ], "score": 1.0, "content": "majority vote (out of 5 ratings)", "type": "text" } ], "index": 8 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 198, 215, 411, 286 ], "group_id": 0, "lines": [ { "bbox": [ 198, 215, 411, 286 ], "spans": [ { "bbox": [ 198, 215, 411, 286 ], "score": 0.979, "html": "
MethodSingle HP Multiple HP
Plug-and-Play2%1%
Tune-a-Video6%6%
Ours34%77%
No good edit / Uncond.58%16%
", "type": "table", "image_path": "793f920703a410915d4fce420678a41a24583494a09b51975f256d631c58c1ee.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 198, 215, 411, 229.2 ], "spans": [], "index": 9 }, { "bbox": [ 198, 229.2, 411, 243.39999999999998 ], "spans": [], "index": 10 }, { "bbox": [ 198, 243.39999999999998, 411, 257.59999999999997 ], "spans": [], "index": 11 }, { "bbox": [ 198, 257.59999999999997, 411, 271.79999999999995 ], "spans": [], "index": 12 }, { "bbox": [ 198, 271.79999999999995, 411, 285.99999999999994 ], "spans": [], "index": 13 } ] } ], "index": 8.75 }, { "type": "title", "bbox": [ 107, 312, 276, 326 ], "lines": [ { "bbox": [ 105, 310, 279, 329 ], "spans": [ { "bbox": [ 105, 310, 279, 329 ], "score": 1.0, "content": "G QUANTITATIVE EVALUATION", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 336, 506, 448 ], "lines": [ { "bbox": [ 105, 336, 505, 350 ], "spans": [ { "bbox": [ 105, 336, 505, 350 ], "score": 1.0, "content": "In the quantitative baseline comparison described in the paper we reported alignment and quality.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 348, 506, 361 ], "spans": [ { "bbox": [ 106, 348, 506, 361 ], "score": 1.0, "content": "We measure alignment by the frame-level CLIP Score Hessel et al. (2021). That is, we compute the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 358, 506, 373 ], "spans": [ { "bbox": [ 105, 358, 506, 373 ], "score": 1.0, "content": "cosine similarity between the CLIP Radford et al. (2021) embedding and the CLIP text embedding", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 370, 506, 383 ], "spans": [ { "bbox": [ 106, 370, 506, 383 ], "score": 1.0, "content": "for each frame. For each video we take the average over all frames, finally we report the mean over", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 381, 506, 394 ], "spans": [ { "bbox": [ 106, 381, 506, 394 ], "score": 1.0, "content": "all the videos. For quality (stability) we compute the LPIPS Zhang et al. (2018) distance between all", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 505, 405 ], "score": 1.0, "content": "pairs of consecutive frames. For each video we take the average over all pairs of consecutive frames,", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 403, 505, 415 ], "spans": [ { "bbox": [ 106, 403, 505, 415 ], "score": 1.0, "content": "finally we report the mean over all the videos. To perform a fair comparison with Tune-a-Video", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 414, 506, 426 ], "spans": [ { "bbox": [ 106, 414, 506, 426 ], "score": 1.0, "content": "(which outputs videos of 24 frames at 5 fps) we subsampled the rest of the methods to match this", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 425, 506, 438 ], "spans": [ { "bbox": [ 106, 425, 506, 438 ], "score": 1.0, "content": "framerate. Additionally, before passing through CLIP and LPIPS, all the frames are preprocessed to", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 435, 470, 449 ], "spans": [ { "bbox": [ 105, 435, 470, 449 ], "score": 1.0, "content": "match the required format (i.e. resize to 224, center crop to 224, ImageNet normalization).", "type": "text" } ], "index": 24 } ], "index": 19.5 }, { "type": "title", "bbox": [ 107, 464, 238, 477 ], "lines": [ { "bbox": [ 105, 462, 240, 478 ], "spans": [ { "bbox": [ 105, 462, 240, 478 ], "score": 1.0, "content": "H IMAGE ATTRIBUTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 133, 488, 578, 650 ], "lines": [ { "bbox": [ 130, 487, 413, 501 ], "spans": [ { "bbox": [ 130, 487, 413, 501 ], "score": 1.0, "content": "• Desert - https://unsplash.com/photos/PP8Escz15d8", "type": "text" } ], "index": 26 }, { "bbox": [ 133, 504, 436, 515 ], "spans": [ { "bbox": [ 133, 504, 436, 515 ], "score": 1.0, "content": "• Fuji mountain https://unsplash.com/photos/9Qwbfa_RM94", "type": "text" } ], "index": 27 }, { "bbox": [ 131, 516, 438, 531 ], "spans": [ { "bbox": [ 131, 516, 438, 531 ], "score": 1.0, "content": "• Tree in snow - https://unsplash.com/photos/aQNy0za7x0k", "type": "text" } ], "index": 28 }, { "bbox": [ 132, 532, 435, 546 ], "spans": [ { "bbox": [ 132, 532, 435, 546 ], "score": 1.0, "content": "• Hut in snow - https://unsplash.com/photos/qV2p17GHKbs", "type": "text" } ], "index": 29 }, { "bbox": [ 132, 546, 448, 561 ], "spans": [ { "bbox": [ 132, 546, 448, 561 ], "score": 1.0, "content": "• Lake with trees - https://unsplash.com/photos/dIQlgwq6V3Y", "type": "text" } ], "index": 30 }, { "bbox": [ 132, 561, 407, 577 ], "spans": [ { "bbox": [ 132, 561, 407, 577 ], "score": 1.0, "content": "• Plant - https://unsplash.com/photos/LrPKL7jOldI", "type": "text" } ], "index": 31 }, { "bbox": [ 132, 578, 410, 591 ], "spans": [ { "bbox": [ 132, 578, 410, 591 ], "score": 1.0, "content": "• Turtle - https://unsplash.com/photos/za9MCg787eI", "type": "text" } ], "index": 32 }, { "bbox": [ 132, 591, 425, 606 ], "spans": [ { "bbox": [ 132, 591, 425, 606 ], "score": 1.0, "content": "• Yosemite - https://unsplash.com/photos/NRQV-hBF10M", "type": "text" } ], "index": 33 }, { "bbox": [ 132, 606, 437, 621 ], "spans": [ { "bbox": [ 132, 606, 437, 621 ], "score": 1.0, "content": "• Foggy forest - https://unsplash.com/photos/pKNqyx_v62s", "type": "text" } ], "index": 34 }, { "bbox": [ 132, 621, 414, 636 ], "spans": [ { "bbox": [ 132, 621, 414, 636 ], "score": 1.0, "content": "• Coffee - https://unsplash.com/photos/SMPe5xfbPT0", "type": "text" } ], "index": 35 }, { "bbox": [ 132, 636, 576, 652 ], "spans": [ { "bbox": [ 132, 636, 576, 652 ], "score": 1.0, "content": "• Monkey - https://www.pexels.com/video/a-brown-monkey-eating-bread-2436088/", "type": "text" } ], "index": 36 } ], "index": 31 }, { "type": "title", "bbox": [ 107, 666, 241, 678 ], "lines": [ { "bbox": [ 104, 664, 241, 680 ], "spans": [ { "bbox": [ 104, 664, 241, 680 ], "score": 1.0, "content": "I ADDITIONAL RESULTS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 105, 690, 502, 702 ], "lines": [ { "bbox": [ 106, 690, 502, 703 ], "spans": [ { "bbox": [ 106, 690, 502, 703 ], "score": 1.0, "content": "Below we present additional results of our method, for the best experience see the included videos.", "type": "text" } ], "index": 38 } ], "index": 38 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 82, 223, 93 ], "lines": [ { "bbox": [ 105, 81, 225, 96 ], "spans": [ { "bbox": [ 105, 81, 225, 96 ], "score": 1.0, "content": "F.3 DIRECT COMPARISON", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 102, 505, 147 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "We also conducted a direct comparisons between the different editing methods. In this comparison", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 506, 127 ], "spans": [ { "bbox": [ 105, 113, 506, 127 ], "score": 1.0, "content": "raters saw the videos simultaneously and selected the best edit. We conducted the comparison once", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 125, 505, 137 ], "spans": [ { "bbox": [ 105, 125, 505, 137 ], "score": 1.0, "content": "with a fixed set of hyperparameters for Dreamix, and once more showing a single Dreamix video", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 136, 388, 148 ], "spans": [ { "bbox": [ 106, 136, 388, 148 ], "score": 1.0, "content": "chosen among 12 hyperparameters sets. Results can be seen in Tab. 5.", "type": "text" } ], "index": 4 } ], "index": 2.5, "bbox_fs": [ 105, 102, 506, 148 ] }, { "type": "table", "bbox": [ 198, 215, 411, 286 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 157, 505, 203 ], "group_id": 0, "lines": [ { "bbox": [ 105, 156, 506, 170 ], "spans": [ { "bbox": [ 105, 156, 506, 170 ], "score": 1.0, "content": "Table 5: Direct comparison of editing methods: Users were shown editing results of different", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 168, 505, 180 ], "spans": [ { "bbox": [ 106, 168, 505, 180 ], "score": 1.0, "content": "editing methods were asked to pick the best one. In the ”Multiple HP” column, the Dreamix video", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 179, 506, 192 ], "spans": [ { "bbox": [ 105, 179, 506, 192 ], "score": 1.0, "content": "was chosen from a set of 12 hyperparameters. We show the number of times each method got a", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 191, 232, 203 ], "spans": [ { "bbox": [ 106, 191, 232, 203 ], "score": 1.0, "content": "majority vote (out of 5 ratings)", "type": "text" } ], "index": 8 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 198, 215, 411, 286 ], "group_id": 0, "lines": [ { "bbox": [ 198, 215, 411, 286 ], "spans": [ { "bbox": [ 198, 215, 411, 286 ], "score": 0.979, "html": "
MethodSingle HP Multiple HP
Plug-and-Play2%1%
Tune-a-Video6%6%
Ours34%77%
No good edit / Uncond.58%16%
", "type": "table", "image_path": "793f920703a410915d4fce420678a41a24583494a09b51975f256d631c58c1ee.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 198, 215, 411, 229.2 ], "spans": [], "index": 9 }, { "bbox": [ 198, 229.2, 411, 243.39999999999998 ], "spans": [], "index": 10 }, { "bbox": [ 198, 243.39999999999998, 411, 257.59999999999997 ], "spans": [], "index": 11 }, { "bbox": [ 198, 257.59999999999997, 411, 271.79999999999995 ], "spans": [], "index": 12 }, { "bbox": [ 198, 271.79999999999995, 411, 285.99999999999994 ], "spans": [], "index": 13 } ] } ], "index": 8.75 }, { "type": "title", "bbox": [ 107, 312, 276, 326 ], "lines": [ { "bbox": [ 105, 310, 279, 329 ], "spans": [ { "bbox": [ 105, 310, 279, 329 ], "score": 1.0, "content": "G QUANTITATIVE EVALUATION", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 336, 506, 448 ], "lines": [ { "bbox": [ 105, 336, 505, 350 ], "spans": [ { "bbox": [ 105, 336, 505, 350 ], "score": 1.0, "content": "In the quantitative baseline comparison described in the paper we reported alignment and quality.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 348, 506, 361 ], "spans": [ { "bbox": [ 106, 348, 506, 361 ], "score": 1.0, "content": "We measure alignment by the frame-level CLIP Score Hessel et al. (2021). That is, we compute the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 358, 506, 373 ], "spans": [ { "bbox": [ 105, 358, 506, 373 ], "score": 1.0, "content": "cosine similarity between the CLIP Radford et al. (2021) embedding and the CLIP text embedding", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 370, 506, 383 ], "spans": [ { "bbox": [ 106, 370, 506, 383 ], "score": 1.0, "content": "for each frame. For each video we take the average over all frames, finally we report the mean over", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 381, 506, 394 ], "spans": [ { "bbox": [ 106, 381, 506, 394 ], "score": 1.0, "content": "all the videos. For quality (stability) we compute the LPIPS Zhang et al. (2018) distance between all", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 505, 405 ], "score": 1.0, "content": "pairs of consecutive frames. For each video we take the average over all pairs of consecutive frames,", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 403, 505, 415 ], "spans": [ { "bbox": [ 106, 403, 505, 415 ], "score": 1.0, "content": "finally we report the mean over all the videos. To perform a fair comparison with Tune-a-Video", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 414, 506, 426 ], "spans": [ { "bbox": [ 106, 414, 506, 426 ], "score": 1.0, "content": "(which outputs videos of 24 frames at 5 fps) we subsampled the rest of the methods to match this", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 425, 506, 438 ], "spans": [ { "bbox": [ 106, 425, 506, 438 ], "score": 1.0, "content": "framerate. Additionally, before passing through CLIP and LPIPS, all the frames are preprocessed to", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 435, 470, 449 ], "spans": [ { "bbox": [ 105, 435, 470, 449 ], "score": 1.0, "content": "match the required format (i.e. resize to 224, center crop to 224, ImageNet normalization).", "type": "text" } ], "index": 24 } ], "index": 19.5, "bbox_fs": [ 105, 336, 506, 449 ] }, { "type": "title", "bbox": [ 107, 464, 238, 477 ], "lines": [ { "bbox": [ 105, 462, 240, 478 ], "spans": [ { "bbox": [ 105, 462, 240, 478 ], "score": 1.0, "content": "H IMAGE ATTRIBUTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "list", "bbox": [ 133, 488, 578, 650 ], "lines": [ { "bbox": [ 130, 487, 413, 501 ], "spans": [ { "bbox": [ 130, 487, 413, 501 ], "score": 1.0, "content": "• Desert - https://unsplash.com/photos/PP8Escz15d8", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 133, 504, 436, 515 ], "spans": [ { "bbox": [ 133, 504, 436, 515 ], "score": 1.0, "content": "• Fuji mountain https://unsplash.com/photos/9Qwbfa_RM94", "type": "text" } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 131, 516, 438, 531 ], "spans": [ { "bbox": [ 131, 516, 438, 531 ], "score": 1.0, "content": "• Tree in snow - https://unsplash.com/photos/aQNy0za7x0k", "type": "text" } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 132, 532, 435, 546 ], "spans": [ { "bbox": [ 132, 532, 435, 546 ], "score": 1.0, "content": "• Hut in snow - https://unsplash.com/photos/qV2p17GHKbs", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 132, 546, 448, 561 ], "spans": [ { "bbox": [ 132, 546, 448, 561 ], "score": 1.0, "content": "• Lake with trees - https://unsplash.com/photos/dIQlgwq6V3Y", "type": "text" } ], "index": 30, "is_list_start_line": true }, { "bbox": [ 132, 561, 407, 577 ], "spans": [ { "bbox": [ 132, 561, 407, 577 ], "score": 1.0, "content": "• Plant - https://unsplash.com/photos/LrPKL7jOldI", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 132, 578, 410, 591 ], "spans": [ { "bbox": [ 132, 578, 410, 591 ], "score": 1.0, "content": "• Turtle - https://unsplash.com/photos/za9MCg787eI", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 132, 591, 425, 606 ], "spans": [ { "bbox": [ 132, 591, 425, 606 ], "score": 1.0, "content": "• Yosemite - https://unsplash.com/photos/NRQV-hBF10M", "type": "text" } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 132, 606, 437, 621 ], "spans": [ { "bbox": [ 132, 606, 437, 621 ], "score": 1.0, "content": "• Foggy forest - https://unsplash.com/photos/pKNqyx_v62s", "type": "text" } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 132, 621, 414, 636 ], "spans": [ { "bbox": [ 132, 621, 414, 636 ], "score": 1.0, "content": "• Coffee - https://unsplash.com/photos/SMPe5xfbPT0", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 132, 636, 576, 652 ], "spans": [ { "bbox": [ 132, 636, 576, 652 ], "score": 1.0, "content": "• Monkey - https://www.pexels.com/video/a-brown-monkey-eating-bread-2436088/", "type": "text" } ], "index": 36, "is_list_start_line": true } ], "index": 31, "bbox_fs": [ 130, 487, 576, 652 ] }, { "type": "title", "bbox": [ 107, 666, 241, 678 ], "lines": [ { "bbox": [ 104, 664, 241, 680 ], "spans": [ { "bbox": [ 104, 664, 241, 680 ], "score": 1.0, "content": "I ADDITIONAL RESULTS", "type": "text" } ], "index": 37 } ], "index": 37 }, { "type": "text", "bbox": [ 105, 690, 502, 702 ], "lines": [ { "bbox": [ 106, 690, 502, 703 ], "spans": [ { "bbox": [ 106, 690, 502, 703 ], "score": 1.0, "content": "Below we present additional results of our method, for the best experience see the included videos.", "type": "text" } ], "index": 38 } ], "index": 38, "bbox_fs": [ 106, 690, 502, 703 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 234, 503, 566 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 234, 503, 566 ], "group_id": 0, "lines": [ { "bbox": [ 107, 234, 503, 566 ], "spans": [ { "bbox": [ 107, 234, 503, 566 ], "score": 0.973, "type": "image", "image_path": "869998a6c334daf2675316863d6e6f1eee0c58683f0a443aa356845be1fdc7b9.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 234, 503, 344.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 107, 344.6666666666667, 503, 455.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 107, 455.33333333333337, 503, 566.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 206, 574, 405, 587 ], "group_id": 0, "lines": [ { "bbox": [ 205, 572, 405, 589 ], "spans": [ { "bbox": [ 205, 572, 405, 589 ], "score": 1.0, "content": "Figure 9: Additional Video Editing Results (1/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 38 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 763 ], "spans": [ { "bbox": [ 299, 750, 313, 763 ], "score": 1.0, "content": "18", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 107, 234, 503, 566 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 234, 503, 566 ], "group_id": 0, "lines": [ { "bbox": [ 107, 234, 503, 566 ], "spans": [ { "bbox": [ 107, 234, 503, 566 ], "score": 0.973, "type": "image", "image_path": "869998a6c334daf2675316863d6e6f1eee0c58683f0a443aa356845be1fdc7b9.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 234, 503, 344.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 107, 344.6666666666667, 503, 455.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 107, 455.33333333333337, 503, 566.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 206, 574, 405, 587 ], "group_id": 0, "lines": [ { "bbox": [ 205, 572, 405, 589 ], "spans": [ { "bbox": [ 205, 572, 405, 589 ], "score": 1.0, "content": "Figure 9: Additional Video Editing Results (1/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 114, 160, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 160, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 114, 160, 509, 630 ], "spans": [ { "bbox": [ 114, 160, 509, 630 ], "score": 0.97, "type": "image", "image_path": "db9fb1b971109f0f5f46afd55be0c667e2e5389f836fd6a1aadc534afb414bfa.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 160, 509, 316.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 114, 316.66666666666663, 509, 473.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 114, 473.33333333333326, 509, 629.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 651 ], "group_id": 0, "lines": [ { "bbox": [ 198, 636, 413, 654 ], "spans": [ { "bbox": [ 198, 636, 413, 654 ], "score": 1.0, "content": "Figure 10: Additional Video Editing Examples (2/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 18, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 765 ], "spans": [ { "bbox": [ 299, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 114, 160, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 160, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 114, 160, 509, 630 ], "spans": [ { "bbox": [ 114, 160, 509, 630 ], "score": 0.97, "type": "image", "image_path": "db9fb1b971109f0f5f46afd55be0c667e2e5389f836fd6a1aadc534afb414bfa.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 160, 509, 316.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 114, 316.66666666666663, 509, 473.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 114, 473.33333333333326, 509, 629.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 651 ], "group_id": 0, "lines": [ { "bbox": [ 198, 636, 413, 654 ], "spans": [ { "bbox": [ 198, 636, 413, 654 ], "score": 1.0, "content": "Figure 10: Additional Video Editing Examples (2/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 113, 166, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 166, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 113, 166, 509, 630 ], "spans": [ { "bbox": [ 113, 166, 509, 630 ], "score": 0.964, "type": "image", "image_path": "ec9e94647162e765b551a40f4bd04f3efbd6aebb1974a2066fadd49b2aed92e2.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 113, 166, 509, 320.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 113, 320.66666666666663, 509, 475.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 113, 475.33333333333326, 509, 629.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 637, 413, 650 ], "group_id": 0, "lines": [ { "bbox": [ 198, 635, 412, 653 ], "spans": [ { "bbox": [ 198, 635, 412, 653 ], "score": 1.0, "content": "Figure 11: Additional Video Editing Examples (3/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 19, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 113, 166, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 166, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 113, 166, 509, 630 ], "spans": [ { "bbox": [ 113, 166, 509, 630 ], "score": 0.964, "type": "image", "image_path": "ec9e94647162e765b551a40f4bd04f3efbd6aebb1974a2066fadd49b2aed92e2.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 113, 166, 509, 320.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 113, 320.66666666666663, 509, 475.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 113, 475.33333333333326, 509, 629.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 637, 413, 650 ], "group_id": 0, "lines": [ { "bbox": [ 198, 635, 412, 653 ], "spans": [ { "bbox": [ 198, 635, 412, 653 ], "score": 1.0, "content": "Figure 11: Additional Video Editing Examples (3/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 114, 162, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 162, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 114, 162, 509, 630 ], "spans": [ { "bbox": [ 114, 162, 509, 630 ], "score": 0.966, "type": "image", "image_path": "39cdb8f30d1e7af28bd598a781b8f615419e02773098d8f9e29532ef9361b376.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 162, 509, 318.0 ], "spans": [], "index": 0 }, { "bbox": [ 114, 318.0, 509, 474.0 ], "spans": [], "index": 1 }, { "bbox": [ 114, 474.0, 509, 630.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 650 ], "group_id": 0, "lines": [ { "bbox": [ 198, 635, 413, 654 ], "spans": [ { "bbox": [ 198, 635, 413, 654 ], "score": 1.0, "content": "Figure 12: Additional Video Editing Examples (4/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 20, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 764 ], "spans": [ { "bbox": [ 298, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 114, 162, 509, 630 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 162, 509, 630 ], "group_id": 0, "lines": [ { "bbox": [ 114, 162, 509, 630 ], "spans": [ { "bbox": [ 114, 162, 509, 630 ], "score": 0.966, "type": "image", "image_path": "39cdb8f30d1e7af28bd598a781b8f615419e02773098d8f9e29532ef9361b376.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 162, 509, 318.0 ], "spans": [], "index": 0 }, { "bbox": [ 114, 318.0, 509, 474.0 ], "spans": [], "index": 1 }, { "bbox": [ 114, 474.0, 509, 630.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 650 ], "group_id": 0, "lines": [ { "bbox": [ 198, 635, 413, 654 ], "spans": [ { "bbox": [ 198, 635, 413, 654 ], "score": 1.0, "content": "Figure 12: Additional Video Editing Examples (4/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 110, 164, 513, 629 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 164, 513, 629 ], "group_id": 0, "lines": [ { "bbox": [ 110, 164, 513, 629 ], "spans": [ { "bbox": [ 110, 164, 513, 629 ], "score": 0.956, "type": "image", "image_path": "b667c0460165b558bc493f346aae63eb9f1d5e35e7d43fdf610c95382cd3e63e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 164, 513, 319.0 ], "spans": [], "index": 0 }, { "bbox": [ 110, 319.0, 513, 474.0 ], "spans": [], "index": 1 }, { "bbox": [ 110, 474.0, 513, 629.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 651 ], "group_id": 0, "lines": [ { "bbox": [ 198, 636, 412, 654 ], "spans": [ { "bbox": [ 198, 636, 412, 654 ], "score": 1.0, "content": "Figure 13: Additional Video Editing Examples (5/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 21, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 765 ], "spans": [ { "bbox": [ 298, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 110, 164, 513, 629 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 164, 513, 629 ], "group_id": 0, "lines": [ { "bbox": [ 110, 164, 513, 629 ], "spans": [ { "bbox": [ 110, 164, 513, 629 ], "score": 0.956, "type": "image", "image_path": "b667c0460165b558bc493f346aae63eb9f1d5e35e7d43fdf610c95382cd3e63e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 164, 513, 319.0 ], "spans": [], "index": 0 }, { "bbox": [ 110, 319.0, 513, 474.0 ], "spans": [], "index": 1 }, { "bbox": [ 110, 474.0, 513, 629.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 198, 638, 413, 651 ], "group_id": 0, "lines": [ { "bbox": [ 198, 636, 412, 654 ], "spans": [ { "bbox": [ 198, 636, 412, 654 ], "score": 1.0, "content": "Figure 13: Additional Video Editing Examples (5/5)", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 114, 175, 509, 619 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 175, 509, 619 ], "group_id": 0, "lines": [ { "bbox": [ 114, 175, 509, 619 ], "spans": [ { "bbox": [ 114, 175, 509, 619 ], "score": 0.948, "type": "image", "image_path": "3585b3fc6f8947a81dc5e42b945c0dc89f7b38164baad4155b4d7d65c4a7685e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 175, 509, 323.0 ], "spans": [], "index": 0 }, { "bbox": [ 114, 323.0, 509, 471.0 ], "spans": [], "index": 1 }, { "bbox": [ 114, 471.0, 509, 619.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 205, 622, 405, 634 ], "group_id": 0, "lines": [ { "bbox": [ 206, 621, 405, 635 ], "spans": [ { "bbox": [ 206, 621, 405, 635 ], "score": 1.0, "content": "Figure 14: Additional Image-to-Video Examples", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 22, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 38 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 114, 175, 509, 619 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 175, 509, 619 ], "group_id": 0, "lines": [ { "bbox": [ 114, 175, 509, 619 ], "spans": [ { "bbox": [ 114, 175, 509, 619 ], "score": 0.948, "type": "image", "image_path": "3585b3fc6f8947a81dc5e42b945c0dc89f7b38164baad4155b4d7d65c4a7685e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 175, 509, 323.0 ], "spans": [], "index": 0 }, { "bbox": [ 114, 323.0, 509, 471.0 ], "spans": [], "index": 1 }, { "bbox": [ 114, 471.0, 509, 619.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 205, 622, 405, 634 ], "group_id": 0, "lines": [ { "bbox": [ 206, 621, 405, 635 ], "spans": [ { "bbox": [ 206, 621, 405, 635 ], "score": 1.0, "content": "Figure 14: Additional Image-to-Video Examples", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 113, 220, 511, 581 ], "blocks": [ { "type": "image_caption", "bbox": [ 258, 205, 375, 217 ], "group_id": 0, "lines": [ { "bbox": [ 257, 204, 376, 218 ], "spans": [ { "bbox": [ 257, 204, 376, 218 ], "score": 1.0, "content": "“A bear is drinking from a glass”", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 113, 220, 511, 581 ], "group_id": 0, "lines": [ { "bbox": [ 113, 220, 511, 581 ], "spans": [ { "bbox": [ 113, 220, 511, 581 ], "score": 0.862, "type": "image", "image_path": "7d05f868b9f211fbf112279508228c0797e83edda976412008c1af5d6dde43b6.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 113, 220, 511, 340.3333333333333 ], "spans": [], "index": 1 }, { "bbox": [ 113, 340.3333333333333, 511, 460.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 113, 460.66666666666663, 511, 581.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 191, 595, 419, 607 ], "group_id": 0, "lines": [ { "bbox": [ 190, 594, 420, 608 ], "spans": [ { "bbox": [ 190, 594, 420, 608 ], "score": 1.0, "content": "Figure 15: Additional Subject-Driven Video Generation", "type": "text" } ], "index": 4 } ], "index": 4 } ], "index": 2 } ], "page_idx": 23, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 38 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 765 ], "spans": [ { "bbox": [ 298, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 113, 220, 511, 581 ], "blocks": [ { "type": "image_caption", "bbox": [ 258, 205, 375, 217 ], "group_id": 0, "lines": [ { "bbox": [ 257, 204, 376, 218 ], "spans": [ { "bbox": [ 257, 204, 376, 218 ], "score": 1.0, "content": "“A bear is drinking from a glass”", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image_body", "bbox": [ 113, 220, 511, 581 ], "group_id": 0, "lines": [ { "bbox": [ 113, 220, 511, 581 ], "spans": [ { "bbox": [ 113, 220, 511, 581 ], "score": 0.862, "type": "image", "image_path": "7d05f868b9f211fbf112279508228c0797e83edda976412008c1af5d6dde43b6.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 113, 220, 511, 340.3333333333333 ], "spans": [], "index": 1 }, { "bbox": [ 113, 340.3333333333333, 511, 460.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 113, 460.66666666666663, 511, 581.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 191, 595, 419, 607 ], "group_id": 0, "lines": [ { "bbox": [ 190, 594, 420, 608 ], "spans": [ { "bbox": [ 190, 594, 420, 608 ], "score": 1.0, "content": "Figure 15: Additional Subject-Driven Video Generation", "type": "text" } ], "index": 4 } ], "index": 4 } ], "index": 2 } ] } ], "_backend": "pipeline", "_version_name": "2.1.11" }