{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
71,
78,
367,
119
],
"lines": [
{
"bbox": [
69,
78,
369,
98
],
"spans": [
{
"bbox": [
69,
78,
369,
98
],
"score": 1.0,
"content": "Imagen Video: High Definition Video",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
70,
99,
338,
119
],
"spans": [
{
"bbox": [
70,
99,
338,
119
],
"score": 1.0,
"content": "Generation with Diffusion Models",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
71,
145,
237,
169
],
"lines": [
{
"bbox": [
71,
146,
173,
158
],
"spans": [
{
"bbox": [
71,
146,
173,
158
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
69,
158,
238,
170
],
"spans": [
{
"bbox": [
69,
158,
238,
170
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "title",
"bbox": [
282,
197,
329,
210
],
"lines": [
{
"bbox": [
280,
195,
331,
212
],
"spans": [
{
"bbox": [
280,
195,
331,
212
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
106,
227,
505,
382
],
"lines": [
{
"bbox": [
106,
227,
505,
239
],
"spans": [
{
"bbox": [
106,
227,
505,
239
],
"score": 1.0,
"content": "We present Imagen Video, a text-conditional video generation system based on a cascade of",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
239,
505,
252
],
"spans": [
{
"bbox": [
106,
239,
505,
252
],
"score": 1.0,
"content": "video diffusion models. Given a text prompt, Imagen Video generates high definition videos",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
251,
505,
263
],
"spans": [
{
"bbox": [
105,
251,
505,
263
],
"score": 1.0,
"content": "using a base video generation model and a sequence of interleaved spatial and temporal",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
263,
504,
275
],
"spans": [
{
"bbox": [
106,
263,
504,
275
],
"score": 1.0,
"content": "video super-resolution models. We describe how we scale up the system as a high defini-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
275,
504,
287
],
"spans": [
{
"bbox": [
106,
275,
504,
287
],
"score": 1.0,
"content": "tion text-to-video model including design decisions such as the choice of fully-convolutional",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
287,
504,
298
],
"spans": [
{
"bbox": [
106,
287,
504,
298
],
"score": 1.0,
"content": "temporal and spatial super-resolution models at certain resolutions, and the choice of the",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
299,
505,
311
],
"spans": [
{
"bbox": [
105,
299,
505,
311
],
"score": 1.0,
"content": "v-parameterization of diffusion models. In addition, we confirm and transfer findings from",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
311,
504,
324
],
"spans": [
{
"bbox": [
105,
311,
504,
324
],
"score": 1.0,
"content": "previous work on diffusion-based image generation to the video generation setting. Finally,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
322,
504,
335
],
"spans": [
{
"bbox": [
106,
322,
504,
335
],
"score": 1.0,
"content": "we apply progressive distillation to our video models with classifier-free guidance for fast,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
335,
505,
348
],
"spans": [
{
"bbox": [
106,
335,
505,
348
],
"score": 1.0,
"content": "high quality sampling. We find Imagen Video not only capable of generating videos of high",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
104,
345,
506,
360
],
"spans": [
{
"bbox": [
104,
345,
506,
360
],
"score": 1.0,
"content": "fidelity, but also having a high degree of controllability and world knowledge, including the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
359,
504,
370
],
"spans": [
{
"bbox": [
106,
359,
504,
370
],
"score": 1.0,
"content": "ability to generate diverse videos and text animations in various artistic styles and with 3D",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
369,
204,
385
],
"spans": [
{
"bbox": [
104,
369,
204,
385
],
"score": 1.0,
"content": "object understanding.",
"type": "text"
}
],
"index": 17
}
],
"index": 11
},
{
"type": "image",
"bbox": [
76,
404,
538,
544
],
"blocks": [
{
"type": "image_body",
"bbox": [
76,
404,
538,
544
],
"group_id": 0,
"lines": [
{
"bbox": [
76,
404,
538,
544
],
"spans": [
{
"bbox": [
76,
404,
538,
544
],
"score": 0.977,
"type": "image",
"image_path": "7278d4c9f33331da8e8af9b951b25b288446ba78053005bae48faadaa9da97cf.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
76,
404,
538,
450.6666666666667
],
"spans": [],
"index": 18
},
{
"bbox": [
76,
450.6666666666667,
538,
497.33333333333337
],
"spans": [],
"index": 19
},
{
"bbox": [
76,
497.33333333333337,
538,
544.0
],
"spans": [],
"index": 20
}
]
},
{
"type": "image_caption",
"bbox": [
72,
553,
541,
588
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
552,
542,
566
],
"spans": [
{
"bbox": [
70,
552,
542,
566
],
"score": 1.0,
"content": "Figure 1: Imagen Video sample for the prompt: “A bunch of autumn leaves falling on a calm lake to form",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
70,
564,
541,
578
],
"spans": [
{
"bbox": [
70,
564,
338,
578
],
"score": 1.0,
"content": "the text ‘Imagen Video’. Smooth.” The generated video is at",
"type": "text"
},
{
"bbox": [
339,
565,
383,
576
],
"score": 0.78,
"content": "1 2 8 0 \\times 7 6 8",
"type": "inline_equation"
},
{
"bbox": [
383,
564,
541,
578
],
"score": 1.0,
"content": "resolution, 5.3 second duration and",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
575,
167,
591
],
"spans": [
{
"bbox": [
69,
575,
167,
591
],
"score": 1.0,
"content": "24 frames per second.",
"type": "text"
}
],
"index": 23
}
],
"index": 22
}
],
"index": 20.5
},
{
"type": "title",
"bbox": [
72,
610,
159,
624
],
"lines": [
{
"bbox": [
68,
609,
160,
627
],
"spans": [
{
"bbox": [
68,
609,
160,
627
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
71,
636,
540,
732
],
"lines": [
{
"bbox": [
70,
636,
541,
649
],
"spans": [
{
"bbox": [
70,
636,
541,
649
],
"score": 1.0,
"content": "Generative modeling has made tremendous progress with recent text-to-image systems like",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
68,
646,
541,
663
],
"spans": [
{
"bbox": [
68,
646,
541,
663
],
"score": 1.0,
"content": "DALL-E 2 (Ramesh et al., 2022), Imagen (Saharia et al., 2022b), Parti (Yu et al., 2022), CogView (Ding",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
659,
541,
674
],
"spans": [
{
"bbox": [
68,
659,
541,
674
],
"score": 1.0,
"content": "et al., 2021) and Latent Diffusion (Rombach et al., 2022). Diffusion models (Sohl-Dickstein et al., 2015;",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "Ho et al., 2020) in particular have found considerable success in multiple generative modeling tasks (Nichol",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
68,
683,
541,
698
],
"spans": [
{
"bbox": [
68,
683,
541,
698
],
"score": 1.0,
"content": "& Dhariwal, 2021; Ho et al., 2022a; Dhariwal & Nichol, 2022) including density estimation (Kingma et al.,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
696,
541,
709
],
"spans": [
{
"bbox": [
69,
696,
541,
709
],
"score": 1.0,
"content": "2021), text-to-speech (Chen et al., 2021a; Kong et al., 2021; Chen et al., 2021b), image-to-image (Saharia",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
707,
541,
721
],
"spans": [
{
"bbox": [
69,
707,
541,
721
],
"score": 1.0,
"content": "et al., 2022c;a; Whang et al., 2022), text-to-image (Rombach et al., 2022; Nichol et al., 2021; Ramesh et al.,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
720,
450,
734
],
"spans": [
{
"bbox": [
69,
720,
450,
734
],
"score": 1.0,
"content": "2022; Saharia et al., 2022b) and 3D synthesis (Poole et al., 2022; Watson et al., 2022).",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
308,
760
],
"lines": [
{
"bbox": [
302,
750,
310,
763
],
"spans": [
{
"bbox": [
302,
750,
310,
763
],
"score": 1.0,
"content": "",
"type": "text",
"height": 13,
"width": 8
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
71,
78,
367,
119
],
"lines": [
{
"bbox": [
69,
78,
369,
98
],
"spans": [
{
"bbox": [
69,
78,
369,
98
],
"score": 1.0,
"content": "Imagen Video: High Definition Video",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
70,
99,
338,
119
],
"spans": [
{
"bbox": [
70,
99,
338,
119
],
"score": 1.0,
"content": "Generation with Diffusion Models",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
71,
145,
237,
169
],
"lines": [
{
"bbox": [
71,
146,
173,
158
],
"spans": [
{
"bbox": [
71,
146,
173,
158
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
69,
158,
238,
170
],
"spans": [
{
"bbox": [
69,
158,
238,
170
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
69,
146,
238,
170
]
},
{
"type": "title",
"bbox": [
282,
197,
329,
210
],
"lines": [
{
"bbox": [
280,
195,
331,
212
],
"spans": [
{
"bbox": [
280,
195,
331,
212
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
106,
227,
505,
382
],
"lines": [
{
"bbox": [
106,
227,
505,
239
],
"spans": [
{
"bbox": [
106,
227,
505,
239
],
"score": 1.0,
"content": "We present Imagen Video, a text-conditional video generation system based on a cascade of",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
239,
505,
252
],
"spans": [
{
"bbox": [
106,
239,
505,
252
],
"score": 1.0,
"content": "video diffusion models. Given a text prompt, Imagen Video generates high definition videos",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
251,
505,
263
],
"spans": [
{
"bbox": [
105,
251,
505,
263
],
"score": 1.0,
"content": "using a base video generation model and a sequence of interleaved spatial and temporal",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
263,
504,
275
],
"spans": [
{
"bbox": [
106,
263,
504,
275
],
"score": 1.0,
"content": "video super-resolution models. We describe how we scale up the system as a high defini-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
275,
504,
287
],
"spans": [
{
"bbox": [
106,
275,
504,
287
],
"score": 1.0,
"content": "tion text-to-video model including design decisions such as the choice of fully-convolutional",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
287,
504,
298
],
"spans": [
{
"bbox": [
106,
287,
504,
298
],
"score": 1.0,
"content": "temporal and spatial super-resolution models at certain resolutions, and the choice of the",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
299,
505,
311
],
"spans": [
{
"bbox": [
105,
299,
505,
311
],
"score": 1.0,
"content": "v-parameterization of diffusion models. In addition, we confirm and transfer findings from",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
311,
504,
324
],
"spans": [
{
"bbox": [
105,
311,
504,
324
],
"score": 1.0,
"content": "previous work on diffusion-based image generation to the video generation setting. Finally,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
322,
504,
335
],
"spans": [
{
"bbox": [
106,
322,
504,
335
],
"score": 1.0,
"content": "we apply progressive distillation to our video models with classifier-free guidance for fast,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
335,
505,
348
],
"spans": [
{
"bbox": [
106,
335,
505,
348
],
"score": 1.0,
"content": "high quality sampling. We find Imagen Video not only capable of generating videos of high",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
104,
345,
506,
360
],
"spans": [
{
"bbox": [
104,
345,
506,
360
],
"score": 1.0,
"content": "fidelity, but also having a high degree of controllability and world knowledge, including the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
359,
504,
370
],
"spans": [
{
"bbox": [
106,
359,
504,
370
],
"score": 1.0,
"content": "ability to generate diverse videos and text animations in various artistic styles and with 3D",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
104,
369,
204,
385
],
"spans": [
{
"bbox": [
104,
369,
204,
385
],
"score": 1.0,
"content": "object understanding.",
"type": "text"
}
],
"index": 17
}
],
"index": 11,
"bbox_fs": [
104,
227,
506,
385
]
},
{
"type": "image",
"bbox": [
76,
404,
538,
544
],
"blocks": [
{
"type": "image_body",
"bbox": [
76,
404,
538,
544
],
"group_id": 0,
"lines": [
{
"bbox": [
76,
404,
538,
544
],
"spans": [
{
"bbox": [
76,
404,
538,
544
],
"score": 0.977,
"type": "image",
"image_path": "7278d4c9f33331da8e8af9b951b25b288446ba78053005bae48faadaa9da97cf.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
76,
404,
538,
450.6666666666667
],
"spans": [],
"index": 18
},
{
"bbox": [
76,
450.6666666666667,
538,
497.33333333333337
],
"spans": [],
"index": 19
},
{
"bbox": [
76,
497.33333333333337,
538,
544.0
],
"spans": [],
"index": 20
}
]
},
{
"type": "image_caption",
"bbox": [
72,
553,
541,
588
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
552,
542,
566
],
"spans": [
{
"bbox": [
70,
552,
542,
566
],
"score": 1.0,
"content": "Figure 1: Imagen Video sample for the prompt: “A bunch of autumn leaves falling on a calm lake to form",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
70,
564,
541,
578
],
"spans": [
{
"bbox": [
70,
564,
338,
578
],
"score": 1.0,
"content": "the text ‘Imagen Video’. Smooth.” The generated video is at",
"type": "text"
},
{
"bbox": [
339,
565,
383,
576
],
"score": 0.78,
"content": "1 2 8 0 \\times 7 6 8",
"type": "inline_equation"
},
{
"bbox": [
383,
564,
541,
578
],
"score": 1.0,
"content": "resolution, 5.3 second duration and",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
575,
167,
591
],
"spans": [
{
"bbox": [
69,
575,
167,
591
],
"score": 1.0,
"content": "24 frames per second.",
"type": "text"
}
],
"index": 23
}
],
"index": 22
}
],
"index": 20.5
},
{
"type": "title",
"bbox": [
72,
610,
159,
624
],
"lines": [
{
"bbox": [
68,
609,
160,
627
],
"spans": [
{
"bbox": [
68,
609,
160,
627
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
71,
636,
540,
732
],
"lines": [
{
"bbox": [
70,
636,
541,
649
],
"spans": [
{
"bbox": [
70,
636,
541,
649
],
"score": 1.0,
"content": "Generative modeling has made tremendous progress with recent text-to-image systems like",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
68,
646,
541,
663
],
"spans": [
{
"bbox": [
68,
646,
541,
663
],
"score": 1.0,
"content": "DALL-E 2 (Ramesh et al., 2022), Imagen (Saharia et al., 2022b), Parti (Yu et al., 2022), CogView (Ding",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
659,
541,
674
],
"spans": [
{
"bbox": [
68,
659,
541,
674
],
"score": 1.0,
"content": "et al., 2021) and Latent Diffusion (Rombach et al., 2022). Diffusion models (Sohl-Dickstein et al., 2015;",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "Ho et al., 2020) in particular have found considerable success in multiple generative modeling tasks (Nichol",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
68,
683,
541,
698
],
"spans": [
{
"bbox": [
68,
683,
541,
698
],
"score": 1.0,
"content": "& Dhariwal, 2021; Ho et al., 2022a; Dhariwal & Nichol, 2022) including density estimation (Kingma et al.,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
696,
541,
709
],
"spans": [
{
"bbox": [
69,
696,
541,
709
],
"score": 1.0,
"content": "2021), text-to-speech (Chen et al., 2021a; Kong et al., 2021; Chen et al., 2021b), image-to-image (Saharia",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
707,
541,
721
],
"spans": [
{
"bbox": [
69,
707,
541,
721
],
"score": 1.0,
"content": "et al., 2022c;a; Whang et al., 2022), text-to-image (Rombach et al., 2022; Nichol et al., 2021; Ramesh et al.,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
720,
450,
734
],
"spans": [
{
"bbox": [
69,
720,
450,
734
],
"score": 1.0,
"content": "2022; Saharia et al., 2022b) and 3D synthesis (Poole et al., 2022; Watson et al., 2022).",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5,
"bbox_fs": [
68,
636,
541,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
73,
51,
605,
121
],
"blocks": [
{
"type": "image_body",
"bbox": [
73,
51,
605,
121
],
"group_id": 0,
"lines": [
{
"bbox": [
73,
51,
605,
121
],
"spans": [
{
"bbox": [
73,
51,
605,
121
],
"score": 0.918,
"type": "image",
"image_path": "1eb960de54321d1c0eb739ad32a1b354b6f33e5c432a29f1dcfd2bd5ee9df629.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
73,
51,
605,
74.33333333333333
],
"spans": [],
"index": 0
},
{
"bbox": [
73,
74.33333333333333,
605,
97.66666666666666
],
"spans": [],
"index": 1
},
{
"bbox": [
73,
97.66666666666666,
605,
120.99999999999999
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
90,
126,
517,
137
],
"group_id": 0,
"lines": [
{
"bbox": [
95,
125,
519,
137
],
"spans": [
{
"bbox": [
95,
125,
519,
137
],
"score": 1.0,
"content": "A colorful professional animated logo for ’Imagen Video’ written using paint brush in cursive. Smooth animation.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
6,
146,
606,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
146,
606,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
6,
146,
606,
218
],
"spans": [
{
"bbox": [
6,
146,
606,
218
],
"score": 0.725,
"type": "image",
"image_path": "8b2f58e72050f92510ccaf16771c8d4531c5e2f54fd0a8c386e3926d4db218f7.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
6,
146,
606,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
6,
170.0,
606,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
6,
194.0,
606,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
181,
222,
434,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
180,
221,
434,
233
],
"spans": [
{
"bbox": [
180,
221,
434,
233
],
"score": 1.0,
"content": "Blue flame transforming into the text “Imagen”. Smooth animation",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
5,
241,
606,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
241,
606,
314
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
241,
606,
314
],
"spans": [
{
"bbox": [
5,
241,
606,
314
],
"score": 0.545,
"type": "image",
"image_path": "883adccdc4dbb074b9587f82f68592dfc4f2020a44d011f24553577566e66d42.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
5,
241,
606,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
5,
265.3333333333333,
606,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
5,
289.66666666666663,
606,
313.99999999999994
],
"spans": [],
"index": 10
}
]
}
],
"index": 9
},
{
"type": "text",
"bbox": [
216,
318,
398,
327
],
"lines": [
{
"bbox": [
216,
317,
398,
329
],
"spans": [
{
"bbox": [
216,
317,
398,
329
],
"score": 1.0,
"content": "Wooden figurine surfing on a surfboard in space.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "image",
"bbox": [
5,
336,
607,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
336,
607,
410
],
"group_id": 4,
"lines": [
{
"bbox": [
5,
336,
607,
410
],
"spans": [
{
"bbox": [
5,
336,
607,
410
],
"score": 0.49,
"type": "image",
"image_path": "ae128f4235e677ebff26bc35628c949da37817a1f991aad51a14a464f6f346b4.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
5,
336,
607,
360.6666666666667
],
"spans": [],
"index": 12
},
{
"bbox": [
5,
360.6666666666667,
607,
385.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
5,
385.33333333333337,
607,
410.00000000000006
],
"spans": [],
"index": 14
}
]
}
],
"index": 13
},
{
"type": "text",
"bbox": [
203,
414,
410,
423
],
"lines": [
{
"bbox": [
203,
413,
412,
424
],
"spans": [
{
"bbox": [
203,
413,
412,
424
],
"score": 1.0,
"content": "Balloon full of water exploding in extreme slow motion.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "image",
"bbox": [
5,
433,
606,
505
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
606,
505
],
"group_id": 5,
"lines": [
{
"bbox": [
5,
433,
606,
505
],
"spans": [
{
"bbox": [
5,
433,
606,
505
],
"score": 0.413,
"type": "image",
"image_path": "70375ded516a5c4193221cee18eb4528633968e19a9a72e007ce242db16e38e5.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
606,
457.0
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.0,
606,
481.0
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.0,
606,
505.0
],
"spans": [],
"index": 18
}
]
}
],
"index": 17
},
{
"type": "text",
"bbox": [
209,
510,
404,
519
],
"lines": [
{
"bbox": [
209,
509,
405,
520
],
"spans": [
{
"bbox": [
209,
509,
405,
520
],
"score": 1.0,
"content": "Melting pistachio ice cream dripping down the cone.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "image",
"bbox": [
6,
528,
608,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
528,
608,
601
],
"group_id": 6,
"lines": [
{
"bbox": [
6,
528,
608,
601
],
"spans": [
{
"bbox": [
6,
528,
608,
601
],
"score": 0.293,
"type": "image",
"image_path": "aa586e8c5fb6ed7653aafe39170a0524cfb04fa279348d10faa61221515b6097.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
6,
528,
608,
552.3333333333334
],
"spans": [],
"index": 20
},
{
"bbox": [
6,
552.3333333333334,
608,
576.6666666666667
],
"spans": [],
"index": 21
},
{
"bbox": [
6,
576.6666666666667,
608,
601.0000000000001
],
"spans": [],
"index": 22
}
]
}
],
"index": 21
},
{
"type": "text",
"bbox": [
230,
605,
385,
615
],
"lines": [
{
"bbox": [
229,
604,
385,
616
],
"spans": [
{
"bbox": [
229,
604,
385,
616
],
"score": 1.0,
"content": "A british shorthair jumping over a couch.",
"type": "text"
}
],
"index": 23
}
],
"index": 23
},
{
"type": "image",
"bbox": [
5,
624,
608,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
624,
608,
696
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
624,
608,
696
],
"spans": [
{
"bbox": [
5,
624,
608,
696
],
"score": 0.272,
"type": "image",
"image_path": "bff20f7b0156aa5827c480f65d371d77d14e8566a89e2766d38c35228533f2ea.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
5,
624,
608,
648.0
],
"spans": [],
"index": 24
},
{
"bbox": [
5,
648.0,
608,
672.0
],
"spans": [],
"index": 25
},
{
"bbox": [
5,
672.0,
608,
696.0
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
258,
701,
355,
711
],
"group_id": 2,
"lines": [
{
"bbox": [
257,
699,
357,
712
],
"spans": [
{
"bbox": [
257,
699,
357,
712
],
"score": 1.0,
"content": "Coffee pouring into a cup.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
68,
721,
538,
746
],
"lines": [
{
"bbox": [
69,
721,
540,
735
],
"spans": [
{
"bbox": [
69,
721,
540,
735
],
"score": 1.0,
"content": "Figure 2: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
71,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "2",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
73,
51,
605,
121
],
"blocks": [
{
"type": "image_body",
"bbox": [
73,
51,
605,
121
],
"group_id": 0,
"lines": [
{
"bbox": [
73,
51,
605,
121
],
"spans": [
{
"bbox": [
73,
51,
605,
121
],
"score": 0.918,
"type": "image",
"image_path": "1eb960de54321d1c0eb739ad32a1b354b6f33e5c432a29f1dcfd2bd5ee9df629.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
73,
51,
605,
74.33333333333333
],
"spans": [],
"index": 0
},
{
"bbox": [
73,
74.33333333333333,
605,
97.66666666666666
],
"spans": [],
"index": 1
},
{
"bbox": [
73,
97.66666666666666,
605,
120.99999999999999
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
90,
126,
517,
137
],
"group_id": 0,
"lines": [
{
"bbox": [
95,
125,
519,
137
],
"spans": [
{
"bbox": [
95,
125,
519,
137
],
"score": 1.0,
"content": "A colorful professional animated logo for ’Imagen Video’ written using paint brush in cursive. Smooth animation.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
6,
146,
606,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
146,
606,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
6,
146,
606,
218
],
"spans": [
{
"bbox": [
6,
146,
606,
218
],
"score": 0.725,
"type": "image",
"image_path": "8b2f58e72050f92510ccaf16771c8d4531c5e2f54fd0a8c386e3926d4db218f7.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
6,
146,
606,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
6,
170.0,
606,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
6,
194.0,
606,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
181,
222,
434,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
180,
221,
434,
233
],
"spans": [
{
"bbox": [
180,
221,
434,
233
],
"score": 1.0,
"content": "Blue flame transforming into the text “Imagen”. Smooth animation",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
5,
241,
606,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
241,
606,
314
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
241,
606,
314
],
"spans": [
{
"bbox": [
5,
241,
606,
314
],
"score": 0.545,
"type": "image",
"image_path": "883adccdc4dbb074b9587f82f68592dfc4f2020a44d011f24553577566e66d42.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
5,
241,
606,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
5,
265.3333333333333,
606,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
5,
289.66666666666663,
606,
313.99999999999994
],
"spans": [],
"index": 10
}
]
}
],
"index": 9
},
{
"type": "text",
"bbox": [
216,
318,
398,
327
],
"lines": [
{
"bbox": [
216,
317,
398,
329
],
"spans": [
{
"bbox": [
216,
317,
398,
329
],
"score": 1.0,
"content": "Wooden figurine surfing on a surfboard in space.",
"type": "text"
}
],
"index": 11
}
],
"index": 11,
"bbox_fs": [
216,
317,
398,
329
]
},
{
"type": "image",
"bbox": [
5,
336,
607,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
336,
607,
410
],
"group_id": 4,
"lines": [
{
"bbox": [
5,
336,
607,
410
],
"spans": [
{
"bbox": [
5,
336,
607,
410
],
"score": 0.49,
"type": "image",
"image_path": "ae128f4235e677ebff26bc35628c949da37817a1f991aad51a14a464f6f346b4.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
5,
336,
607,
360.6666666666667
],
"spans": [],
"index": 12
},
{
"bbox": [
5,
360.6666666666667,
607,
385.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
5,
385.33333333333337,
607,
410.00000000000006
],
"spans": [],
"index": 14
}
]
}
],
"index": 13
},
{
"type": "text",
"bbox": [
203,
414,
410,
423
],
"lines": [
{
"bbox": [
203,
413,
412,
424
],
"spans": [
{
"bbox": [
203,
413,
412,
424
],
"score": 1.0,
"content": "Balloon full of water exploding in extreme slow motion.",
"type": "text"
}
],
"index": 15
}
],
"index": 15,
"bbox_fs": [
203,
413,
412,
424
]
},
{
"type": "image",
"bbox": [
5,
433,
606,
505
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
606,
505
],
"group_id": 5,
"lines": [
{
"bbox": [
5,
433,
606,
505
],
"spans": [
{
"bbox": [
5,
433,
606,
505
],
"score": 0.413,
"type": "image",
"image_path": "70375ded516a5c4193221cee18eb4528633968e19a9a72e007ce242db16e38e5.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
606,
457.0
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.0,
606,
481.0
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.0,
606,
505.0
],
"spans": [],
"index": 18
}
]
}
],
"index": 17
},
{
"type": "text",
"bbox": [
209,
510,
404,
519
],
"lines": [
{
"bbox": [
209,
509,
405,
520
],
"spans": [
{
"bbox": [
209,
509,
405,
520
],
"score": 1.0,
"content": "Melting pistachio ice cream dripping down the cone.",
"type": "text"
}
],
"index": 19
}
],
"index": 19,
"bbox_fs": [
209,
509,
405,
520
]
},
{
"type": "image",
"bbox": [
6,
528,
608,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
528,
608,
601
],
"group_id": 6,
"lines": [
{
"bbox": [
6,
528,
608,
601
],
"spans": [
{
"bbox": [
6,
528,
608,
601
],
"score": 0.293,
"type": "image",
"image_path": "aa586e8c5fb6ed7653aafe39170a0524cfb04fa279348d10faa61221515b6097.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
6,
528,
608,
552.3333333333334
],
"spans": [],
"index": 20
},
{
"bbox": [
6,
552.3333333333334,
608,
576.6666666666667
],
"spans": [],
"index": 21
},
{
"bbox": [
6,
576.6666666666667,
608,
601.0000000000001
],
"spans": [],
"index": 22
}
]
}
],
"index": 21
},
{
"type": "text",
"bbox": [
230,
605,
385,
615
],
"lines": [
{
"bbox": [
229,
604,
385,
616
],
"spans": [
{
"bbox": [
229,
604,
385,
616
],
"score": 1.0,
"content": "A british shorthair jumping over a couch.",
"type": "text"
}
],
"index": 23
}
],
"index": 23,
"bbox_fs": [
229,
604,
385,
616
]
},
{
"type": "image",
"bbox": [
5,
624,
608,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
624,
608,
696
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
624,
608,
696
],
"spans": [
{
"bbox": [
5,
624,
608,
696
],
"score": 0.272,
"type": "image",
"image_path": "bff20f7b0156aa5827c480f65d371d77d14e8566a89e2766d38c35228533f2ea.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
5,
624,
608,
648.0
],
"spans": [],
"index": 24
},
{
"bbox": [
5,
648.0,
608,
672.0
],
"spans": [],
"index": 25
},
{
"bbox": [
5,
672.0,
608,
696.0
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
258,
701,
355,
711
],
"group_id": 2,
"lines": [
{
"bbox": [
257,
699,
357,
712
],
"spans": [
{
"bbox": [
257,
699,
357,
712
],
"score": 1.0,
"content": "Coffee pouring into a cup.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
68,
721,
538,
746
],
"lines": [
{
"bbox": [
69,
721,
540,
735
],
"spans": [
{
"bbox": [
69,
721,
540,
735
],
"score": 1.0,
"content": "Figure 2: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
69,
721,
540,
747
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
4,
51,
607,
122
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
51,
607,
122
],
"group_id": 0,
"lines": [
{
"bbox": [
4,
51,
607,
122
],
"spans": [
{
"bbox": [
4,
51,
607,
122
],
"score": 0.667,
"type": "image",
"image_path": "4d7e688166466d6c141b0ffdd85c547d4a2347c3358652df89c986300b16c69e.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
4,
51,
607,
74.66666666666667
],
"spans": [],
"index": 0
},
{
"bbox": [
4,
74.66666666666667,
607,
98.33333333333334
],
"spans": [],
"index": 1
},
{
"bbox": [
4,
98.33333333333334,
607,
122.00000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
205,
127,
408,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
204,
124,
410,
138
],
"spans": [
{
"bbox": [
204,
124,
410,
138
],
"score": 1.0,
"content": "A small hand-crafted wooden boat taking off to space.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
606,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
606,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
5,
146,
606,
218
],
"spans": [
{
"bbox": [
5,
146,
606,
218
],
"score": 0.353,
"type": "image",
"image_path": "6f662944f0642b3bad14ef4d2b7a028ffac4828ab102982b6abbcb67bce14d5b.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
606,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
606,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
606,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
240,
222,
376,
231
],
"group_id": 1,
"lines": [
{
"bbox": [
237,
221,
376,
232
],
"spans": [
{
"bbox": [
237,
221,
376,
232
],
"score": 1.0,
"content": "A person riding a bike in the sunset.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
5,
241,
607,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
241,
607,
314
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
241,
607,
314
],
"spans": [
{
"bbox": [
5,
241,
607,
314
],
"score": 0.351,
"type": "image",
"image_path": "e7915d1eed1b5078a741ab2a4e7fd4aa5e0d034439b24753a12a479bf2b181d6.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
5,
241,
607,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
5,
265.3333333333333,
607,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
5,
289.66666666666663,
607,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
204,
318,
410,
327
],
"group_id": 2,
"lines": [
{
"bbox": [
204,
317,
410,
328
],
"spans": [
{
"bbox": [
204,
317,
410,
328
],
"score": 1.0,
"content": "Drone flythrough interior of Sagrada Familia cathedral",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
6,
337,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
337,
606,
410
],
"group_id": 6,
"lines": [
{
"bbox": [
6,
337,
606,
410
],
"spans": [
{
"bbox": [
6,
337,
606,
410
],
"score": 0.537,
"type": "image",
"image_path": "bd441bd546a618fce6339bf087375abda89ac8f0ebaf7e7136d6682f3739361f.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
6,
337,
606,
361.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
6,
361.3333333333333,
606,
385.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
6,
385.66666666666663,
606,
409.99999999999994
],
"spans": [],
"index": 14
}
]
}
],
"index": 13
},
{
"type": "text",
"bbox": [
183,
413,
430,
423
],
"lines": [
{
"bbox": [
183,
412,
431,
424
],
"spans": [
{
"bbox": [
183,
412,
431,
424
],
"score": 1.0,
"content": "Wooden figurine walking on a treadmill made out of exercise mat.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "image",
"bbox": [
5,
433,
607,
506
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
607,
506
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
433,
607,
506
],
"spans": [
{
"bbox": [
5,
433,
607,
506
],
"score": 0.785,
"type": "image",
"image_path": "a501fe5f233094e4b84e1c1686fd3ca4f120dcde28ac2c91bb9df555009797f1.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
607,
457.3333333333333
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.3333333333333,
607,
481.66666666666663
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.66666666666663,
607,
505.99999999999994
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
84,
508,
531,
520
],
"group_id": 3,
"lines": [
{
"bbox": [
88,
507,
525,
520
],
"spans": [
{
"bbox": [
88,
507,
525,
520
],
"score": 1.0,
"content": "Origami dancers in white paper, 3D render, ultra-detailed, on white background, studio shot, dancing modern dance.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "image",
"bbox": [
7,
529,
606,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
7,
529,
606,
601
],
"group_id": 4,
"lines": [
{
"bbox": [
7,
529,
606,
601
],
"spans": [
{
"bbox": [
7,
529,
606,
601
],
"score": 0.554,
"type": "image",
"image_path": "de7ad70f69cca72ba1027beec906864801b5442478c1026f8e558fd8f832c360.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
7,
529,
606,
553.0
],
"spans": [],
"index": 20
},
{
"bbox": [
7,
553.0,
606,
577.0
],
"spans": [],
"index": 21
},
{
"bbox": [
7,
577.0,
606,
601.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
173,
605,
439,
615
],
"group_id": 4,
"lines": [
{
"bbox": [
174,
604,
440,
616
],
"spans": [
{
"bbox": [
174,
604,
440,
616
],
"score": 1.0,
"content": "Campfire at night in a snowy forest with starry sky in the background.",
"type": "text"
}
],
"index": 23
}
],
"index": 23
}
],
"index": 22.0
},
{
"type": "image",
"bbox": [
5,
625,
608,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
625,
608,
696
],
"group_id": 5,
"lines": [
{
"bbox": [
5,
625,
608,
696
],
"spans": [
{
"bbox": [
5,
625,
608,
696
],
"score": 0.342,
"type": "image",
"image_path": "190bd2b24d98ab38702d79260a77f51ecd29f3569f68e2defb117a11ddb8fcc0.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
5,
625,
608,
648.6666666666666
],
"spans": [],
"index": 24
},
{
"bbox": [
5,
648.6666666666666,
608,
672.3333333333333
],
"spans": [],
"index": 25
},
{
"bbox": [
5,
672.3333333333333,
608,
695.9999999999999
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
255,
701,
360,
710
],
"group_id": 5,
"lines": [
{
"bbox": [
254,
700,
361,
711
],
"spans": [
{
"bbox": [
254,
700,
361,
711
],
"score": 1.0,
"content": "An astronaut riding a horse.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
68,
721,
537,
746
],
"lines": [
{
"bbox": [
69,
721,
540,
735
],
"spans": [
{
"bbox": [
69,
721,
540,
735
],
"score": 1.0,
"content": "Figure 3: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
}
],
"page_idx": 2,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": [
4,
51,
607,
122
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
51,
607,
122
],
"group_id": 0,
"lines": [
{
"bbox": [
4,
51,
607,
122
],
"spans": [
{
"bbox": [
4,
51,
607,
122
],
"score": 0.667,
"type": "image",
"image_path": "4d7e688166466d6c141b0ffdd85c547d4a2347c3358652df89c986300b16c69e.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
4,
51,
607,
74.66666666666667
],
"spans": [],
"index": 0
},
{
"bbox": [
4,
74.66666666666667,
607,
98.33333333333334
],
"spans": [],
"index": 1
},
{
"bbox": [
4,
98.33333333333334,
607,
122.00000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
205,
127,
408,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
204,
124,
410,
138
],
"spans": [
{
"bbox": [
204,
124,
410,
138
],
"score": 1.0,
"content": "A small hand-crafted wooden boat taking off to space.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
606,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
606,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
5,
146,
606,
218
],
"spans": [
{
"bbox": [
5,
146,
606,
218
],
"score": 0.353,
"type": "image",
"image_path": "6f662944f0642b3bad14ef4d2b7a028ffac4828ab102982b6abbcb67bce14d5b.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
606,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
606,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
606,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
240,
222,
376,
231
],
"group_id": 1,
"lines": [
{
"bbox": [
237,
221,
376,
232
],
"spans": [
{
"bbox": [
237,
221,
376,
232
],
"score": 1.0,
"content": "A person riding a bike in the sunset.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
5,
241,
607,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
241,
607,
314
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
241,
607,
314
],
"spans": [
{
"bbox": [
5,
241,
607,
314
],
"score": 0.351,
"type": "image",
"image_path": "e7915d1eed1b5078a741ab2a4e7fd4aa5e0d034439b24753a12a479bf2b181d6.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
5,
241,
607,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
5,
265.3333333333333,
607,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
5,
289.66666666666663,
607,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
204,
318,
410,
327
],
"group_id": 2,
"lines": [
{
"bbox": [
204,
317,
410,
328
],
"spans": [
{
"bbox": [
204,
317,
410,
328
],
"score": 1.0,
"content": "Drone flythrough interior of Sagrada Familia cathedral",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
6,
337,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
6,
337,
606,
410
],
"group_id": 6,
"lines": [
{
"bbox": [
6,
337,
606,
410
],
"spans": [
{
"bbox": [
6,
337,
606,
410
],
"score": 0.537,
"type": "image",
"image_path": "bd441bd546a618fce6339bf087375abda89ac8f0ebaf7e7136d6682f3739361f.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
6,
337,
606,
361.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
6,
361.3333333333333,
606,
385.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
6,
385.66666666666663,
606,
409.99999999999994
],
"spans": [],
"index": 14
}
]
}
],
"index": 13
},
{
"type": "text",
"bbox": [
183,
413,
430,
423
],
"lines": [
{
"bbox": [
183,
412,
431,
424
],
"spans": [
{
"bbox": [
183,
412,
431,
424
],
"score": 1.0,
"content": "Wooden figurine walking on a treadmill made out of exercise mat.",
"type": "text"
}
],
"index": 15
}
],
"index": 15,
"bbox_fs": [
183,
412,
431,
424
]
},
{
"type": "image",
"bbox": [
5,
433,
607,
506
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
607,
506
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
433,
607,
506
],
"spans": [
{
"bbox": [
5,
433,
607,
506
],
"score": 0.785,
"type": "image",
"image_path": "a501fe5f233094e4b84e1c1686fd3ca4f120dcde28ac2c91bb9df555009797f1.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
607,
457.3333333333333
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.3333333333333,
607,
481.66666666666663
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.66666666666663,
607,
505.99999999999994
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
84,
508,
531,
520
],
"group_id": 3,
"lines": [
{
"bbox": [
88,
507,
525,
520
],
"spans": [
{
"bbox": [
88,
507,
525,
520
],
"score": 1.0,
"content": "Origami dancers in white paper, 3D render, ultra-detailed, on white background, studio shot, dancing modern dance.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "image",
"bbox": [
7,
529,
606,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
7,
529,
606,
601
],
"group_id": 4,
"lines": [
{
"bbox": [
7,
529,
606,
601
],
"spans": [
{
"bbox": [
7,
529,
606,
601
],
"score": 0.554,
"type": "image",
"image_path": "de7ad70f69cca72ba1027beec906864801b5442478c1026f8e558fd8f832c360.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
7,
529,
606,
553.0
],
"spans": [],
"index": 20
},
{
"bbox": [
7,
553.0,
606,
577.0
],
"spans": [],
"index": 21
},
{
"bbox": [
7,
577.0,
606,
601.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
173,
605,
439,
615
],
"group_id": 4,
"lines": [
{
"bbox": [
174,
604,
440,
616
],
"spans": [
{
"bbox": [
174,
604,
440,
616
],
"score": 1.0,
"content": "Campfire at night in a snowy forest with starry sky in the background.",
"type": "text"
}
],
"index": 23
}
],
"index": 23
}
],
"index": 22.0
},
{
"type": "image",
"bbox": [
5,
625,
608,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
625,
608,
696
],
"group_id": 5,
"lines": [
{
"bbox": [
5,
625,
608,
696
],
"spans": [
{
"bbox": [
5,
625,
608,
696
],
"score": 0.342,
"type": "image",
"image_path": "190bd2b24d98ab38702d79260a77f51ecd29f3569f68e2defb117a11ddb8fcc0.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
5,
625,
608,
648.6666666666666
],
"spans": [],
"index": 24
},
{
"bbox": [
5,
648.6666666666666,
608,
672.3333333333333
],
"spans": [],
"index": 25
},
{
"bbox": [
5,
672.3333333333333,
608,
695.9999999999999
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
255,
701,
360,
710
],
"group_id": 5,
"lines": [
{
"bbox": [
254,
700,
361,
711
],
"spans": [
{
"bbox": [
254,
700,
361,
711
],
"score": 1.0,
"content": "An astronaut riding a horse.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
68,
721,
537,
746
],
"lines": [
{
"bbox": [
69,
721,
540,
735
],
"spans": [
{
"bbox": [
69,
721,
540,
735
],
"score": 1.0,
"content": "Figure 3: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
69,
721,
540,
747
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
4,
51,
606,
123
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
51,
606,
123
],
"group_id": 0,
"lines": [
{
"bbox": [
4,
51,
606,
123
],
"spans": [
{
"bbox": [
4,
51,
606,
123
],
"score": 0.422,
"type": "image",
"image_path": "db86e191c3d0d8ac566432a1b352855849607a56afa2eb528793dd13ee33c57c.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
4,
51,
606,
75.0
],
"spans": [],
"index": 0
},
{
"bbox": [
4,
75.0,
606,
99.0
],
"spans": [],
"index": 1
},
{
"bbox": [
4,
99.0,
606,
123.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
235,
127,
379,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
234,
125,
379,
137
],
"spans": [
{
"bbox": [
234,
125,
379,
137
],
"score": 1.0,
"content": "A person riding a horse in the sunrise.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
607,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
607,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
5,
146,
607,
218
],
"spans": [
{
"bbox": [
5,
146,
607,
218
],
"score": 0.543,
"type": "image",
"image_path": "18eec6a916fd82c5ef51a71cea291b74714119896a22f76139f1290312139811.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
607,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
607,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
607,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
184,
222,
426,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
186,
221,
428,
233
],
"spans": [
{
"bbox": [
186,
221,
428,
233
],
"score": 1.0,
"content": "A happy elephant wearing a birthday hat walking under the sea.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
4,
241,
608,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
241,
608,
314
],
"group_id": 2,
"lines": [
{
"bbox": [
4,
241,
608,
314
],
"spans": [
{
"bbox": [
4,
241,
608,
314
],
"score": 0.903,
"type": "image",
"image_path": "a438281876e28778822daefc812fdca67163a7c82a34505ac5b1fd898d75c6e9.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
4,
241,
608,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
4,
265.3333333333333,
608,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
4,
289.66666666666663,
608,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
113,
317,
497,
327
],
"group_id": 2,
"lines": [
{
"bbox": [
116,
316,
497,
328
],
"spans": [
{
"bbox": [
116,
316,
497,
328
],
"score": 1.0,
"content": "Studio shot of minimal kinetic sculpture made from thin wire shaped like a bird on white background.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
27,
336,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
27,
336,
606,
410
],
"group_id": 3,
"lines": [
{
"bbox": [
27,
336,
606,
410
],
"spans": [
{
"bbox": [
27,
336,
606,
410
],
"score": 0.945,
"type": "image",
"image_path": "3383c92edaf0e151809f19912dc489314fc382f079a14a9b74b7d1a3743b2f39.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
27,
336,
606,
360.6666666666667
],
"spans": [],
"index": 12
},
{
"bbox": [
27,
360.6666666666667,
606,
385.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
27,
385.33333333333337,
606,
410.00000000000006
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
123,
413,
489,
424
],
"group_id": 3,
"lines": [
{
"bbox": [
125,
412,
489,
424
],
"spans": [
{
"bbox": [
125,
412,
489,
424
],
"score": 1.0,
"content": "A bunch of colorful candies falling into a tray in the shape of text ’Imagen Video’. Smooth video.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
}
],
"index": 14.0
},
{
"type": "image",
"bbox": [
2,
432,
609,
698
],
"blocks": [
{
"type": "image_body",
"bbox": [
2,
432,
609,
698
],
"group_id": 4,
"lines": [
{
"bbox": [
2,
432,
609,
698
],
"spans": [
{
"bbox": [
2,
432,
609,
698
],
"score": 0.415,
"type": "image",
"image_path": "b88e7ef8b5d9d21c925572bba93936c5e91a31e3ec645f28425896ea4f5e7570.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
2,
432,
609,
520.6666666666666
],
"spans": [],
"index": 16
},
{
"bbox": [
2,
520.6666666666666,
609,
609.3333333333333
],
"spans": [],
"index": 17
},
{
"bbox": [
2,
609.3333333333333,
609,
697.9999999999999
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
126,
700,
485,
711
],
"group_id": 4,
"lines": [
{
"bbox": [
128,
700,
486,
711
],
"spans": [
{
"bbox": [
128,
700,
486,
711
],
"score": 1.0,
"content": "Incredibly detailed science fiction scene set on an alien planet, view of a marketplace. Pixel art.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "text",
"bbox": [
68,
721,
539,
746
],
"lines": [
{
"bbox": [
69,
720,
540,
735
],
"spans": [
{
"bbox": [
69,
720,
540,
735
],
"score": 1.0,
"content": "Figure 4: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
71,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "4",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
4,
51,
606,
123
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
51,
606,
123
],
"group_id": 0,
"lines": [
{
"bbox": [
4,
51,
606,
123
],
"spans": [
{
"bbox": [
4,
51,
606,
123
],
"score": 0.422,
"type": "image",
"image_path": "db86e191c3d0d8ac566432a1b352855849607a56afa2eb528793dd13ee33c57c.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
4,
51,
606,
75.0
],
"spans": [],
"index": 0
},
{
"bbox": [
4,
75.0,
606,
99.0
],
"spans": [],
"index": 1
},
{
"bbox": [
4,
99.0,
606,
123.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
235,
127,
379,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
234,
125,
379,
137
],
"spans": [
{
"bbox": [
234,
125,
379,
137
],
"score": 1.0,
"content": "A person riding a horse in the sunrise.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
607,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
607,
218
],
"group_id": 1,
"lines": [
{
"bbox": [
5,
146,
607,
218
],
"spans": [
{
"bbox": [
5,
146,
607,
218
],
"score": 0.543,
"type": "image",
"image_path": "18eec6a916fd82c5ef51a71cea291b74714119896a22f76139f1290312139811.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
607,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
607,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
607,
218.0
],
"spans": [],
"index": 6
}
]
},
{
"type": "image_caption",
"bbox": [
184,
222,
426,
232
],
"group_id": 1,
"lines": [
{
"bbox": [
186,
221,
428,
233
],
"spans": [
{
"bbox": [
186,
221,
428,
233
],
"score": 1.0,
"content": "A happy elephant wearing a birthday hat walking under the sea.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
}
],
"index": 6.0
},
{
"type": "image",
"bbox": [
4,
241,
608,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
241,
608,
314
],
"group_id": 2,
"lines": [
{
"bbox": [
4,
241,
608,
314
],
"spans": [
{
"bbox": [
4,
241,
608,
314
],
"score": 0.903,
"type": "image",
"image_path": "a438281876e28778822daefc812fdca67163a7c82a34505ac5b1fd898d75c6e9.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
4,
241,
608,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
4,
265.3333333333333,
608,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
4,
289.66666666666663,
608,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
113,
317,
497,
327
],
"group_id": 2,
"lines": [
{
"bbox": [
116,
316,
497,
328
],
"spans": [
{
"bbox": [
116,
316,
497,
328
],
"score": 1.0,
"content": "Studio shot of minimal kinetic sculpture made from thin wire shaped like a bird on white background.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
27,
336,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
27,
336,
606,
410
],
"group_id": 3,
"lines": [
{
"bbox": [
27,
336,
606,
410
],
"spans": [
{
"bbox": [
27,
336,
606,
410
],
"score": 0.945,
"type": "image",
"image_path": "3383c92edaf0e151809f19912dc489314fc382f079a14a9b74b7d1a3743b2f39.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
27,
336,
606,
360.6666666666667
],
"spans": [],
"index": 12
},
{
"bbox": [
27,
360.6666666666667,
606,
385.33333333333337
],
"spans": [],
"index": 13
},
{
"bbox": [
27,
385.33333333333337,
606,
410.00000000000006
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
123,
413,
489,
424
],
"group_id": 3,
"lines": [
{
"bbox": [
125,
412,
489,
424
],
"spans": [
{
"bbox": [
125,
412,
489,
424
],
"score": 1.0,
"content": "A bunch of colorful candies falling into a tray in the shape of text ’Imagen Video’. Smooth video.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
}
],
"index": 14.0
},
{
"type": "image",
"bbox": [
2,
432,
609,
698
],
"blocks": [
{
"type": "image_body",
"bbox": [
2,
432,
609,
698
],
"group_id": 4,
"lines": [
{
"bbox": [
2,
432,
609,
698
],
"spans": [
{
"bbox": [
2,
432,
609,
698
],
"score": 0.415,
"type": "image",
"image_path": "b88e7ef8b5d9d21c925572bba93936c5e91a31e3ec645f28425896ea4f5e7570.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
2,
432,
609,
520.6666666666666
],
"spans": [],
"index": 16
},
{
"bbox": [
2,
520.6666666666666,
609,
609.3333333333333
],
"spans": [],
"index": 17
},
{
"bbox": [
2,
609.3333333333333,
609,
697.9999999999999
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
126,
700,
485,
711
],
"group_id": 4,
"lines": [
{
"bbox": [
128,
700,
486,
711
],
"spans": [
{
"bbox": [
128,
700,
486,
711
],
"score": 1.0,
"content": "Incredibly detailed science fiction scene set on an alien planet, view of a marketplace. Pixel art.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "text",
"bbox": [
68,
721,
539,
746
],
"lines": [
{
"bbox": [
69,
720,
540,
735
],
"spans": [
{
"bbox": [
69,
720,
540,
735
],
"score": 1.0,
"content": "Figure 4: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5,
"bbox_fs": [
69,
720,
540,
747
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
5,
51,
607,
123
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
51,
607,
123
],
"group_id": 0,
"lines": [
{
"bbox": [
5,
51,
607,
123
],
"spans": [
{
"bbox": [
5,
51,
607,
123
],
"score": 0.85,
"type": "image",
"image_path": "ca9d9cd34b1bd01ca5374fdceb29bf7158a693cbdf1a83e7994261de1dcd8d07.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
5,
51,
607,
75.0
],
"spans": [],
"index": 0
},
{
"bbox": [
5,
75.0,
607,
99.0
],
"spans": [],
"index": 1
},
{
"bbox": [
5,
99.0,
607,
123.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
138,
126,
475,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
137,
125,
477,
137
],
"spans": [
{
"bbox": [
137,
125,
477,
137
],
"score": 1.0,
"content": "A bunch of autumn leaves falling on a calm lake to form the text ’Imagen Video’. Smooth.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
608,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
608,
218
],
"group_id": 6,
"lines": [
{
"bbox": [
5,
146,
608,
218
],
"spans": [
{
"bbox": [
5,
146,
608,
218
],
"score": 0.886,
"type": "image",
"image_path": "ba9bda2d9f8d0aa6326ecf96e3f1d211e9fcb480080e986750b9e5b1617bf453.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
608,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
608,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
608,
218.0
],
"spans": [],
"index": 6
}
]
}
],
"index": 5
},
{
"type": "text",
"bbox": [
183,
222,
427,
232
],
"lines": [
{
"bbox": [
184,
221,
428,
233
],
"spans": [
{
"bbox": [
184,
221,
428,
233
],
"score": 1.0,
"content": "Pouring latte art into a silver cup with a golden spoon next to it.",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "image",
"bbox": [
3,
241,
607,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
3,
241,
607,
314
],
"group_id": 1,
"lines": [
{
"bbox": [
3,
241,
607,
314
],
"spans": [
{
"bbox": [
3,
241,
607,
314
],
"score": 0.512,
"type": "image",
"image_path": "686c7925afa30dc99a85f5511bccd4f983092f8dfce212f04f11bdedff4530da.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
3,
241,
607,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
3,
265.3333333333333,
607,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
3,
289.66666666666663,
607,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
277,
318,
336,
327
],
"group_id": 1,
"lines": [
{
"bbox": [
276,
316,
338,
329
],
"spans": [
{
"bbox": [
276,
316,
338,
329
],
"score": 1.0,
"content": "Shoveling snow.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
5,
337,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
337,
606,
410
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
337,
606,
410
],
"spans": [
{
"bbox": [
5,
337,
606,
410
],
"score": 0.502,
"type": "image",
"image_path": "47e6607f5ef4a202e1a6b5542e2a72e88f000e640c5afb5744dbe7286c6c9b33.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
5,
337,
606,
361.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
5,
361.3333333333333,
606,
385.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
5,
385.66666666666663,
606,
409.99999999999994
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
207,
414,
406,
423
],
"group_id": 2,
"lines": [
{
"bbox": [
207,
412,
407,
424
],
"spans": [
{
"bbox": [
207,
412,
407,
424
],
"score": 1.0,
"content": "Drone flythrough of a tropical jungle covered in snow",
"type": "text"
}
],
"index": 15
}
],
"index": 15
}
],
"index": 14.0
},
{
"type": "image",
"bbox": [
5,
433,
607,
505
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
607,
505
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
433,
607,
505
],
"spans": [
{
"bbox": [
5,
433,
607,
505
],
"score": 0.819,
"type": "image",
"image_path": "9865925f1d533f84b920f6910e84ca39a9b180d3a050f0e57ae20f6c2eed92e3.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
607,
457.0
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.0,
607,
481.0
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.0,
607,
505.0
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
137,
509,
474,
519
],
"group_id": 3,
"lines": [
{
"bbox": [
142,
508,
473,
520
],
"spans": [
{
"bbox": [
142,
508,
473,
520
],
"score": 1.0,
"content": "A beautiful sunrise on mars, Curiosity rover. High definition, timelapse, dramatic colors",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "image",
"bbox": [
4,
529,
608,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
529,
608,
601
],
"group_id": 4,
"lines": [
{
"bbox": [
4,
529,
608,
601
],
"spans": [
{
"bbox": [
4,
529,
608,
601
],
"score": 0.875,
"type": "image",
"image_path": "681a6e74ef7c23a49b08b685a9c128659f022ef24774d950978133d4d8642e53.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
4,
529,
608,
553.0
],
"spans": [],
"index": 20
},
{
"bbox": [
4,
553.0,
608,
577.0
],
"spans": [],
"index": 21
},
{
"bbox": [
4,
577.0,
608,
601.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
224,
605,
390,
615
],
"group_id": 4,
"lines": [
{
"bbox": [
223,
604,
391,
615
],
"spans": [
{
"bbox": [
223,
604,
391,
615
],
"score": 1.0,
"content": "A shark swimming in clear Carribean ocean.",
"type": "text"
}
],
"index": 23
}
],
"index": 23
}
],
"index": 22.0
},
{
"type": "image",
"bbox": [
3,
625,
607,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
3,
625,
607,
696
],
"group_id": 5,
"lines": [
{
"bbox": [
3,
625,
607,
696
],
"spans": [
{
"bbox": [
3,
625,
607,
696
],
"score": 0.931,
"type": "image",
"image_path": "88b40bd17207d00bcbec5d2f8f5d5de4c7955b0797a0b4bd8d0547a7ecfc9c59.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
3,
625,
607,
648.6666666666666
],
"spans": [],
"index": 24
},
{
"bbox": [
3,
648.6666666666666,
607,
672.3333333333333
],
"spans": [],
"index": 25
},
{
"bbox": [
3,
672.3333333333333,
607,
695.9999999999999
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
272,
700,
342,
710
],
"group_id": 5,
"lines": [
{
"bbox": [
271,
699,
344,
712
],
"spans": [
{
"bbox": [
271,
699,
344,
712
],
"score": 1.0,
"content": "A hand lifts a cup.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
69,
721,
538,
746
],
"lines": [
{
"bbox": [
68,
720,
541,
735
],
"spans": [
{
"bbox": [
68,
720,
541,
735
],
"score": 1.0,
"content": "Figure 5: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
71,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
751,
309,
762
],
"spans": [
{
"bbox": [
302,
751,
309,
762
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
5,
51,
607,
123
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
51,
607,
123
],
"group_id": 0,
"lines": [
{
"bbox": [
5,
51,
607,
123
],
"spans": [
{
"bbox": [
5,
51,
607,
123
],
"score": 0.85,
"type": "image",
"image_path": "ca9d9cd34b1bd01ca5374fdceb29bf7158a693cbdf1a83e7994261de1dcd8d07.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
5,
51,
607,
75.0
],
"spans": [],
"index": 0
},
{
"bbox": [
5,
75.0,
607,
99.0
],
"spans": [],
"index": 1
},
{
"bbox": [
5,
99.0,
607,
123.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
138,
126,
475,
136
],
"group_id": 0,
"lines": [
{
"bbox": [
137,
125,
477,
137
],
"spans": [
{
"bbox": [
137,
125,
477,
137
],
"score": 1.0,
"content": "A bunch of autumn leaves falling on a calm lake to form the text ’Imagen Video’. Smooth.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "image",
"bbox": [
5,
146,
608,
218
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
146,
608,
218
],
"group_id": 6,
"lines": [
{
"bbox": [
5,
146,
608,
218
],
"spans": [
{
"bbox": [
5,
146,
608,
218
],
"score": 0.886,
"type": "image",
"image_path": "ba9bda2d9f8d0aa6326ecf96e3f1d211e9fcb480080e986750b9e5b1617bf453.jpg"
}
]
}
],
"index": 5,
"virtual_lines": [
{
"bbox": [
5,
146,
608,
170.0
],
"spans": [],
"index": 4
},
{
"bbox": [
5,
170.0,
608,
194.0
],
"spans": [],
"index": 5
},
{
"bbox": [
5,
194.0,
608,
218.0
],
"spans": [],
"index": 6
}
]
}
],
"index": 5
},
{
"type": "text",
"bbox": [
183,
222,
427,
232
],
"lines": [
{
"bbox": [
184,
221,
428,
233
],
"spans": [
{
"bbox": [
184,
221,
428,
233
],
"score": 1.0,
"content": "Pouring latte art into a silver cup with a golden spoon next to it.",
"type": "text"
}
],
"index": 7
}
],
"index": 7,
"bbox_fs": [
184,
221,
428,
233
]
},
{
"type": "image",
"bbox": [
3,
241,
607,
314
],
"blocks": [
{
"type": "image_body",
"bbox": [
3,
241,
607,
314
],
"group_id": 1,
"lines": [
{
"bbox": [
3,
241,
607,
314
],
"spans": [
{
"bbox": [
3,
241,
607,
314
],
"score": 0.512,
"type": "image",
"image_path": "686c7925afa30dc99a85f5511bccd4f983092f8dfce212f04f11bdedff4530da.jpg"
}
]
}
],
"index": 9,
"virtual_lines": [
{
"bbox": [
3,
241,
607,
265.3333333333333
],
"spans": [],
"index": 8
},
{
"bbox": [
3,
265.3333333333333,
607,
289.66666666666663
],
"spans": [],
"index": 9
},
{
"bbox": [
3,
289.66666666666663,
607,
313.99999999999994
],
"spans": [],
"index": 10
}
]
},
{
"type": "image_caption",
"bbox": [
277,
318,
336,
327
],
"group_id": 1,
"lines": [
{
"bbox": [
276,
316,
338,
329
],
"spans": [
{
"bbox": [
276,
316,
338,
329
],
"score": 1.0,
"content": "Shoveling snow.",
"type": "text"
}
],
"index": 11
}
],
"index": 11
}
],
"index": 10.0
},
{
"type": "image",
"bbox": [
5,
337,
606,
410
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
337,
606,
410
],
"group_id": 2,
"lines": [
{
"bbox": [
5,
337,
606,
410
],
"spans": [
{
"bbox": [
5,
337,
606,
410
],
"score": 0.502,
"type": "image",
"image_path": "47e6607f5ef4a202e1a6b5542e2a72e88f000e640c5afb5744dbe7286c6c9b33.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
5,
337,
606,
361.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
5,
361.3333333333333,
606,
385.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
5,
385.66666666666663,
606,
409.99999999999994
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
207,
414,
406,
423
],
"group_id": 2,
"lines": [
{
"bbox": [
207,
412,
407,
424
],
"spans": [
{
"bbox": [
207,
412,
407,
424
],
"score": 1.0,
"content": "Drone flythrough of a tropical jungle covered in snow",
"type": "text"
}
],
"index": 15
}
],
"index": 15
}
],
"index": 14.0
},
{
"type": "image",
"bbox": [
5,
433,
607,
505
],
"blocks": [
{
"type": "image_body",
"bbox": [
5,
433,
607,
505
],
"group_id": 3,
"lines": [
{
"bbox": [
5,
433,
607,
505
],
"spans": [
{
"bbox": [
5,
433,
607,
505
],
"score": 0.819,
"type": "image",
"image_path": "9865925f1d533f84b920f6910e84ca39a9b180d3a050f0e57ae20f6c2eed92e3.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
5,
433,
607,
457.0
],
"spans": [],
"index": 16
},
{
"bbox": [
5,
457.0,
607,
481.0
],
"spans": [],
"index": 17
},
{
"bbox": [
5,
481.0,
607,
505.0
],
"spans": [],
"index": 18
}
]
},
{
"type": "image_caption",
"bbox": [
137,
509,
474,
519
],
"group_id": 3,
"lines": [
{
"bbox": [
142,
508,
473,
520
],
"spans": [
{
"bbox": [
142,
508,
473,
520
],
"score": 1.0,
"content": "A beautiful sunrise on mars, Curiosity rover. High definition, timelapse, dramatic colors",
"type": "text"
}
],
"index": 19
}
],
"index": 19
}
],
"index": 18.0
},
{
"type": "image",
"bbox": [
4,
529,
608,
601
],
"blocks": [
{
"type": "image_body",
"bbox": [
4,
529,
608,
601
],
"group_id": 4,
"lines": [
{
"bbox": [
4,
529,
608,
601
],
"spans": [
{
"bbox": [
4,
529,
608,
601
],
"score": 0.875,
"type": "image",
"image_path": "681a6e74ef7c23a49b08b685a9c128659f022ef24774d950978133d4d8642e53.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
4,
529,
608,
553.0
],
"spans": [],
"index": 20
},
{
"bbox": [
4,
553.0,
608,
577.0
],
"spans": [],
"index": 21
},
{
"bbox": [
4,
577.0,
608,
601.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
224,
605,
390,
615
],
"group_id": 4,
"lines": [
{
"bbox": [
223,
604,
391,
615
],
"spans": [
{
"bbox": [
223,
604,
391,
615
],
"score": 1.0,
"content": "A shark swimming in clear Carribean ocean.",
"type": "text"
}
],
"index": 23
}
],
"index": 23
}
],
"index": 22.0
},
{
"type": "image",
"bbox": [
3,
625,
607,
696
],
"blocks": [
{
"type": "image_body",
"bbox": [
3,
625,
607,
696
],
"group_id": 5,
"lines": [
{
"bbox": [
3,
625,
607,
696
],
"spans": [
{
"bbox": [
3,
625,
607,
696
],
"score": 0.931,
"type": "image",
"image_path": "88b40bd17207d00bcbec5d2f8f5d5de4c7955b0797a0b4bd8d0547a7ecfc9c59.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
3,
625,
607,
648.6666666666666
],
"spans": [],
"index": 24
},
{
"bbox": [
3,
648.6666666666666,
607,
672.3333333333333
],
"spans": [],
"index": 25
},
{
"bbox": [
3,
672.3333333333333,
607,
695.9999999999999
],
"spans": [],
"index": 26
}
]
},
{
"type": "image_caption",
"bbox": [
272,
700,
342,
710
],
"group_id": 5,
"lines": [
{
"bbox": [
271,
699,
344,
712
],
"spans": [
{
"bbox": [
271,
699,
344,
712
],
"score": 1.0,
"content": "A hand lifts a cup.",
"type": "text"
}
],
"index": 27
}
],
"index": 27
}
],
"index": 26.0
},
{
"type": "text",
"bbox": [
69,
721,
538,
746
],
"lines": [
{
"bbox": [
68,
720,
541,
735
],
"spans": [
{
"bbox": [
68,
720,
541,
735
],
"score": 1.0,
"content": "Figure 5: Videos generated from various text prompts. Imagen Video produces diverse and temporally-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
732,
337,
747
],
"spans": [
{
"bbox": [
69,
732,
337,
747
],
"score": 1.0,
"content": "coherent videos that are well-aligned with the given prompt.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
68,
720,
541,
747
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
201
],
"lines": [
{
"bbox": [
70,
83,
541,
95
],
"spans": [
{
"bbox": [
70,
83,
541,
95
],
"score": 1.0,
"content": "Our work aims to generate videos from text. Prior work on video generation has focused on more restricted",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
69,
94,
542,
108
],
"spans": [
{
"bbox": [
69,
94,
542,
108
],
"score": 1.0,
"content": "datasets with autoregressive models (Ranzato et al., 2014; Shi et al., 2015; Finn et al., 2016; Kalchbrenner",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
68,
105,
541,
119
],
"spans": [
{
"bbox": [
68,
105,
541,
119
],
"score": 1.0,
"content": "et al., 2017; Babaeizadeh et al., 2021), latent-variable models with autoregressive priors (Mathieu et al., 2016;",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
69,
117,
542,
132
],
"spans": [
{
"bbox": [
69,
117,
542,
132
],
"score": 1.0,
"content": "Vondrick et al., 2016; Babaeizadeh et al., 2018; Kumar et al., 2020), and more recently non-autoregressive",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
69,
129,
541,
144
],
"spans": [
{
"bbox": [
69,
129,
541,
144
],
"score": 1.0,
"content": "latent-variable approaches (Gupta et al., 2022). Diffusion models have also shown promise for video gener-",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
141,
542,
155
],
"spans": [
{
"bbox": [
69,
141,
542,
155
],
"score": 1.0,
"content": "ation (Ho et al., 2022b) at moderate resolution. Yang et al. (2022) showed autoregressive generation with",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
154,
541,
168
],
"spans": [
{
"bbox": [
69,
154,
541,
168
],
"score": 1.0,
"content": "a RNN-based model with conditional diffusion observations. The concurrent work of Singer et al. (2022)",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
167,
540,
178
],
"spans": [
{
"bbox": [
70,
167,
540,
178
],
"score": 1.0,
"content": "also applied text-to-video modelling with diffusion models, but built on a pretrained text-to-image model.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
177,
542,
192
],
"spans": [
{
"bbox": [
69,
177,
542,
192
],
"score": 1.0,
"content": "Harvey et al. (2022) generates videos up to 25 minutes in length with video diffusion models, however the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
70,
190,
162,
202
],
"spans": [
{
"bbox": [
70,
190,
162,
202
],
"score": 1.0,
"content": "domain is restricted.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5
},
{
"type": "text",
"bbox": [
71,
207,
540,
291
],
"lines": [
{
"bbox": [
69,
208,
541,
220
],
"spans": [
{
"bbox": [
69,
208,
541,
220
],
"score": 1.0,
"content": "In this work, we introduce Imagen Video, a text-to-video generation system based on video diffusion models",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
220,
541,
233
],
"spans": [
{
"bbox": [
70,
220,
541,
233
],
"score": 1.0,
"content": "(Ho et al., 2022b) that is capable of generating high definition videos with high frame fidelity, strong temporal",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
231,
541,
244
],
"spans": [
{
"bbox": [
69,
231,
516,
244
],
"score": 1.0,
"content": "consistency, and deep language understanding. Imagen Video scales from prior work of 64-frame 128",
"type": "text"
},
{
"bbox": [
516,
235,
524,
242
],
"score": 0.62,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
525,
231,
541,
244
],
"score": 1.0,
"content": "128",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
70,
243,
541,
257
],
"spans": [
{
"bbox": [
70,
243,
281,
257
],
"score": 1.0,
"content": "videos at 24 frames per second to 128 frame 1280",
"type": "text"
},
{
"bbox": [
282,
247,
290,
254
],
"score": 0.71,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
290,
243,
541,
257
],
"score": 1.0,
"content": "768 high-definition video at 24 frames per second. Imagen",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
254,
542,
269
],
"spans": [
{
"bbox": [
69,
254,
542,
269
],
"score": 1.0,
"content": "Video has a simple architecture: The model consists of a frozen T5 text encoder (Raffel et al., 2020), a",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
267,
540,
281
],
"spans": [
{
"bbox": [
69,
267,
540,
281
],
"score": 1.0,
"content": "base video diffusion model, and interleaved spatial and temporal super-resolution diffusion models. Our key",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
70,
280,
195,
292
],
"spans": [
{
"bbox": [
70,
280,
195,
292
],
"score": 1.0,
"content": "contributions are as follows:",
"type": "text"
}
],
"index": 16
}
],
"index": 13
},
{
"type": "text",
"bbox": [
91,
306,
542,
452
],
"lines": [
{
"bbox": [
92,
304,
541,
320
],
"spans": [
{
"bbox": [
92,
304,
541,
320
],
"score": 1.0,
"content": "1. We demonstrate the simplicity and effectiveness of cascaded diffusion video models for high definition",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
319,
183,
330
],
"spans": [
{
"bbox": [
106,
319,
183,
330
],
"score": 1.0,
"content": "video generation.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
92,
337,
541,
351
],
"spans": [
{
"bbox": [
92,
337,
541,
351
],
"score": 1.0,
"content": "2. We confirm that recent findings in the text-to-image setting transfer to video generation, such as",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
350,
457,
363
],
"spans": [
{
"bbox": [
105,
350,
457,
363
],
"score": 1.0,
"content": "the effectiveness of frozen encoder text conditioning and classifier-free guidance.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
92,
370,
541,
384
],
"spans": [
{
"bbox": [
92,
370,
541,
384
],
"score": 1.0,
"content": "3. We show new findings for video diffusion models that have implications for diffusion models in",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
383,
541,
396
],
"spans": [
{
"bbox": [
105,
383,
541,
396
],
"score": 1.0,
"content": "general, such as the effectiveness of the v-prediction parameterization for sample quality and the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
395,
540,
408
],
"spans": [
{
"bbox": [
106,
395,
540,
408
],
"score": 1.0,
"content": "effectiveness of progressive distillation of guided diffusion models for the text-conditioned video",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
408,
189,
420
],
"spans": [
{
"bbox": [
105,
408,
189,
420
],
"score": 1.0,
"content": "generation setting.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
91,
426,
540,
442
],
"spans": [
{
"bbox": [
91,
426,
540,
442
],
"score": 1.0,
"content": "4. We demonstrate qualitative controllability in Imagen Video, such as 3D object understanding, gen-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
439,
444,
452
],
"spans": [
{
"bbox": [
105,
439,
444,
452
],
"score": 1.0,
"content": "eration of text animations, and generation of videos in various artistic styles.",
"type": "text"
}
],
"index": 26
}
],
"index": 21.5
},
{
"type": "title",
"bbox": [
71,
466,
168,
480
],
"lines": [
{
"bbox": [
68,
464,
169,
484
],
"spans": [
{
"bbox": [
68,
464,
169,
484
],
"score": 1.0,
"content": "2 Imagen Video",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
71,
492,
540,
552
],
"lines": [
{
"bbox": [
70,
493,
541,
505
],
"spans": [
{
"bbox": [
70,
493,
541,
505
],
"score": 1.0,
"content": "Our model, Imagen Video, is a cascade of video diffusion models (Ho et al., 2022a;b). It consists of 7",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
504,
541,
518
],
"spans": [
{
"bbox": [
69,
504,
541,
518
],
"score": 1.0,
"content": "sub-models which perform text-conditional video generation, spatial super-resolution, and temporal super-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
70,
516,
541,
530
],
"spans": [
{
"bbox": [
70,
516,
436,
530
],
"score": 1.0,
"content": "resolution. With the entire cascade, Imagen Video generates high definition 1280",
"type": "text"
},
{
"bbox": [
436,
520,
444,
527
],
"score": 0.5,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
444,
516,
497,
530
],
"score": 1.0,
"content": "768 (width",
"type": "text"
},
{
"bbox": [
497,
520,
504,
527
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
505,
516,
541,
530
],
"score": 1.0,
"content": "height)",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
70,
529,
541,
541
],
"spans": [
{
"bbox": [
70,
529,
291,
541
],
"score": 1.0,
"content": "videos at 24 frames per second, for 128 frames (",
"type": "text"
},
{
"bbox": [
292,
532,
318,
538
],
"score": 0.58,
"content": "\\approx 5 . 3",
"type": "inline_equation"
},
{
"bbox": [
318,
529,
541,
541
],
"score": 1.0,
"content": "seconds)—approximately 126 million pixels. We",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
70,
541,
490,
554
],
"spans": [
{
"bbox": [
70,
541,
490,
554
],
"score": 1.0,
"content": "describe the components and techniques that constitute Imagen Video in the following sections.",
"type": "text"
}
],
"index": 32
}
],
"index": 30
},
{
"type": "title",
"bbox": [
72,
566,
173,
578
],
"lines": [
{
"bbox": [
69,
564,
174,
580
],
"spans": [
{
"bbox": [
69,
564,
174,
580
],
"score": 1.0,
"content": "2.1 Diffusion models",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "text",
"bbox": [
71,
587,
540,
648
],
"lines": [
{
"bbox": [
69,
587,
541,
601
],
"spans": [
{
"bbox": [
69,
587,
541,
601
],
"score": 1.0,
"content": "Imagen Video is built from diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2019; Ho et al.,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
599,
542,
613
],
"spans": [
{
"bbox": [
69,
599,
250,
613
],
"score": 1.0,
"content": "2020) specified in continuous time (Tzen",
"type": "text"
},
{
"bbox": [
250,
602,
258,
610
],
"score": 0.34,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
258,
599,
542,
613
],
"score": 1.0,
"content": "Raginsky, 2019; Song et al., 2021; Kingma et al., 2021). We use",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
610,
540,
626
],
"spans": [
{
"bbox": [
69,
610,
460,
626
],
"score": 1.0,
"content": "the formulation of Kingma et al. (2021): the model is a latent variable model with latents",
"type": "text"
},
{
"bbox": [
460,
614,
540,
624
],
"score": 0.93,
"content": "\\mathbf { z } = \\{ \\mathbf { z } _ { t } | t \\in [ 0 , 1 ] \\}",
"type": "inline_equation"
}
],
"index": 36
},
{
"bbox": [
69,
623,
541,
637
],
"spans": [
{
"bbox": [
69,
623,
191,
637
],
"score": 1.0,
"content": "following a forward process",
"type": "text"
},
{
"bbox": [
191,
626,
218,
636
],
"score": 0.94,
"content": "q ( \\mathbf { z } | \\mathbf { x } )",
"type": "inline_equation"
},
{
"bbox": [
218,
623,
291,
637
],
"score": 1.0,
"content": "starting at data",
"type": "text"
},
{
"bbox": [
291,
626,
329,
636
],
"score": 0.94,
"content": "\\mathbf { x } \\sim p ( \\mathbf { x } )",
"type": "inline_equation"
},
{
"bbox": [
330,
623,
541,
637
],
"score": 1.0,
"content": ". The forward process is a Gaussian process that",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
636,
218,
649
],
"spans": [
{
"bbox": [
69,
636,
218,
649
],
"score": 1.0,
"content": "satisfies the Markovian structure:",
"type": "text"
}
],
"index": 38
}
],
"index": 36
},
{
"type": "interline_equation",
"bbox": [
172,
657,
440,
672
],
"lines": [
{
"bbox": [
172,
657,
440,
672
],
"spans": [
{
"bbox": [
172,
657,
440,
672
],
"score": 0.86,
"content": "\\begin{array} { r } { q ( \\mathbf { z } _ { t } | \\mathbf { x } ) = \\mathcal { N } ( \\mathbf { z } _ { t } ; \\alpha _ { t } \\mathbf { x } , \\sigma _ { t } ^ { 2 } \\mathbf { I } ) , \\quad q ( \\mathbf { z } _ { t } | \\mathbf { z } _ { s } ) = \\mathcal { N } ( \\mathbf { z } _ { t } ; ( \\alpha _ { t } / \\alpha _ { s } ) \\mathbf { z } _ { s } , \\sigma _ { t | s } ^ { 2 } \\mathbf { I } ) } \\end{array}",
"type": "interline_equation",
"image_path": "91e32dd5f663fe175f4fe9874f7db7b0f12fcefcf05a707499a151a1035386b8.jpg"
}
]
}
],
"index": 39,
"virtual_lines": [
{
"bbox": [
172,
657,
440,
672
],
"spans": [],
"index": 39
}
]
},
{
"type": "text",
"bbox": [
70,
681,
540,
732
],
"lines": [
{
"bbox": [
68,
680,
543,
697
],
"spans": [
{
"bbox": [
68,
680,
99,
696
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
100,
685,
158,
694
],
"score": 0.9,
"content": "0 \\leq s < t \\leq 1",
"type": "inline_equation"
},
{
"bbox": [
158,
680,
163,
696
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
163,
684,
254,
697
],
"score": 0.94,
"content": "\\sigma _ { t | s } ^ { 2 } = ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\sigma _ { t } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
254,
680,
278,
696
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
278,
687,
302,
694
],
"score": 0.91,
"content": "\\alpha _ { t } , \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
302,
680,
543,
696
],
"score": 1.0,
"content": "specify a noise schedule whose log signal-to-noise-ratio",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
71,
695,
542,
711
],
"spans": [
{
"bbox": [
71,
698,
140,
709
],
"score": 0.94,
"content": "\\lambda _ { t } = \\log [ \\alpha _ { t } ^ { 2 } / \\sigma _ { t } ^ { 2 } ]",
"type": "inline_equation"
},
{
"bbox": [
140,
695,
273,
711
],
"score": 1.0,
"content": "decreases monotonically with",
"type": "text"
},
{
"bbox": [
274,
700,
278,
706
],
"score": 0.87,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
695,
305,
711
],
"score": 1.0,
"content": "until",
"type": "text"
},
{
"bbox": [
305,
698,
373,
709
],
"score": 0.94,
"content": "q ( \\mathbf { z } _ { 1 } ) \\approx \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { I } )",
"type": "inline_equation"
},
{
"bbox": [
374,
695,
542,
711
],
"score": 1.0,
"content": ". We use a continuous time version of",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
708,
541,
722
],
"spans": [
{
"bbox": [
70,
708,
541,
722
],
"score": 1.0,
"content": "the cosine noise schedule (Nichol & Dhariwal, 2021). The generative model is a learned model that matches",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
720,
528,
734
],
"spans": [
{
"bbox": [
69,
720,
339,
734
],
"score": 1.0,
"content": "this forward process in the reverse time direction, generating",
"type": "text"
},
{
"bbox": [
340,
725,
348,
732
],
"score": 0.89,
"content": "\\mathbf { z } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
349,
720,
412,
734
],
"score": 1.0,
"content": "starting from",
"type": "text"
},
{
"bbox": [
412,
723,
434,
730
],
"score": 0.92,
"content": "t = 1",
"type": "inline_equation"
},
{
"bbox": [
435,
720,
501,
734
],
"score": 1.0,
"content": "and ending at",
"type": "text"
},
{
"bbox": [
501,
723,
523,
730
],
"score": 0.92,
"content": "t = 0",
"type": "inline_equation"
},
{
"bbox": [
524,
720,
528,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 43
}
],
"index": 41.5
}
],
"page_idx": 5,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
70,
25,
239,
38
],
"spans": [
{
"bbox": [
70,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "6",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
201
],
"lines": [
{
"bbox": [
70,
83,
541,
95
],
"spans": [
{
"bbox": [
70,
83,
541,
95
],
"score": 1.0,
"content": "Our work aims to generate videos from text. Prior work on video generation has focused on more restricted",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
69,
94,
542,
108
],
"spans": [
{
"bbox": [
69,
94,
542,
108
],
"score": 1.0,
"content": "datasets with autoregressive models (Ranzato et al., 2014; Shi et al., 2015; Finn et al., 2016; Kalchbrenner",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
68,
105,
541,
119
],
"spans": [
{
"bbox": [
68,
105,
541,
119
],
"score": 1.0,
"content": "et al., 2017; Babaeizadeh et al., 2021), latent-variable models with autoregressive priors (Mathieu et al., 2016;",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
69,
117,
542,
132
],
"spans": [
{
"bbox": [
69,
117,
542,
132
],
"score": 1.0,
"content": "Vondrick et al., 2016; Babaeizadeh et al., 2018; Kumar et al., 2020), and more recently non-autoregressive",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
69,
129,
541,
144
],
"spans": [
{
"bbox": [
69,
129,
541,
144
],
"score": 1.0,
"content": "latent-variable approaches (Gupta et al., 2022). Diffusion models have also shown promise for video gener-",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
141,
542,
155
],
"spans": [
{
"bbox": [
69,
141,
542,
155
],
"score": 1.0,
"content": "ation (Ho et al., 2022b) at moderate resolution. Yang et al. (2022) showed autoregressive generation with",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
154,
541,
168
],
"spans": [
{
"bbox": [
69,
154,
541,
168
],
"score": 1.0,
"content": "a RNN-based model with conditional diffusion observations. The concurrent work of Singer et al. (2022)",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
167,
540,
178
],
"spans": [
{
"bbox": [
70,
167,
540,
178
],
"score": 1.0,
"content": "also applied text-to-video modelling with diffusion models, but built on a pretrained text-to-image model.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
177,
542,
192
],
"spans": [
{
"bbox": [
69,
177,
542,
192
],
"score": 1.0,
"content": "Harvey et al. (2022) generates videos up to 25 minutes in length with video diffusion models, however the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
70,
190,
162,
202
],
"spans": [
{
"bbox": [
70,
190,
162,
202
],
"score": 1.0,
"content": "domain is restricted.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5,
"bbox_fs": [
68,
83,
542,
202
]
},
{
"type": "text",
"bbox": [
71,
207,
540,
291
],
"lines": [
{
"bbox": [
69,
208,
541,
220
],
"spans": [
{
"bbox": [
69,
208,
541,
220
],
"score": 1.0,
"content": "In this work, we introduce Imagen Video, a text-to-video generation system based on video diffusion models",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
220,
541,
233
],
"spans": [
{
"bbox": [
70,
220,
541,
233
],
"score": 1.0,
"content": "(Ho et al., 2022b) that is capable of generating high definition videos with high frame fidelity, strong temporal",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
231,
541,
244
],
"spans": [
{
"bbox": [
69,
231,
516,
244
],
"score": 1.0,
"content": "consistency, and deep language understanding. Imagen Video scales from prior work of 64-frame 128",
"type": "text"
},
{
"bbox": [
516,
235,
524,
242
],
"score": 0.62,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
525,
231,
541,
244
],
"score": 1.0,
"content": "128",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
70,
243,
541,
257
],
"spans": [
{
"bbox": [
70,
243,
281,
257
],
"score": 1.0,
"content": "videos at 24 frames per second to 128 frame 1280",
"type": "text"
},
{
"bbox": [
282,
247,
290,
254
],
"score": 0.71,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
290,
243,
541,
257
],
"score": 1.0,
"content": "768 high-definition video at 24 frames per second. Imagen",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
254,
542,
269
],
"spans": [
{
"bbox": [
69,
254,
542,
269
],
"score": 1.0,
"content": "Video has a simple architecture: The model consists of a frozen T5 text encoder (Raffel et al., 2020), a",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
267,
540,
281
],
"spans": [
{
"bbox": [
69,
267,
540,
281
],
"score": 1.0,
"content": "base video diffusion model, and interleaved spatial and temporal super-resolution diffusion models. Our key",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
70,
280,
195,
292
],
"spans": [
{
"bbox": [
70,
280,
195,
292
],
"score": 1.0,
"content": "contributions are as follows:",
"type": "text"
}
],
"index": 16
}
],
"index": 13,
"bbox_fs": [
69,
208,
542,
292
]
},
{
"type": "list",
"bbox": [
91,
306,
542,
452
],
"lines": [
{
"bbox": [
92,
304,
541,
320
],
"spans": [
{
"bbox": [
92,
304,
541,
320
],
"score": 1.0,
"content": "1. We demonstrate the simplicity and effectiveness of cascaded diffusion video models for high definition",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
106,
319,
183,
330
],
"spans": [
{
"bbox": [
106,
319,
183,
330
],
"score": 1.0,
"content": "video generation.",
"type": "text"
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
92,
337,
541,
351
],
"spans": [
{
"bbox": [
92,
337,
541,
351
],
"score": 1.0,
"content": "2. We confirm that recent findings in the text-to-image setting transfer to video generation, such as",
"type": "text"
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
105,
350,
457,
363
],
"spans": [
{
"bbox": [
105,
350,
457,
363
],
"score": 1.0,
"content": "the effectiveness of frozen encoder text conditioning and classifier-free guidance.",
"type": "text"
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
92,
370,
541,
384
],
"spans": [
{
"bbox": [
92,
370,
541,
384
],
"score": 1.0,
"content": "3. We show new findings for video diffusion models that have implications for diffusion models in",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true
},
{
"bbox": [
105,
383,
541,
396
],
"spans": [
{
"bbox": [
105,
383,
541,
396
],
"score": 1.0,
"content": "general, such as the effectiveness of the v-prediction parameterization for sample quality and the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
395,
540,
408
],
"spans": [
{
"bbox": [
106,
395,
540,
408
],
"score": 1.0,
"content": "effectiveness of progressive distillation of guided diffusion models for the text-conditioned video",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
408,
189,
420
],
"spans": [
{
"bbox": [
105,
408,
189,
420
],
"score": 1.0,
"content": "generation setting.",
"type": "text"
}
],
"index": 24,
"is_list_end_line": true
},
{
"bbox": [
91,
426,
540,
442
],
"spans": [
{
"bbox": [
91,
426,
540,
442
],
"score": 1.0,
"content": "4. We demonstrate qualitative controllability in Imagen Video, such as 3D object understanding, gen-",
"type": "text"
}
],
"index": 25,
"is_list_start_line": true
},
{
"bbox": [
105,
439,
444,
452
],
"spans": [
{
"bbox": [
105,
439,
444,
452
],
"score": 1.0,
"content": "eration of text animations, and generation of videos in various artistic styles.",
"type": "text"
}
],
"index": 26,
"is_list_end_line": true
}
],
"index": 21.5,
"bbox_fs": [
91,
304,
541,
452
]
},
{
"type": "title",
"bbox": [
71,
466,
168,
480
],
"lines": [
{
"bbox": [
68,
464,
169,
484
],
"spans": [
{
"bbox": [
68,
464,
169,
484
],
"score": 1.0,
"content": "2 Imagen Video",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
71,
492,
540,
552
],
"lines": [
{
"bbox": [
70,
493,
541,
505
],
"spans": [
{
"bbox": [
70,
493,
541,
505
],
"score": 1.0,
"content": "Our model, Imagen Video, is a cascade of video diffusion models (Ho et al., 2022a;b). It consists of 7",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
504,
541,
518
],
"spans": [
{
"bbox": [
69,
504,
541,
518
],
"score": 1.0,
"content": "sub-models which perform text-conditional video generation, spatial super-resolution, and temporal super-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
70,
516,
541,
530
],
"spans": [
{
"bbox": [
70,
516,
436,
530
],
"score": 1.0,
"content": "resolution. With the entire cascade, Imagen Video generates high definition 1280",
"type": "text"
},
{
"bbox": [
436,
520,
444,
527
],
"score": 0.5,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
444,
516,
497,
530
],
"score": 1.0,
"content": "768 (width",
"type": "text"
},
{
"bbox": [
497,
520,
504,
527
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
505,
516,
541,
530
],
"score": 1.0,
"content": "height)",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
70,
529,
541,
541
],
"spans": [
{
"bbox": [
70,
529,
291,
541
],
"score": 1.0,
"content": "videos at 24 frames per second, for 128 frames (",
"type": "text"
},
{
"bbox": [
292,
532,
318,
538
],
"score": 0.58,
"content": "\\approx 5 . 3",
"type": "inline_equation"
},
{
"bbox": [
318,
529,
541,
541
],
"score": 1.0,
"content": "seconds)—approximately 126 million pixels. We",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
70,
541,
490,
554
],
"spans": [
{
"bbox": [
70,
541,
490,
554
],
"score": 1.0,
"content": "describe the components and techniques that constitute Imagen Video in the following sections.",
"type": "text"
}
],
"index": 32
}
],
"index": 30,
"bbox_fs": [
69,
493,
541,
554
]
},
{
"type": "title",
"bbox": [
72,
566,
173,
578
],
"lines": [
{
"bbox": [
69,
564,
174,
580
],
"spans": [
{
"bbox": [
69,
564,
174,
580
],
"score": 1.0,
"content": "2.1 Diffusion models",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "text",
"bbox": [
71,
587,
540,
648
],
"lines": [
{
"bbox": [
69,
587,
541,
601
],
"spans": [
{
"bbox": [
69,
587,
541,
601
],
"score": 1.0,
"content": "Imagen Video is built from diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2019; Ho et al.,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
599,
542,
613
],
"spans": [
{
"bbox": [
69,
599,
250,
613
],
"score": 1.0,
"content": "2020) specified in continuous time (Tzen",
"type": "text"
},
{
"bbox": [
250,
602,
258,
610
],
"score": 0.34,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
258,
599,
542,
613
],
"score": 1.0,
"content": "Raginsky, 2019; Song et al., 2021; Kingma et al., 2021). We use",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
610,
540,
626
],
"spans": [
{
"bbox": [
69,
610,
460,
626
],
"score": 1.0,
"content": "the formulation of Kingma et al. (2021): the model is a latent variable model with latents",
"type": "text"
},
{
"bbox": [
460,
614,
540,
624
],
"score": 0.93,
"content": "\\mathbf { z } = \\{ \\mathbf { z } _ { t } | t \\in [ 0 , 1 ] \\}",
"type": "inline_equation"
}
],
"index": 36
},
{
"bbox": [
69,
623,
541,
637
],
"spans": [
{
"bbox": [
69,
623,
191,
637
],
"score": 1.0,
"content": "following a forward process",
"type": "text"
},
{
"bbox": [
191,
626,
218,
636
],
"score": 0.94,
"content": "q ( \\mathbf { z } | \\mathbf { x } )",
"type": "inline_equation"
},
{
"bbox": [
218,
623,
291,
637
],
"score": 1.0,
"content": "starting at data",
"type": "text"
},
{
"bbox": [
291,
626,
329,
636
],
"score": 0.94,
"content": "\\mathbf { x } \\sim p ( \\mathbf { x } )",
"type": "inline_equation"
},
{
"bbox": [
330,
623,
541,
637
],
"score": 1.0,
"content": ". The forward process is a Gaussian process that",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
636,
218,
649
],
"spans": [
{
"bbox": [
69,
636,
218,
649
],
"score": 1.0,
"content": "satisfies the Markovian structure:",
"type": "text"
}
],
"index": 38
}
],
"index": 36,
"bbox_fs": [
69,
587,
542,
649
]
},
{
"type": "interline_equation",
"bbox": [
172,
657,
440,
672
],
"lines": [
{
"bbox": [
172,
657,
440,
672
],
"spans": [
{
"bbox": [
172,
657,
440,
672
],
"score": 0.86,
"content": "\\begin{array} { r } { q ( \\mathbf { z } _ { t } | \\mathbf { x } ) = \\mathcal { N } ( \\mathbf { z } _ { t } ; \\alpha _ { t } \\mathbf { x } , \\sigma _ { t } ^ { 2 } \\mathbf { I } ) , \\quad q ( \\mathbf { z } _ { t } | \\mathbf { z } _ { s } ) = \\mathcal { N } ( \\mathbf { z } _ { t } ; ( \\alpha _ { t } / \\alpha _ { s } ) \\mathbf { z } _ { s } , \\sigma _ { t | s } ^ { 2 } \\mathbf { I } ) } \\end{array}",
"type": "interline_equation",
"image_path": "91e32dd5f663fe175f4fe9874f7db7b0f12fcefcf05a707499a151a1035386b8.jpg"
}
]
}
],
"index": 39,
"virtual_lines": [
{
"bbox": [
172,
657,
440,
672
],
"spans": [],
"index": 39
}
]
},
{
"type": "text",
"bbox": [
70,
681,
540,
732
],
"lines": [
{
"bbox": [
68,
680,
543,
697
],
"spans": [
{
"bbox": [
68,
680,
99,
696
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
100,
685,
158,
694
],
"score": 0.9,
"content": "0 \\leq s < t \\leq 1",
"type": "inline_equation"
},
{
"bbox": [
158,
680,
163,
696
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
163,
684,
254,
697
],
"score": 0.94,
"content": "\\sigma _ { t | s } ^ { 2 } = ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\sigma _ { t } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
254,
680,
278,
696
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
278,
687,
302,
694
],
"score": 0.91,
"content": "\\alpha _ { t } , \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
302,
680,
543,
696
],
"score": 1.0,
"content": "specify a noise schedule whose log signal-to-noise-ratio",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
71,
695,
542,
711
],
"spans": [
{
"bbox": [
71,
698,
140,
709
],
"score": 0.94,
"content": "\\lambda _ { t } = \\log [ \\alpha _ { t } ^ { 2 } / \\sigma _ { t } ^ { 2 } ]",
"type": "inline_equation"
},
{
"bbox": [
140,
695,
273,
711
],
"score": 1.0,
"content": "decreases monotonically with",
"type": "text"
},
{
"bbox": [
274,
700,
278,
706
],
"score": 0.87,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
695,
305,
711
],
"score": 1.0,
"content": "until",
"type": "text"
},
{
"bbox": [
305,
698,
373,
709
],
"score": 0.94,
"content": "q ( \\mathbf { z } _ { 1 } ) \\approx \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { I } )",
"type": "inline_equation"
},
{
"bbox": [
374,
695,
542,
711
],
"score": 1.0,
"content": ". We use a continuous time version of",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
708,
541,
722
],
"spans": [
{
"bbox": [
70,
708,
541,
722
],
"score": 1.0,
"content": "the cosine noise schedule (Nichol & Dhariwal, 2021). The generative model is a learned model that matches",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
720,
528,
734
],
"spans": [
{
"bbox": [
69,
720,
339,
734
],
"score": 1.0,
"content": "this forward process in the reverse time direction, generating",
"type": "text"
},
{
"bbox": [
340,
725,
348,
732
],
"score": 0.89,
"content": "\\mathbf { z } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
349,
720,
412,
734
],
"score": 1.0,
"content": "starting from",
"type": "text"
},
{
"bbox": [
412,
723,
434,
730
],
"score": 0.92,
"content": "t = 1",
"type": "inline_equation"
},
{
"bbox": [
435,
720,
501,
734
],
"score": 1.0,
"content": "and ending at",
"type": "text"
},
{
"bbox": [
501,
723,
523,
730
],
"score": 0.92,
"content": "t = 0",
"type": "inline_equation"
},
{
"bbox": [
524,
720,
528,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 43
}
],
"index": 41.5,
"bbox_fs": [
68,
680,
543,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
118
],
"lines": [
{
"bbox": [
69,
81,
540,
96
],
"spans": [
{
"bbox": [
69,
81,
485,
96
],
"score": 1.0,
"content": "Learning to reverse the forward process for generation can be reduced to learning to denoise",
"type": "text"
},
{
"bbox": [
485,
84,
540,
95
],
"score": 0.94,
"content": "{ \\mathbf z } _ { t } \\sim q ( { \\mathbf z } _ { t } | { \\mathbf x } )",
"type": "inline_equation"
}
],
"index": 0
},
{
"bbox": [
68,
91,
542,
110
],
"spans": [
{
"bbox": [
68,
91,
146,
110
],
"score": 1.0,
"content": "into an estimate",
"type": "text"
},
{
"bbox": [
147,
96,
209,
106
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( { \\mathbf z } _ { t } , \\lambda _ { t } ) \\approx { \\mathbf x }",
"type": "inline_equation"
},
{
"bbox": [
209,
91,
243,
110
],
"score": 1.0,
"content": "for all",
"type": "text"
},
{
"bbox": [
243,
97,
247,
104
],
"score": 0.87,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
247,
91,
542,
110
],
"score": 1.0,
"content": ". Like (Song & Ermon, 2019; Ho et al., 2020) and most follow-up",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
70,
106,
397,
119
],
"spans": [
{
"bbox": [
70,
106,
397,
119
],
"score": 1.0,
"content": "work, we optimize the model by minimizing a simple noise-prediction loss:",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "interline_equation",
"bbox": [
210,
128,
400,
142
],
"lines": [
{
"bbox": [
210,
128,
400,
142
],
"spans": [
{
"bbox": [
210,
128,
400,
142
],
"score": 0.91,
"content": "\\mathcal { L } ( \\mathbf { x } ) = \\mathbb { E } _ { \\epsilon \\sim \\mathcal { N } ( 0 , \\mathbf { I } ) , t \\sim U ( 0 , 1 ) } \\big [ \\| \\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) - \\epsilon \\| _ { 2 } ^ { 2 } \\big ]",
"type": "interline_equation",
"image_path": "db4bd1a4a4cd023b03677a900aa10cb38dbaf8e9b28230ce0e6a05cc97c09a70.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
210,
128,
400,
142
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
71,
152,
540,
189
],
"lines": [
{
"bbox": [
69,
151,
541,
168
],
"spans": [
{
"bbox": [
69,
151,
99,
168
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
99,
157,
162,
164
],
"score": 0.92,
"content": "\\mathbf { z } _ { t } = \\alpha _ { t } \\mathbf { x } + \\sigma _ { t } \\mathbf { \\epsilon } ",
"type": "inline_equation"
},
{
"bbox": [
162,
151,
186,
168
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
186,
154,
333,
165
],
"score": 0.93,
"content": "\\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) = \\sigma _ { t } ^ { - 1 } ( \\mathbf { z } _ { t } - \\alpha _ { t } \\hat { \\mathbf { x } } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) )",
"type": "inline_equation"
},
{
"bbox": [
333,
151,
480,
168
],
"score": 1.0,
"content": ". We will drop the dependence on",
"type": "text"
},
{
"bbox": [
481,
156,
490,
164
],
"score": 0.91,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
491,
151,
541,
168
],
"score": 1.0,
"content": "to simplify",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
164,
541,
179
],
"spans": [
{
"bbox": [
69,
164,
353,
179
],
"score": 1.0,
"content": "notation. In practice, we parameterize our models in terms of the",
"type": "text"
},
{
"bbox": [
354,
171,
360,
175
],
"score": 0.8,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
360,
164,
541,
179
],
"score": 1.0,
"content": "-parameterization (Salimans & Ho, 2022),",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
70,
177,
308,
190
],
"spans": [
{
"bbox": [
70,
177,
171,
190
],
"score": 1.0,
"content": "rather than predicting",
"type": "text"
},
{
"bbox": [
172,
182,
177,
187
],
"score": 0.89,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
177,
177,
191,
190
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
192,
182,
198,
187
],
"score": 0.89,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
199,
177,
308,
190
],
"score": 1.0,
"content": "directly; see Section 2.4.",
"type": "text"
}
],
"index": 6
}
],
"index": 5
},
{
"type": "text",
"bbox": [
70,
195,
540,
267
],
"lines": [
{
"bbox": [
69,
194,
542,
209
],
"spans": [
{
"bbox": [
69,
194,
411,
209
],
"score": 1.0,
"content": "For conditional generative modeling, we provide the conditioning information",
"type": "text"
},
{
"bbox": [
412,
200,
417,
205
],
"score": 0.72,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
417,
194,
504,
209
],
"score": 1.0,
"content": "drawn jointly with",
"type": "text"
},
{
"bbox": [
504,
200,
510,
205
],
"score": 0.81,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
511,
194,
542,
209
],
"score": 1.0,
"content": "to the",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
207,
541,
221
],
"spans": [
{
"bbox": [
69,
207,
113,
221
],
"score": 1.0,
"content": "model as",
"type": "text"
},
{
"bbox": [
113,
209,
154,
219
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
154,
207,
541,
221
],
"score": 1.0,
"content": ". We use these conditional diffusion models for spatial and temporal super-resolution in",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
68,
218,
542,
232
],
"spans": [
{
"bbox": [
68,
218,
275,
232
],
"score": 1.0,
"content": "our pipeline of diffusion models: in these cases,",
"type": "text"
},
{
"bbox": [
275,
224,
281,
229
],
"score": 0.6,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
281,
218,
542,
232
],
"score": 1.0,
"content": "includes both the text and the previous stage low resolution",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
230,
542,
244
],
"spans": [
{
"bbox": [
69,
230,
181,
244
],
"score": 1.0,
"content": "video as well as a signal",
"type": "text"
},
{
"bbox": [
182,
233,
191,
243
],
"score": 0.92,
"content": "\\lambda _ { t } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
191,
230,
492,
244
],
"score": 1.0,
"content": "that describes the strength of conditioning augmentation added to",
"type": "text"
},
{
"bbox": [
493,
236,
498,
241
],
"score": 0.66,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
498,
230,
542,
244
],
"score": 1.0,
"content": ". Saharia",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
68,
241,
541,
256
],
"spans": [
{
"bbox": [
68,
241,
541,
256
],
"score": 1.0,
"content": "et al. (2022b) found it critical to condition all the super-resolution models with the text embedding, and we",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
254,
163,
268
],
"spans": [
{
"bbox": [
69,
254,
163,
268
],
"score": 1.0,
"content": "follow this approach.",
"type": "text"
}
],
"index": 12
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
70,
272,
540,
322
],
"lines": [
{
"bbox": [
70,
272,
541,
286
],
"spans": [
{
"bbox": [
70,
272,
541,
286
],
"score": 1.0,
"content": "We use the discrete time ancestral sampler (Ho et al., 2020), with sampling variances derived from lower and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
68,
284,
541,
298
],
"spans": [
{
"bbox": [
68,
284,
485,
298
],
"score": 1.0,
"content": "upper bounds on reverse process entropy (Sohl-Dickstein et al., 2015; Ho et al., 2020; Nichol",
"type": "text"
},
{
"bbox": [
485,
287,
493,
295
],
"score": 0.67,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
493,
284,
541,
298
],
"score": 1.0,
"content": "Dhariwal,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
296,
541,
310
],
"spans": [
{
"bbox": [
69,
296,
484,
310
],
"score": 1.0,
"content": "2021). This sampler can be formulated by using a reversed description of the forward process as",
"type": "text"
},
{
"bbox": [
485,
299,
541,
309
],
"score": 0.93,
"content": "q ( \\mathbf { z } _ { s } | \\mathbf { z } _ { t } , \\mathbf { x } ) =",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
71,
307,
269,
324
],
"spans": [
{
"bbox": [
71,
310,
170,
324
],
"score": 0.92,
"content": "\\mathcal { N } ( \\mathbf { z } _ { s } ; \\tilde { \\pmb { \\mu } } _ { s | t } ( \\mathbf { z } _ { t } , \\mathbf { x } ) , \\tilde { \\sigma } _ { s | t } ^ { 2 } \\mathbf { I } )",
"type": "inline_equation"
},
{
"bbox": [
171,
307,
208,
324
],
"score": 1.0,
"content": "(noting",
"type": "text"
},
{
"bbox": [
208,
312,
230,
319
],
"score": 0.88,
"content": "s < t",
"type": "inline_equation"
},
{
"bbox": [
230,
307,
269,
324
],
"score": 1.0,
"content": "), where",
"type": "text"
}
],
"index": 16
}
],
"index": 14.5
},
{
"type": "interline_equation",
"bbox": [
134,
333,
477,
348
],
"lines": [
{
"bbox": [
134,
333,
477,
348
],
"spans": [
{
"bbox": [
134,
333,
477,
348
],
"score": 0.89,
"content": "\\tilde { \\mu } _ { s | t } ( \\mathbf { z } _ { t } , \\mathbf { x } ) = e ^ { \\lambda _ { t } - \\lambda _ { s } } ( \\alpha _ { s } / \\alpha _ { t } ) \\mathbf { z } _ { t } + ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\alpha _ { s } \\mathbf { x } \\quad \\mathrm { a n d } \\quad \\tilde { \\sigma } _ { s | t } ^ { 2 } = ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\sigma _ { s } ^ { 2 } .",
"type": "interline_equation",
"image_path": "a7b880d8eadfc62c0f843b873189b1cb8c4885c743fc1fd90542e4ff2258635f.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
134,
333,
477,
348
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
71,
357,
348,
369
],
"lines": [
{
"bbox": [
70,
357,
348,
370
],
"spans": [
{
"bbox": [
70,
357,
122,
370
],
"score": 1.0,
"content": "Starting at",
"type": "text"
},
{
"bbox": [
122,
359,
177,
370
],
"score": 0.95,
"content": "{ \\bf z } _ { 1 } \\sim \\mathcal { N } ( { \\bf 0 } , { \\bf I } )",
"type": "inline_equation"
},
{
"bbox": [
177,
357,
348,
370
],
"score": 1.0,
"content": ", the ancestral sampler follows the rule",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "interline_equation",
"bbox": [
214,
379,
397,
400
],
"lines": [
{
"bbox": [
214,
379,
397,
400
],
"spans": [
{
"bbox": [
214,
379,
397,
400
],
"score": 0.94,
"content": "\\mathbf { z } _ { s } = \\tilde { \\mu } _ { s | t } ( \\mathbf { z } _ { t } , \\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ) + \\sqrt { ( \\tilde { \\sigma } _ { s | t } ^ { 2 } ) ^ { 1 - \\gamma } ( \\sigma _ { t | s } ^ { 2 } ) ^ { \\gamma } } \\epsilon",
"type": "interline_equation",
"image_path": "9a4f0f1639000ce1432966e3bf67c71de0417417382e4f2966860af208c6efa2.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
214,
379,
397,
400
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
70,
408,
540,
445
],
"lines": [
{
"bbox": [
69,
407,
540,
422
],
"spans": [
{
"bbox": [
69,
407,
101,
422
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
102,
414,
107,
419
],
"score": 0.87,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
107,
407,
240,
422
],
"score": 1.0,
"content": "is standard Gaussian noise,",
"type": "text"
},
{
"bbox": [
240,
414,
246,
421
],
"score": 0.92,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
247,
407,
540,
422
],
"score": 1.0,
"content": "is a hyperparameter that controls the stochasticity of the sam-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
70,
420,
541,
434
],
"spans": [
{
"bbox": [
70,
420,
233,
434
],
"score": 1.0,
"content": "pler (Nichol & Dhariwal, 2021), and",
"type": "text"
},
{
"bbox": [
234,
424,
246,
432
],
"score": 0.92,
"content": "s , t",
"type": "inline_equation"
},
{
"bbox": [
247,
420,
452,
434
],
"score": 1.0,
"content": "follow a uniformly spaced sequence from 1 to",
"type": "text"
},
{
"bbox": [
452,
424,
457,
431
],
"score": 0.29,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
458,
420,
541,
434
],
"score": 1.0,
"content": ". See Section 3 for",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
433,
218,
447
],
"spans": [
{
"bbox": [
69,
433,
218,
447
],
"score": 1.0,
"content": "sampler hyperparameter settings.",
"type": "text"
}
],
"index": 22
}
],
"index": 21
},
{
"type": "text",
"bbox": [
70,
450,
540,
510
],
"lines": [
{
"bbox": [
70,
451,
540,
463
],
"spans": [
{
"bbox": [
70,
451,
540,
463
],
"score": 1.0,
"content": "Alternatively, the deterministic DDIM sampler (Song et al., 2020) can be used for sampling. This sampler",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
463,
540,
475
],
"spans": [
{
"bbox": [
69,
463,
540,
475
],
"score": 1.0,
"content": "is a numerical integration rule for the probability flow ODE (Song et al., 2021; Salimans & Ho, 2022),",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
474,
541,
487
],
"spans": [
{
"bbox": [
69,
474,
541,
487
],
"score": 1.0,
"content": "which describes how a sample from a standard normal distribution can be deterministically transformed",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
70,
487,
540,
499
],
"spans": [
{
"bbox": [
70,
487,
540,
499
],
"score": 1.0,
"content": "into a sample from the video data distribution using the denoising model. The DDIM sampler is useful for",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
498,
372,
511
],
"spans": [
{
"bbox": [
69,
498,
372,
511
],
"score": 1.0,
"content": "progressive distillation for fast sampling, as described in Section 2.7.",
"type": "text"
}
],
"index": 27
}
],
"index": 25
},
{
"type": "title",
"bbox": [
71,
524,
322,
537
],
"lines": [
{
"bbox": [
69,
523,
323,
540
],
"spans": [
{
"bbox": [
69,
523,
323,
540
],
"score": 1.0,
"content": "2.2 Cascaded Diffusion Models and text conditioning",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
70,
546,
540,
655
],
"lines": [
{
"bbox": [
70,
547,
540,
560
],
"spans": [
{
"bbox": [
70,
547,
540,
560
],
"score": 1.0,
"content": "Cascaded Diffusion Models (Ho et al., 2022a) are an effective method for scaling diffusion models to high",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
559,
541,
572
],
"spans": [
{
"bbox": [
69,
559,
541,
572
],
"score": 1.0,
"content": "resolution outputs, finding considerable success in both class-conditional ImageNet (Ho et al., 2022a) and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
68,
570,
541,
584
],
"spans": [
{
"bbox": [
68,
570,
541,
584
],
"score": 1.0,
"content": "text-to-image generation (Ramesh et al., 2022; Saharia et al., 2022b). Cascaded diffusion models generate",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
68,
582,
541,
596
],
"spans": [
{
"bbox": [
68,
582,
541,
596
],
"score": 1.0,
"content": "an image or video at a low resolution, then sequentially increase the resolution of the image or video through",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
594,
540,
607
],
"spans": [
{
"bbox": [
69,
594,
540,
607
],
"score": 1.0,
"content": "a series of super-resolution diffusion models. Cascaded Diffusion Models can model very high dimensional",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
607,
541,
619
],
"spans": [
{
"bbox": [
70,
607,
541,
619
],
"score": 1.0,
"content": "problems while still keeping each sub-model relatively simple. Imagen (Saharia et al., 2022b) also showed",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
619,
541,
632
],
"spans": [
{
"bbox": [
70,
619,
541,
632
],
"score": 1.0,
"content": "that by conditioning on text embeddings from a large frozen language model in conjunction with cascaded",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
279,
644
],
"score": 1.0,
"content": "diffusion models, one can generate high quality",
"type": "text"
},
{
"bbox": [
280,
633,
332,
641
],
"score": 0.88,
"content": "1 0 2 4 \\times 1 0 2 4",
"type": "inline_equation"
},
{
"bbox": [
333,
630,
542,
644
],
"score": 1.0,
"content": "images from text descriptions. In this work we",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
69,
643,
253,
656
],
"spans": [
{
"bbox": [
69,
643,
253,
656
],
"score": 1.0,
"content": "extend this approach to video generation.",
"type": "text"
}
],
"index": 37
}
],
"index": 33
},
{
"type": "text",
"bbox": [
70,
660,
540,
732
],
"lines": [
{
"bbox": [
70,
660,
540,
673
],
"spans": [
{
"bbox": [
70,
660,
540,
673
],
"score": 1.0,
"content": "Figure 6 summarizes the entire cascading pipeline of Imagen Video. In total, we have 1 frozen text encoder, 1",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "base video diffusion model, 3 SSR (spatial super-resolution), and 3 TSR (temporal super-resolution) models",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
76,
685,
541,
697
],
"spans": [
{
"bbox": [
76,
685,
541,
697
],
"score": 1.0,
"content": "for a total of 7 video diffusion models, with a total of 11.6B diffusion model parameters. The data used to",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
69,
696,
541,
710
],
"spans": [
{
"bbox": [
69,
696,
541,
710
],
"score": 1.0,
"content": "train these models is processed to the appropriate spatial and temporal resolutions by spatial resizing and",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
709,
541,
721
],
"spans": [
{
"bbox": [
70,
709,
541,
721
],
"score": 1.0,
"content": "frame skipping. At generation time, the SSR models increase spatial resolution for all input frames, whereas",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
720,
542,
734
],
"spans": [
{
"bbox": [
69,
720,
542,
734
],
"score": 1.0,
"content": "the TSR models increase temporal resolution by filling in intermediate frames between input frames. All",
"type": "text"
}
],
"index": 43
}
],
"index": 40.5
}
],
"page_idx": 6,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
26,
237,
37
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "7",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
118
],
"lines": [
{
"bbox": [
69,
81,
540,
96
],
"spans": [
{
"bbox": [
69,
81,
485,
96
],
"score": 1.0,
"content": "Learning to reverse the forward process for generation can be reduced to learning to denoise",
"type": "text"
},
{
"bbox": [
485,
84,
540,
95
],
"score": 0.94,
"content": "{ \\mathbf z } _ { t } \\sim q ( { \\mathbf z } _ { t } | { \\mathbf x } )",
"type": "inline_equation"
}
],
"index": 0
},
{
"bbox": [
68,
91,
542,
110
],
"spans": [
{
"bbox": [
68,
91,
146,
110
],
"score": 1.0,
"content": "into an estimate",
"type": "text"
},
{
"bbox": [
147,
96,
209,
106
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( { \\mathbf z } _ { t } , \\lambda _ { t } ) \\approx { \\mathbf x }",
"type": "inline_equation"
},
{
"bbox": [
209,
91,
243,
110
],
"score": 1.0,
"content": "for all",
"type": "text"
},
{
"bbox": [
243,
97,
247,
104
],
"score": 0.87,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
247,
91,
542,
110
],
"score": 1.0,
"content": ". Like (Song & Ermon, 2019; Ho et al., 2020) and most follow-up",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
70,
106,
397,
119
],
"spans": [
{
"bbox": [
70,
106,
397,
119
],
"score": 1.0,
"content": "work, we optimize the model by minimizing a simple noise-prediction loss:",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
68,
81,
542,
119
]
},
{
"type": "interline_equation",
"bbox": [
210,
128,
400,
142
],
"lines": [
{
"bbox": [
210,
128,
400,
142
],
"spans": [
{
"bbox": [
210,
128,
400,
142
],
"score": 0.91,
"content": "\\mathcal { L } ( \\mathbf { x } ) = \\mathbb { E } _ { \\epsilon \\sim \\mathcal { N } ( 0 , \\mathbf { I } ) , t \\sim U ( 0 , 1 ) } \\big [ \\| \\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) - \\epsilon \\| _ { 2 } ^ { 2 } \\big ]",
"type": "interline_equation",
"image_path": "db4bd1a4a4cd023b03677a900aa10cb38dbaf8e9b28230ce0e6a05cc97c09a70.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
210,
128,
400,
142
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
71,
152,
540,
189
],
"lines": [
{
"bbox": [
69,
151,
541,
168
],
"spans": [
{
"bbox": [
69,
151,
99,
168
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
99,
157,
162,
164
],
"score": 0.92,
"content": "\\mathbf { z } _ { t } = \\alpha _ { t } \\mathbf { x } + \\sigma _ { t } \\mathbf { \\epsilon } ",
"type": "inline_equation"
},
{
"bbox": [
162,
151,
186,
168
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
186,
154,
333,
165
],
"score": 0.93,
"content": "\\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) = \\sigma _ { t } ^ { - 1 } ( \\mathbf { z } _ { t } - \\alpha _ { t } \\hat { \\mathbf { x } } _ { \\boldsymbol { \\theta } } ( \\mathbf { z } _ { t } , \\lambda _ { t } ) )",
"type": "inline_equation"
},
{
"bbox": [
333,
151,
480,
168
],
"score": 1.0,
"content": ". We will drop the dependence on",
"type": "text"
},
{
"bbox": [
481,
156,
490,
164
],
"score": 0.91,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
491,
151,
541,
168
],
"score": 1.0,
"content": "to simplify",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
164,
541,
179
],
"spans": [
{
"bbox": [
69,
164,
353,
179
],
"score": 1.0,
"content": "notation. In practice, we parameterize our models in terms of the",
"type": "text"
},
{
"bbox": [
354,
171,
360,
175
],
"score": 0.8,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
360,
164,
541,
179
],
"score": 1.0,
"content": "-parameterization (Salimans & Ho, 2022),",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
70,
177,
308,
190
],
"spans": [
{
"bbox": [
70,
177,
171,
190
],
"score": 1.0,
"content": "rather than predicting",
"type": "text"
},
{
"bbox": [
172,
182,
177,
187
],
"score": 0.89,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
177,
177,
191,
190
],
"score": 1.0,
"content": "or",
"type": "text"
},
{
"bbox": [
192,
182,
198,
187
],
"score": 0.89,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
199,
177,
308,
190
],
"score": 1.0,
"content": "directly; see Section 2.4.",
"type": "text"
}
],
"index": 6
}
],
"index": 5,
"bbox_fs": [
69,
151,
541,
190
]
},
{
"type": "text",
"bbox": [
70,
195,
540,
267
],
"lines": [
{
"bbox": [
69,
194,
542,
209
],
"spans": [
{
"bbox": [
69,
194,
411,
209
],
"score": 1.0,
"content": "For conditional generative modeling, we provide the conditioning information",
"type": "text"
},
{
"bbox": [
412,
200,
417,
205
],
"score": 0.72,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
417,
194,
504,
209
],
"score": 1.0,
"content": "drawn jointly with",
"type": "text"
},
{
"bbox": [
504,
200,
510,
205
],
"score": 0.81,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
511,
194,
542,
209
],
"score": 1.0,
"content": "to the",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
207,
541,
221
],
"spans": [
{
"bbox": [
69,
207,
113,
221
],
"score": 1.0,
"content": "model as",
"type": "text"
},
{
"bbox": [
113,
209,
154,
219
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
154,
207,
541,
221
],
"score": 1.0,
"content": ". We use these conditional diffusion models for spatial and temporal super-resolution in",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
68,
218,
542,
232
],
"spans": [
{
"bbox": [
68,
218,
275,
232
],
"score": 1.0,
"content": "our pipeline of diffusion models: in these cases,",
"type": "text"
},
{
"bbox": [
275,
224,
281,
229
],
"score": 0.6,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
281,
218,
542,
232
],
"score": 1.0,
"content": "includes both the text and the previous stage low resolution",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
230,
542,
244
],
"spans": [
{
"bbox": [
69,
230,
181,
244
],
"score": 1.0,
"content": "video as well as a signal",
"type": "text"
},
{
"bbox": [
182,
233,
191,
243
],
"score": 0.92,
"content": "\\lambda _ { t } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
191,
230,
492,
244
],
"score": 1.0,
"content": "that describes the strength of conditioning augmentation added to",
"type": "text"
},
{
"bbox": [
493,
236,
498,
241
],
"score": 0.66,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
498,
230,
542,
244
],
"score": 1.0,
"content": ". Saharia",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
68,
241,
541,
256
],
"spans": [
{
"bbox": [
68,
241,
541,
256
],
"score": 1.0,
"content": "et al. (2022b) found it critical to condition all the super-resolution models with the text embedding, and we",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
254,
163,
268
],
"spans": [
{
"bbox": [
69,
254,
163,
268
],
"score": 1.0,
"content": "follow this approach.",
"type": "text"
}
],
"index": 12
}
],
"index": 9.5,
"bbox_fs": [
68,
194,
542,
268
]
},
{
"type": "text",
"bbox": [
70,
272,
540,
322
],
"lines": [
{
"bbox": [
70,
272,
541,
286
],
"spans": [
{
"bbox": [
70,
272,
541,
286
],
"score": 1.0,
"content": "We use the discrete time ancestral sampler (Ho et al., 2020), with sampling variances derived from lower and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
68,
284,
541,
298
],
"spans": [
{
"bbox": [
68,
284,
485,
298
],
"score": 1.0,
"content": "upper bounds on reverse process entropy (Sohl-Dickstein et al., 2015; Ho et al., 2020; Nichol",
"type": "text"
},
{
"bbox": [
485,
287,
493,
295
],
"score": 0.67,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
493,
284,
541,
298
],
"score": 1.0,
"content": "Dhariwal,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
296,
541,
310
],
"spans": [
{
"bbox": [
69,
296,
484,
310
],
"score": 1.0,
"content": "2021). This sampler can be formulated by using a reversed description of the forward process as",
"type": "text"
},
{
"bbox": [
485,
299,
541,
309
],
"score": 0.93,
"content": "q ( \\mathbf { z } _ { s } | \\mathbf { z } _ { t } , \\mathbf { x } ) =",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
71,
307,
269,
324
],
"spans": [
{
"bbox": [
71,
310,
170,
324
],
"score": 0.92,
"content": "\\mathcal { N } ( \\mathbf { z } _ { s } ; \\tilde { \\pmb { \\mu } } _ { s | t } ( \\mathbf { z } _ { t } , \\mathbf { x } ) , \\tilde { \\sigma } _ { s | t } ^ { 2 } \\mathbf { I } )",
"type": "inline_equation"
},
{
"bbox": [
171,
307,
208,
324
],
"score": 1.0,
"content": "(noting",
"type": "text"
},
{
"bbox": [
208,
312,
230,
319
],
"score": 0.88,
"content": "s < t",
"type": "inline_equation"
},
{
"bbox": [
230,
307,
269,
324
],
"score": 1.0,
"content": "), where",
"type": "text"
}
],
"index": 16
}
],
"index": 14.5,
"bbox_fs": [
68,
272,
541,
324
]
},
{
"type": "interline_equation",
"bbox": [
134,
333,
477,
348
],
"lines": [
{
"bbox": [
134,
333,
477,
348
],
"spans": [
{
"bbox": [
134,
333,
477,
348
],
"score": 0.89,
"content": "\\tilde { \\mu } _ { s | t } ( \\mathbf { z } _ { t } , \\mathbf { x } ) = e ^ { \\lambda _ { t } - \\lambda _ { s } } ( \\alpha _ { s } / \\alpha _ { t } ) \\mathbf { z } _ { t } + ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\alpha _ { s } \\mathbf { x } \\quad \\mathrm { a n d } \\quad \\tilde { \\sigma } _ { s | t } ^ { 2 } = ( 1 - e ^ { \\lambda _ { t } - \\lambda _ { s } } ) \\sigma _ { s } ^ { 2 } .",
"type": "interline_equation",
"image_path": "a7b880d8eadfc62c0f843b873189b1cb8c4885c743fc1fd90542e4ff2258635f.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
134,
333,
477,
348
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
71,
357,
348,
369
],
"lines": [
{
"bbox": [
70,
357,
348,
370
],
"spans": [
{
"bbox": [
70,
357,
122,
370
],
"score": 1.0,
"content": "Starting at",
"type": "text"
},
{
"bbox": [
122,
359,
177,
370
],
"score": 0.95,
"content": "{ \\bf z } _ { 1 } \\sim \\mathcal { N } ( { \\bf 0 } , { \\bf I } )",
"type": "inline_equation"
},
{
"bbox": [
177,
357,
348,
370
],
"score": 1.0,
"content": ", the ancestral sampler follows the rule",
"type": "text"
}
],
"index": 18
}
],
"index": 18,
"bbox_fs": [
70,
357,
348,
370
]
},
{
"type": "interline_equation",
"bbox": [
214,
379,
397,
400
],
"lines": [
{
"bbox": [
214,
379,
397,
400
],
"spans": [
{
"bbox": [
214,
379,
397,
400
],
"score": 0.94,
"content": "\\mathbf { z } _ { s } = \\tilde { \\mu } _ { s | t } ( \\mathbf { z } _ { t } , \\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ) + \\sqrt { ( \\tilde { \\sigma } _ { s | t } ^ { 2 } ) ^ { 1 - \\gamma } ( \\sigma _ { t | s } ^ { 2 } ) ^ { \\gamma } } \\epsilon",
"type": "interline_equation",
"image_path": "9a4f0f1639000ce1432966e3bf67c71de0417417382e4f2966860af208c6efa2.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
214,
379,
397,
400
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
70,
408,
540,
445
],
"lines": [
{
"bbox": [
69,
407,
540,
422
],
"spans": [
{
"bbox": [
69,
407,
101,
422
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
102,
414,
107,
419
],
"score": 0.87,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
107,
407,
240,
422
],
"score": 1.0,
"content": "is standard Gaussian noise,",
"type": "text"
},
{
"bbox": [
240,
414,
246,
421
],
"score": 0.92,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
247,
407,
540,
422
],
"score": 1.0,
"content": "is a hyperparameter that controls the stochasticity of the sam-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
70,
420,
541,
434
],
"spans": [
{
"bbox": [
70,
420,
233,
434
],
"score": 1.0,
"content": "pler (Nichol & Dhariwal, 2021), and",
"type": "text"
},
{
"bbox": [
234,
424,
246,
432
],
"score": 0.92,
"content": "s , t",
"type": "inline_equation"
},
{
"bbox": [
247,
420,
452,
434
],
"score": 1.0,
"content": "follow a uniformly spaced sequence from 1 to",
"type": "text"
},
{
"bbox": [
452,
424,
457,
431
],
"score": 0.29,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
458,
420,
541,
434
],
"score": 1.0,
"content": ". See Section 3 for",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
433,
218,
447
],
"spans": [
{
"bbox": [
69,
433,
218,
447
],
"score": 1.0,
"content": "sampler hyperparameter settings.",
"type": "text"
}
],
"index": 22
}
],
"index": 21,
"bbox_fs": [
69,
407,
541,
447
]
},
{
"type": "text",
"bbox": [
70,
450,
540,
510
],
"lines": [
{
"bbox": [
70,
451,
540,
463
],
"spans": [
{
"bbox": [
70,
451,
540,
463
],
"score": 1.0,
"content": "Alternatively, the deterministic DDIM sampler (Song et al., 2020) can be used for sampling. This sampler",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
463,
540,
475
],
"spans": [
{
"bbox": [
69,
463,
540,
475
],
"score": 1.0,
"content": "is a numerical integration rule for the probability flow ODE (Song et al., 2021; Salimans & Ho, 2022),",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
474,
541,
487
],
"spans": [
{
"bbox": [
69,
474,
541,
487
],
"score": 1.0,
"content": "which describes how a sample from a standard normal distribution can be deterministically transformed",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
70,
487,
540,
499
],
"spans": [
{
"bbox": [
70,
487,
540,
499
],
"score": 1.0,
"content": "into a sample from the video data distribution using the denoising model. The DDIM sampler is useful for",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
498,
372,
511
],
"spans": [
{
"bbox": [
69,
498,
372,
511
],
"score": 1.0,
"content": "progressive distillation for fast sampling, as described in Section 2.7.",
"type": "text"
}
],
"index": 27
}
],
"index": 25,
"bbox_fs": [
69,
451,
541,
511
]
},
{
"type": "title",
"bbox": [
71,
524,
322,
537
],
"lines": [
{
"bbox": [
69,
523,
323,
540
],
"spans": [
{
"bbox": [
69,
523,
323,
540
],
"score": 1.0,
"content": "2.2 Cascaded Diffusion Models and text conditioning",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
70,
546,
540,
655
],
"lines": [
{
"bbox": [
70,
547,
540,
560
],
"spans": [
{
"bbox": [
70,
547,
540,
560
],
"score": 1.0,
"content": "Cascaded Diffusion Models (Ho et al., 2022a) are an effective method for scaling diffusion models to high",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
559,
541,
572
],
"spans": [
{
"bbox": [
69,
559,
541,
572
],
"score": 1.0,
"content": "resolution outputs, finding considerable success in both class-conditional ImageNet (Ho et al., 2022a) and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
68,
570,
541,
584
],
"spans": [
{
"bbox": [
68,
570,
541,
584
],
"score": 1.0,
"content": "text-to-image generation (Ramesh et al., 2022; Saharia et al., 2022b). Cascaded diffusion models generate",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
68,
582,
541,
596
],
"spans": [
{
"bbox": [
68,
582,
541,
596
],
"score": 1.0,
"content": "an image or video at a low resolution, then sequentially increase the resolution of the image or video through",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
594,
540,
607
],
"spans": [
{
"bbox": [
69,
594,
540,
607
],
"score": 1.0,
"content": "a series of super-resolution diffusion models. Cascaded Diffusion Models can model very high dimensional",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
607,
541,
619
],
"spans": [
{
"bbox": [
70,
607,
541,
619
],
"score": 1.0,
"content": "problems while still keeping each sub-model relatively simple. Imagen (Saharia et al., 2022b) also showed",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
619,
541,
632
],
"spans": [
{
"bbox": [
70,
619,
541,
632
],
"score": 1.0,
"content": "that by conditioning on text embeddings from a large frozen language model in conjunction with cascaded",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
279,
644
],
"score": 1.0,
"content": "diffusion models, one can generate high quality",
"type": "text"
},
{
"bbox": [
280,
633,
332,
641
],
"score": 0.88,
"content": "1 0 2 4 \\times 1 0 2 4",
"type": "inline_equation"
},
{
"bbox": [
333,
630,
542,
644
],
"score": 1.0,
"content": "images from text descriptions. In this work we",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
69,
643,
253,
656
],
"spans": [
{
"bbox": [
69,
643,
253,
656
],
"score": 1.0,
"content": "extend this approach to video generation.",
"type": "text"
}
],
"index": 37
}
],
"index": 33,
"bbox_fs": [
68,
547,
542,
656
]
},
{
"type": "text",
"bbox": [
70,
660,
540,
732
],
"lines": [
{
"bbox": [
70,
660,
540,
673
],
"spans": [
{
"bbox": [
70,
660,
540,
673
],
"score": 1.0,
"content": "Figure 6 summarizes the entire cascading pipeline of Imagen Video. In total, we have 1 frozen text encoder, 1",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "base video diffusion model, 3 SSR (spatial super-resolution), and 3 TSR (temporal super-resolution) models",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
76,
685,
541,
697
],
"spans": [
{
"bbox": [
76,
685,
541,
697
],
"score": 1.0,
"content": "for a total of 7 video diffusion models, with a total of 11.6B diffusion model parameters. The data used to",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
69,
696,
541,
710
],
"spans": [
{
"bbox": [
69,
696,
541,
710
],
"score": 1.0,
"content": "train these models is processed to the appropriate spatial and temporal resolutions by spatial resizing and",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
709,
541,
721
],
"spans": [
{
"bbox": [
70,
709,
541,
721
],
"score": 1.0,
"content": "frame skipping. At generation time, the SSR models increase spatial resolution for all input frames, whereas",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
720,
542,
734
],
"spans": [
{
"bbox": [
69,
720,
542,
734
],
"score": 1.0,
"content": "the TSR models increase temporal resolution by filling in intermediate frames between input frames. All",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
69,
302,
540,
315
],
"spans": [
{
"bbox": [
69,
302,
540,
315
],
"score": 1.0,
"content": "models generate an entire block of frames simultaneously – so for instance, our SSR models do not suffer",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
69,
313,
518,
327
],
"spans": [
{
"bbox": [
69,
313,
518,
327
],
"score": 1.0,
"content": "from obvious artifacts that would occur from naively running super-resolution on independent frames.",
"type": "text",
"cross_page": true
}
],
"index": 8
}
],
"index": 40.5,
"bbox_fs": [
69,
660,
542,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
81,
91,
530,
208
],
"blocks": [
{
"type": "image_body",
"bbox": [
81,
91,
530,
208
],
"group_id": 0,
"lines": [
{
"bbox": [
81,
91,
530,
208
],
"spans": [
{
"bbox": [
81,
91,
530,
208
],
"score": 0.964,
"type": "image",
"image_path": "6b59996bb4d772372726e02789aea0eee1b3f9ba382f99a8566a6e8d81ec47a7.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
81,
91,
530,
130.0
],
"spans": [],
"index": 0
},
{
"bbox": [
81,
130.0,
530,
169.0
],
"spans": [],
"index": 1
},
{
"bbox": [
81,
169.0,
530,
208.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
70,
228,
540,
275
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
228,
540,
241
],
"spans": [
{
"bbox": [
70,
228,
540,
241
],
"score": 1.0,
"content": "Figure 6: The cascaded sampling pipeline starting from a text prompt input to generating a 5.3-second,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
71,
239,
541,
254
],
"spans": [
{
"bbox": [
71,
241,
114,
251
],
"score": 0.74,
"content": "1 2 8 0 \\times 7 6 8",
"type": "inline_equation"
},
{
"bbox": [
115,
239,
541,
254
],
"score": 1.0,
"content": "video at 24fps. “SSR” and “TSR” denote spatial and temporal super-resolution respectively, and",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
251,
541,
266
],
"spans": [
{
"bbox": [
69,
251,
194,
266
],
"score": 1.0,
"content": "videos are labeled as frames",
"type": "text"
},
{
"bbox": [
194,
254,
202,
263
],
"score": 0.7,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
203,
251,
227,
266
],
"score": 1.0,
"content": "width",
"type": "text"
},
{
"bbox": [
227,
255,
235,
263
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
235,
251,
541,
266
],
"score": 1.0,
"content": "height. In practice, the text embeddings are injected into all models,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
264,
179,
277
],
"spans": [
{
"bbox": [
69,
264,
179,
277
],
"score": 1.0,
"content": "not just the base model.",
"type": "text"
}
],
"index": 6
}
],
"index": 4.5
}
],
"index": 2.75
},
{
"type": "text",
"bbox": [
71,
302,
538,
326
],
"lines": [
{
"bbox": [
69,
302,
540,
315
],
"spans": [
{
"bbox": [
69,
302,
540,
315
],
"score": 1.0,
"content": "models generate an entire block of frames simultaneously – so for instance, our SSR models do not suffer",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
313,
518,
327
],
"spans": [
{
"bbox": [
69,
313,
518,
327
],
"score": 1.0,
"content": "from obvious artifacts that would occur from naively running super-resolution on independent frames.",
"type": "text"
}
],
"index": 8
}
],
"index": 7.5
},
{
"type": "text",
"bbox": [
70,
332,
540,
404
],
"lines": [
{
"bbox": [
69,
330,
542,
347
],
"spans": [
{
"bbox": [
69,
330,
542,
347
],
"score": 1.0,
"content": "One benefit of cascaded models is that each diffusion model can be trained independently, allowing one to",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
343,
541,
358
],
"spans": [
{
"bbox": [
69,
343,
541,
358
],
"score": 1.0,
"content": "train all 7 models in parallel. Additionally, our super-resolution models are general purpose video super-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
69,
356,
542,
370
],
"spans": [
{
"bbox": [
69,
356,
542,
370
],
"score": 1.0,
"content": "resolution models, and they can be applied to real videos or samples from generative models other than the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
367,
542,
381
],
"spans": [
{
"bbox": [
69,
367,
542,
381
],
"score": 1.0,
"content": "ones presented in this paper. This is similar to how Imagen’s super-resolution models helped improve the",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
379,
540,
393
],
"spans": [
{
"bbox": [
69,
379,
540,
393
],
"score": 1.0,
"content": "fidelity of the images generated by Parti (Yu et al., 2022), which is an autoregressive text-to-image model.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
70,
392,
450,
406
],
"spans": [
{
"bbox": [
70,
392,
450,
406
],
"score": 1.0,
"content": "We intend to explore hybrid pipelines of multiple model classes further in future work.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5
},
{
"type": "text",
"bbox": [
71,
409,
540,
469
],
"lines": [
{
"bbox": [
69,
408,
541,
423
],
"spans": [
{
"bbox": [
69,
408,
541,
423
],
"score": 1.0,
"content": "Similar to Saharia et al. (2022b), we utilize contextual embeddings from a frozen T5-XXL text encoder",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
70,
420,
541,
435
],
"spans": [
{
"bbox": [
70,
420,
541,
435
],
"score": 1.0,
"content": "(Raffel et al., 2020) for conditioning on the input text prompt. We find these embeddings to be critical for",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
70,
433,
541,
446
],
"spans": [
{
"bbox": [
70,
433,
541,
446
],
"score": 1.0,
"content": "alignment between generated video and the text prompt. Similar to the findings of Saharia et al. (2022b), we",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
68,
444,
542,
459
],
"spans": [
{
"bbox": [
68,
444,
542,
459
],
"score": 1.0,
"content": "observe evidence of deeper language understanding, enabling us to generate the videos displayed in Figs. 2",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
457,
94,
470
],
"spans": [
{
"bbox": [
69,
457,
94,
470
],
"score": 1.0,
"content": "to 5.",
"type": "text"
}
],
"index": 19
}
],
"index": 17
},
{
"type": "title",
"bbox": [
72,
488,
226,
500
],
"lines": [
{
"bbox": [
69,
486,
228,
502
],
"spans": [
{
"bbox": [
69,
486,
228,
502
],
"score": 1.0,
"content": "2.3 Video diffusion architectures",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
70,
511,
540,
690
],
"lines": [
{
"bbox": [
68,
510,
542,
525
],
"spans": [
{
"bbox": [
68,
510,
542,
525
],
"score": 1.0,
"content": "Diffusion models for image generation typically use a 2D U-Net architecture (Ronneberger et al., 2015;",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
522,
541,
535
],
"spans": [
{
"bbox": [
69,
522,
399,
535
],
"score": 1.0,
"content": "Salimans et al., 2017; Ho et al., 2020) to represent the denoising model",
"type": "text"
},
{
"bbox": [
399,
525,
410,
534
],
"score": 0.89,
"content": "\\hat { \\mathbf { x } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
410,
522,
541,
535
],
"score": 1.0,
"content": ". This is a multiscale model",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
70,
536,
541,
547
],
"spans": [
{
"bbox": [
70,
536,
541,
547
],
"score": 1.0,
"content": "consisting of multiple layers of spatial attention and convolution at each resolution, combined with shortcuts",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
547,
541,
560
],
"spans": [
{
"bbox": [
69,
547,
541,
560
],
"score": 1.0,
"content": "between layers at the same resolution. In earlier work on Video Diffusion Models, Ho et al. (2022b) introduced",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
70,
559,
542,
572
],
"spans": [
{
"bbox": [
70,
559,
542,
572
],
"score": 1.0,
"content": "the Video U-Net , which generalizes the 2D diffusion model architecture to 3D in a space-time separable",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
570,
541,
584
],
"spans": [
{
"bbox": [
69,
570,
541,
584
],
"score": 1.0,
"content": "fashion using temporal attention and convolution layers interleaved within spatial attention and convolution",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
582,
542,
596
],
"spans": [
{
"bbox": [
69,
582,
542,
596
],
"score": 1.0,
"content": "layers to capture dependencies between video frames. Our work builds on the Video U-Net architecture: see",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
595,
542,
608
],
"spans": [
{
"bbox": [
69,
595,
386,
608
],
"score": 1.0,
"content": "Figure 7. Following Video Diffusion Models, each of our denoising models",
"type": "text"
},
{
"bbox": [
387,
597,
397,
606
],
"score": 0.91,
"content": "\\hat { \\mathbf { x } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
397,
595,
542,
608
],
"score": 1.0,
"content": "operate on multiple video frames",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
607,
541,
620
],
"spans": [
{
"bbox": [
69,
607,
541,
620
],
"score": 1.0,
"content": "simultaneously and thereby generate entire blocks of video frames at a time, which we find to be important",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
68,
619,
541,
631
],
"spans": [
{
"bbox": [
68,
619,
541,
631
],
"score": 1.0,
"content": "to capture the temporal coherence of the generated video compared to frame-autoregressive approaches. Our",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
542,
644
],
"score": 1.0,
"content": "spatial super-resolution (SSR) and temporal super-resolution (TSR) models condition on their input videos",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
642,
541,
655
],
"spans": [
{
"bbox": [
69,
642,
423,
655
],
"score": 1.0,
"content": "by concatenating an upsampled conditioning input channelwise to the noisy data",
"type": "text"
},
{
"bbox": [
424,
648,
433,
654
],
"score": 0.89,
"content": "\\mathbf { z } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
433,
642,
541,
655
],
"score": 1.0,
"content": ", the same mechanism as",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
654,
542,
667
],
"spans": [
{
"bbox": [
69,
654,
542,
667
],
"score": 1.0,
"content": "SR3 (Saharia et al., 2022c) and Palette (Saharia et al., 2022a): spatial upsampling before concatenation is",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
68,
665,
542,
682
],
"spans": [
{
"bbox": [
68,
665,
542,
682
],
"score": 1.0,
"content": "performed using bilinear resizing, and temporal upsampling before concatenation is performed by repeating",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
678,
228,
691
],
"spans": [
{
"bbox": [
69,
678,
228,
691
],
"score": 1.0,
"content": "frames or by filling in blank frames.",
"type": "text"
}
],
"index": 35
}
],
"index": 28
},
{
"type": "text",
"bbox": [
71,
696,
540,
732
],
"lines": [
{
"bbox": [
69,
695,
541,
710
],
"spans": [
{
"bbox": [
69,
695,
541,
710
],
"score": 1.0,
"content": "Our base video model, which is the first model in the pipeline that generates data at the lowest frame count",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
68,
707,
541,
722
],
"spans": [
{
"bbox": [
68,
707,
541,
722
],
"score": 1.0,
"content": "and spatial resolution, uses temporal attention to mix information across time. Our SSR and TSR models, on",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
720,
541,
733
],
"spans": [
{
"bbox": [
69,
720,
541,
733
],
"score": 1.0,
"content": "the other hand, use temporal convolutions instead of temporal attention. The temporal attention in the base",
"type": "text"
}
],
"index": 38
}
],
"index": 37
}
],
"page_idx": 7,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
36
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
308,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
761
],
"spans": [
{
"bbox": [
302,
750,
309,
761
],
"score": 1.0,
"content": "8",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
81,
91,
530,
208
],
"blocks": [
{
"type": "image_body",
"bbox": [
81,
91,
530,
208
],
"group_id": 0,
"lines": [
{
"bbox": [
81,
91,
530,
208
],
"spans": [
{
"bbox": [
81,
91,
530,
208
],
"score": 0.964,
"type": "image",
"image_path": "6b59996bb4d772372726e02789aea0eee1b3f9ba382f99a8566a6e8d81ec47a7.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
81,
91,
530,
130.0
],
"spans": [],
"index": 0
},
{
"bbox": [
81,
130.0,
530,
169.0
],
"spans": [],
"index": 1
},
{
"bbox": [
81,
169.0,
530,
208.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
70,
228,
540,
275
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
228,
540,
241
],
"spans": [
{
"bbox": [
70,
228,
540,
241
],
"score": 1.0,
"content": "Figure 6: The cascaded sampling pipeline starting from a text prompt input to generating a 5.3-second,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
71,
239,
541,
254
],
"spans": [
{
"bbox": [
71,
241,
114,
251
],
"score": 0.74,
"content": "1 2 8 0 \\times 7 6 8",
"type": "inline_equation"
},
{
"bbox": [
115,
239,
541,
254
],
"score": 1.0,
"content": "video at 24fps. “SSR” and “TSR” denote spatial and temporal super-resolution respectively, and",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
251,
541,
266
],
"spans": [
{
"bbox": [
69,
251,
194,
266
],
"score": 1.0,
"content": "videos are labeled as frames",
"type": "text"
},
{
"bbox": [
194,
254,
202,
263
],
"score": 0.7,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
203,
251,
227,
266
],
"score": 1.0,
"content": "width",
"type": "text"
},
{
"bbox": [
227,
255,
235,
263
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
235,
251,
541,
266
],
"score": 1.0,
"content": "height. In practice, the text embeddings are injected into all models,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
264,
179,
277
],
"spans": [
{
"bbox": [
69,
264,
179,
277
],
"score": 1.0,
"content": "not just the base model.",
"type": "text"
}
],
"index": 6
}
],
"index": 4.5
}
],
"index": 2.75
},
{
"type": "text",
"bbox": [
71,
302,
538,
326
],
"lines": [],
"index": 7.5,
"bbox_fs": [
69,
302,
540,
327
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
70,
332,
540,
404
],
"lines": [
{
"bbox": [
69,
330,
542,
347
],
"spans": [
{
"bbox": [
69,
330,
542,
347
],
"score": 1.0,
"content": "One benefit of cascaded models is that each diffusion model can be trained independently, allowing one to",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
343,
541,
358
],
"spans": [
{
"bbox": [
69,
343,
541,
358
],
"score": 1.0,
"content": "train all 7 models in parallel. Additionally, our super-resolution models are general purpose video super-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
69,
356,
542,
370
],
"spans": [
{
"bbox": [
69,
356,
542,
370
],
"score": 1.0,
"content": "resolution models, and they can be applied to real videos or samples from generative models other than the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
367,
542,
381
],
"spans": [
{
"bbox": [
69,
367,
542,
381
],
"score": 1.0,
"content": "ones presented in this paper. This is similar to how Imagen’s super-resolution models helped improve the",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
379,
540,
393
],
"spans": [
{
"bbox": [
69,
379,
540,
393
],
"score": 1.0,
"content": "fidelity of the images generated by Parti (Yu et al., 2022), which is an autoregressive text-to-image model.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
70,
392,
450,
406
],
"spans": [
{
"bbox": [
70,
392,
450,
406
],
"score": 1.0,
"content": "We intend to explore hybrid pipelines of multiple model classes further in future work.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5,
"bbox_fs": [
69,
330,
542,
406
]
},
{
"type": "text",
"bbox": [
71,
409,
540,
469
],
"lines": [
{
"bbox": [
69,
408,
541,
423
],
"spans": [
{
"bbox": [
69,
408,
541,
423
],
"score": 1.0,
"content": "Similar to Saharia et al. (2022b), we utilize contextual embeddings from a frozen T5-XXL text encoder",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
70,
420,
541,
435
],
"spans": [
{
"bbox": [
70,
420,
541,
435
],
"score": 1.0,
"content": "(Raffel et al., 2020) for conditioning on the input text prompt. We find these embeddings to be critical for",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
70,
433,
541,
446
],
"spans": [
{
"bbox": [
70,
433,
541,
446
],
"score": 1.0,
"content": "alignment between generated video and the text prompt. Similar to the findings of Saharia et al. (2022b), we",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
68,
444,
542,
459
],
"spans": [
{
"bbox": [
68,
444,
542,
459
],
"score": 1.0,
"content": "observe evidence of deeper language understanding, enabling us to generate the videos displayed in Figs. 2",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
457,
94,
470
],
"spans": [
{
"bbox": [
69,
457,
94,
470
],
"score": 1.0,
"content": "to 5.",
"type": "text"
}
],
"index": 19
}
],
"index": 17,
"bbox_fs": [
68,
408,
542,
470
]
},
{
"type": "title",
"bbox": [
72,
488,
226,
500
],
"lines": [
{
"bbox": [
69,
486,
228,
502
],
"spans": [
{
"bbox": [
69,
486,
228,
502
],
"score": 1.0,
"content": "2.3 Video diffusion architectures",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
70,
511,
540,
690
],
"lines": [
{
"bbox": [
68,
510,
542,
525
],
"spans": [
{
"bbox": [
68,
510,
542,
525
],
"score": 1.0,
"content": "Diffusion models for image generation typically use a 2D U-Net architecture (Ronneberger et al., 2015;",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
522,
541,
535
],
"spans": [
{
"bbox": [
69,
522,
399,
535
],
"score": 1.0,
"content": "Salimans et al., 2017; Ho et al., 2020) to represent the denoising model",
"type": "text"
},
{
"bbox": [
399,
525,
410,
534
],
"score": 0.89,
"content": "\\hat { \\mathbf { x } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
410,
522,
541,
535
],
"score": 1.0,
"content": ". This is a multiscale model",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
70,
536,
541,
547
],
"spans": [
{
"bbox": [
70,
536,
541,
547
],
"score": 1.0,
"content": "consisting of multiple layers of spatial attention and convolution at each resolution, combined with shortcuts",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
547,
541,
560
],
"spans": [
{
"bbox": [
69,
547,
541,
560
],
"score": 1.0,
"content": "between layers at the same resolution. In earlier work on Video Diffusion Models, Ho et al. (2022b) introduced",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
70,
559,
542,
572
],
"spans": [
{
"bbox": [
70,
559,
542,
572
],
"score": 1.0,
"content": "the Video U-Net , which generalizes the 2D diffusion model architecture to 3D in a space-time separable",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
570,
541,
584
],
"spans": [
{
"bbox": [
69,
570,
541,
584
],
"score": 1.0,
"content": "fashion using temporal attention and convolution layers interleaved within spatial attention and convolution",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
582,
542,
596
],
"spans": [
{
"bbox": [
69,
582,
542,
596
],
"score": 1.0,
"content": "layers to capture dependencies between video frames. Our work builds on the Video U-Net architecture: see",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
595,
542,
608
],
"spans": [
{
"bbox": [
69,
595,
386,
608
],
"score": 1.0,
"content": "Figure 7. Following Video Diffusion Models, each of our denoising models",
"type": "text"
},
{
"bbox": [
387,
597,
397,
606
],
"score": 0.91,
"content": "\\hat { \\mathbf { x } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
397,
595,
542,
608
],
"score": 1.0,
"content": "operate on multiple video frames",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
607,
541,
620
],
"spans": [
{
"bbox": [
69,
607,
541,
620
],
"score": 1.0,
"content": "simultaneously and thereby generate entire blocks of video frames at a time, which we find to be important",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
68,
619,
541,
631
],
"spans": [
{
"bbox": [
68,
619,
541,
631
],
"score": 1.0,
"content": "to capture the temporal coherence of the generated video compared to frame-autoregressive approaches. Our",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
542,
644
],
"score": 1.0,
"content": "spatial super-resolution (SSR) and temporal super-resolution (TSR) models condition on their input videos",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
642,
541,
655
],
"spans": [
{
"bbox": [
69,
642,
423,
655
],
"score": 1.0,
"content": "by concatenating an upsampled conditioning input channelwise to the noisy data",
"type": "text"
},
{
"bbox": [
424,
648,
433,
654
],
"score": 0.89,
"content": "\\mathbf { z } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
433,
642,
541,
655
],
"score": 1.0,
"content": ", the same mechanism as",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
654,
542,
667
],
"spans": [
{
"bbox": [
69,
654,
542,
667
],
"score": 1.0,
"content": "SR3 (Saharia et al., 2022c) and Palette (Saharia et al., 2022a): spatial upsampling before concatenation is",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
68,
665,
542,
682
],
"spans": [
{
"bbox": [
68,
665,
542,
682
],
"score": 1.0,
"content": "performed using bilinear resizing, and temporal upsampling before concatenation is performed by repeating",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
678,
228,
691
],
"spans": [
{
"bbox": [
69,
678,
228,
691
],
"score": 1.0,
"content": "frames or by filling in blank frames.",
"type": "text"
}
],
"index": 35
}
],
"index": 28,
"bbox_fs": [
68,
510,
542,
691
]
},
{
"type": "text",
"bbox": [
71,
696,
540,
732
],
"lines": [
{
"bbox": [
69,
695,
541,
710
],
"spans": [
{
"bbox": [
69,
695,
541,
710
],
"score": 1.0,
"content": "Our base video model, which is the first model in the pipeline that generates data at the lowest frame count",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
68,
707,
541,
722
],
"spans": [
{
"bbox": [
68,
707,
541,
722
],
"score": 1.0,
"content": "and spatial resolution, uses temporal attention to mix information across time. Our SSR and TSR models, on",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
720,
541,
733
],
"spans": [
{
"bbox": [
69,
720,
541,
733
],
"score": 1.0,
"content": "the other hand, use temporal convolutions instead of temporal attention. The temporal attention in the base",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
300,
542,
315
],
"spans": [
{
"bbox": [
69,
300,
542,
315
],
"score": 1.0,
"content": "model enables Imagen Video to model long term temporal dependencies, while the temporal convolutions in",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
69,
313,
541,
327
],
"spans": [
{
"bbox": [
69,
313,
541,
327
],
"score": 1.0,
"content": "the SSR and TSR models allow Imagen Video to maintain local temporal consistency during upsampling.",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
70,
325,
542,
339
],
"spans": [
{
"bbox": [
70,
325,
542,
339
],
"score": 1.0,
"content": "The use of temporal convolutions lowers memory and computation costs over temporal attention—this is",
"type": "text",
"cross_page": true
}
],
"index": 10
},
{
"bbox": [
69,
338,
541,
351
],
"spans": [
{
"bbox": [
69,
338,
541,
351
],
"score": 1.0,
"content": "crucial because the very purpose of the TSR and SSR models is to operate at high frame rates and spatial",
"type": "text",
"cross_page": true
}
],
"index": 11
},
{
"bbox": [
69,
349,
542,
362
],
"spans": [
{
"bbox": [
69,
349,
542,
362
],
"score": 1.0,
"content": "resolutions. In our initial experiments, we did not find any significant improvements when using temporal",
"type": "text",
"cross_page": true
}
],
"index": 12
},
{
"bbox": [
70,
362,
541,
374
],
"spans": [
{
"bbox": [
70,
362,
541,
374
],
"score": 1.0,
"content": "attention over temporal convolutions in our SSR and TSR models, which we hypothesize is due to the",
"type": "text",
"cross_page": true
}
],
"index": 13
},
{
"bbox": [
69,
373,
513,
386
],
"spans": [
{
"bbox": [
69,
373,
513,
386
],
"score": 1.0,
"content": "significant amount of temporal correlation already present in the conditioning input to these models.",
"type": "text",
"cross_page": true
}
],
"index": 14
}
],
"index": 37,
"bbox_fs": [
68,
695,
541,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
85,
82,
527,
187
],
"blocks": [
{
"type": "image_body",
"bbox": [
85,
82,
527,
187
],
"group_id": 0,
"lines": [
{
"bbox": [
85,
82,
527,
187
],
"spans": [
{
"bbox": [
85,
82,
527,
187
],
"score": 0.964,
"type": "image",
"image_path": "2910e48845f75da3725a555562623a8a34ff8e36072bcf47c79400b54567845e.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
85,
82,
527,
117.0
],
"spans": [],
"index": 0
},
{
"bbox": [
85,
117.0,
527,
152.0
],
"spans": [],
"index": 1
},
{
"bbox": [
85,
152.0,
527,
187.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
70,
219,
540,
279
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
220,
541,
232
],
"spans": [
{
"bbox": [
70,
220,
541,
232
],
"score": 1.0,
"content": "Figure 7: Video U-Net space-time separable block. Spatial operations are performed independently over",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
232,
541,
244
],
"spans": [
{
"bbox": [
70,
232,
541,
244
],
"score": 1.0,
"content": "frames with shared parameters, whereas the temporal operation mixes activations over frames. Our base",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
68,
241,
542,
259
],
"spans": [
{
"bbox": [
68,
241,
542,
259
],
"score": 1.0,
"content": "model uses spatial convolutions, spatial self-attention and temporal self-attention. For memory efficiency,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
255,
541,
269
],
"spans": [
{
"bbox": [
69,
255,
541,
269
],
"score": 1.0,
"content": "our spatial and temporal super-resolution models use temporal convolutions instead of attention, and our",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
69,
267,
377,
281
],
"spans": [
{
"bbox": [
69,
267,
377,
281
],
"score": 1.0,
"content": "models at the highest spatial resolution do not have spatial attention.",
"type": "text"
}
],
"index": 7
}
],
"index": 5
}
],
"index": 3.0
},
{
"type": "text",
"bbox": [
71,
301,
540,
385
],
"lines": [
{
"bbox": [
69,
300,
542,
315
],
"spans": [
{
"bbox": [
69,
300,
542,
315
],
"score": 1.0,
"content": "model enables Imagen Video to model long term temporal dependencies, while the temporal convolutions in",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
313,
541,
327
],
"spans": [
{
"bbox": [
69,
313,
541,
327
],
"score": 1.0,
"content": "the SSR and TSR models allow Imagen Video to maintain local temporal consistency during upsampling.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
70,
325,
542,
339
],
"spans": [
{
"bbox": [
70,
325,
542,
339
],
"score": 1.0,
"content": "The use of temporal convolutions lowers memory and computation costs over temporal attention—this is",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
69,
338,
541,
351
],
"spans": [
{
"bbox": [
69,
338,
541,
351
],
"score": 1.0,
"content": "crucial because the very purpose of the TSR and SSR models is to operate at high frame rates and spatial",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
349,
542,
362
],
"spans": [
{
"bbox": [
69,
349,
542,
362
],
"score": 1.0,
"content": "resolutions. In our initial experiments, we did not find any significant improvements when using temporal",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
70,
362,
541,
374
],
"spans": [
{
"bbox": [
70,
362,
541,
374
],
"score": 1.0,
"content": "attention over temporal convolutions in our SSR and TSR models, which we hypothesize is due to the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
373,
513,
386
],
"spans": [
{
"bbox": [
69,
373,
513,
386
],
"score": 1.0,
"content": "significant amount of temporal correlation already present in the conditioning input to these models.",
"type": "text"
}
],
"index": 14
}
],
"index": 11
},
{
"type": "text",
"bbox": [
71,
391,
540,
475
],
"lines": [
{
"bbox": [
71,
392,
540,
403
],
"spans": [
{
"bbox": [
71,
392,
540,
403
],
"score": 1.0,
"content": "Our models also use spatial attention and spatial convolutions. The base model and the first two spatial",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
68,
403,
542,
417
],
"spans": [
{
"bbox": [
68,
403,
542,
417
],
"score": 1.0,
"content": "super-resolution models have spatial attention in addition to spatial convolutions. We found this to improve",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
415,
542,
428
],
"spans": [
{
"bbox": [
69,
415,
542,
428
],
"score": 1.0,
"content": "sample fidelity. However, as we move to higher resolutions, we switch to fully convolutional architectures, like",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
69,
426,
542,
441
],
"spans": [
{
"bbox": [
69,
426,
446,
441
],
"score": 1.0,
"content": "Saharia et al. (2022b), to minimize memory and compute costs in order to generate 1280",
"type": "text"
},
{
"bbox": [
447,
430,
455,
438
],
"score": 0.73,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
455,
426,
542,
441
],
"score": 1.0,
"content": "768 resolution data.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
71,
439,
540,
451
],
"spans": [
{
"bbox": [
71,
439,
540,
451
],
"score": 1.0,
"content": "The highest resolution SSR model in our pipeline is a fully convolutional model trained on random lower",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
70,
451,
541,
465
],
"spans": [
{
"bbox": [
70,
451,
541,
465
],
"score": 1.0,
"content": "resolution spatial crops for training time memory efficiency, and we find that the model easily generalizes to",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
462,
249,
477
],
"spans": [
{
"bbox": [
69,
462,
249,
477
],
"score": 1.0,
"content": "the full resolution during sampling time.",
"type": "text"
}
],
"index": 21
}
],
"index": 18
},
{
"type": "title",
"bbox": [
71,
489,
151,
501
],
"lines": [
{
"bbox": [
69,
488,
153,
504
],
"spans": [
{
"bbox": [
69,
488,
153,
504
],
"score": 1.0,
"content": "2.4 v-prediction",
"type": "text"
}
],
"index": 22
}
],
"index": 22
},
{
"type": "text",
"bbox": [
71,
511,
540,
594
],
"lines": [
{
"bbox": [
70,
511,
541,
524
],
"spans": [
{
"bbox": [
70,
511,
388,
524
],
"score": 1.0,
"content": "We follow Salimans & Ho (2022) and use v-prediction parameterization (",
"type": "text"
},
{
"bbox": [
388,
514,
455,
523
],
"score": 0.79,
"content": "\\mathbf { v } _ { t } \\equiv \\alpha _ { t } \\mathbf { \\epsilon } \\epsilon - \\sigma _ { t } \\mathbf { x } ,",
"type": "inline_equation"
},
{
"bbox": [
456,
511,
541,
524
],
"score": 1.0,
"content": "for all our models.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
70,
523,
542,
536
],
"spans": [
{
"bbox": [
70,
523,
92,
536
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
93,
528,
100,
533
],
"score": 0.51,
"content": "\\mathbf { v } .",
"type": "inline_equation"
},
{
"bbox": [
100,
523,
542,
536
],
"score": 1.0,
"content": "-parameterization is particularly useful for numerical stability throughout the diffusion process to",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
70,
534,
541,
548
],
"spans": [
{
"bbox": [
70,
534,
541,
548
],
"score": 1.0,
"content": "enable progressive distillation for our models. For models that operate at higher resolution in our pipeline,",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
70,
547,
541,
560
],
"spans": [
{
"bbox": [
70,
547,
541,
560
],
"score": 1.0,
"content": "we also discovered that the v-parameterization avoids color shifting artifacts that are known to affect high",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
558,
542,
572
],
"spans": [
{
"bbox": [
69,
558,
542,
572
],
"score": 1.0,
"content": "resolution diffusion models, and in the video setting it avoids temporal color shifting that sometimes appears",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
70,
570,
541,
584
],
"spans": [
{
"bbox": [
70,
570,
93,
584
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
93,
576,
99,
581
],
"score": 0.88,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
99,
570,
541,
584
],
"score": 1.0,
"content": "-prediction models. Our use of v-parameterization also has the benefit of faster convergence of sample",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
582,
212,
596
],
"spans": [
{
"bbox": [
69,
582,
212,
596
],
"score": 1.0,
"content": "quality metrics: see Section 3.3.",
"type": "text"
}
],
"index": 29
}
],
"index": 26
},
{
"type": "title",
"bbox": [
72,
608,
222,
621
],
"lines": [
{
"bbox": [
69,
606,
224,
624
],
"spans": [
{
"bbox": [
69,
606,
224,
624
],
"score": 1.0,
"content": "2.5 Conditioning Augmentation",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
70,
630,
540,
703
],
"lines": [
{
"bbox": [
70,
631,
541,
643
],
"spans": [
{
"bbox": [
70,
631,
541,
643
],
"score": 1.0,
"content": "We use noise conditioning augmentation (Ho et al., 2022a) for all our temporal and spatial super-resolution",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
70,
643,
541,
655
],
"spans": [
{
"bbox": [
70,
643,
541,
655
],
"score": 1.0,
"content": "models. Noise conditioning augmentation has been found to be critical for cascaded diffusion models for",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
70,
654,
541,
667
],
"spans": [
{
"bbox": [
70,
654,
541,
667
],
"score": 1.0,
"content": "class-conditional generation (Ho et al., 2022a) as well as text-to-image models (Saharia et al., 2022b). In",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
666,
542,
680
],
"spans": [
{
"bbox": [
70,
666,
542,
680
],
"score": 1.0,
"content": "particular, it facilitates parallel training of different models in the cascade, as it reduces the sensitivity to",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
68,
677,
541,
692
],
"spans": [
{
"bbox": [
68,
677,
541,
692
],
"score": 1.0,
"content": "domain gaps between the output of one stage of the cascade and the inputs used in training the subsequent",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
690,
99,
705
],
"spans": [
{
"bbox": [
69,
690,
99,
705
],
"score": 1.0,
"content": "stage.",
"type": "text"
}
],
"index": 36
}
],
"index": 33.5
},
{
"type": "text",
"bbox": [
70,
708,
538,
732
],
"lines": [
{
"bbox": [
69,
708,
541,
721
],
"spans": [
{
"bbox": [
69,
708,
541,
721
],
"score": 1.0,
"content": "Following Ho et al. (2022a), we apply Gaussian noise augmentation with a random signal-to-noise ratio to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
70,
720,
541,
733
],
"spans": [
{
"bbox": [
70,
720,
541,
733
],
"score": 1.0,
"content": "the conditioning input video during training, and this sampled signal-to-noise ratio is provided to the model",
"type": "text"
}
],
"index": 38
}
],
"index": 37.5
}
],
"page_idx": 8,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
36
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
85,
82,
527,
187
],
"blocks": [
{
"type": "image_body",
"bbox": [
85,
82,
527,
187
],
"group_id": 0,
"lines": [
{
"bbox": [
85,
82,
527,
187
],
"spans": [
{
"bbox": [
85,
82,
527,
187
],
"score": 0.964,
"type": "image",
"image_path": "2910e48845f75da3725a555562623a8a34ff8e36072bcf47c79400b54567845e.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
85,
82,
527,
117.0
],
"spans": [],
"index": 0
},
{
"bbox": [
85,
117.0,
527,
152.0
],
"spans": [],
"index": 1
},
{
"bbox": [
85,
152.0,
527,
187.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
70,
219,
540,
279
],
"group_id": 0,
"lines": [
{
"bbox": [
70,
220,
541,
232
],
"spans": [
{
"bbox": [
70,
220,
541,
232
],
"score": 1.0,
"content": "Figure 7: Video U-Net space-time separable block. Spatial operations are performed independently over",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
232,
541,
244
],
"spans": [
{
"bbox": [
70,
232,
541,
244
],
"score": 1.0,
"content": "frames with shared parameters, whereas the temporal operation mixes activations over frames. Our base",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
68,
241,
542,
259
],
"spans": [
{
"bbox": [
68,
241,
542,
259
],
"score": 1.0,
"content": "model uses spatial convolutions, spatial self-attention and temporal self-attention. For memory efficiency,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
255,
541,
269
],
"spans": [
{
"bbox": [
69,
255,
541,
269
],
"score": 1.0,
"content": "our spatial and temporal super-resolution models use temporal convolutions instead of attention, and our",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
69,
267,
377,
281
],
"spans": [
{
"bbox": [
69,
267,
377,
281
],
"score": 1.0,
"content": "models at the highest spatial resolution do not have spatial attention.",
"type": "text"
}
],
"index": 7
}
],
"index": 5
}
],
"index": 3.0
},
{
"type": "text",
"bbox": [
71,
301,
540,
385
],
"lines": [],
"index": 11,
"bbox_fs": [
69,
300,
542,
386
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
71,
391,
540,
475
],
"lines": [
{
"bbox": [
71,
392,
540,
403
],
"spans": [
{
"bbox": [
71,
392,
540,
403
],
"score": 1.0,
"content": "Our models also use spatial attention and spatial convolutions. The base model and the first two spatial",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
68,
403,
542,
417
],
"spans": [
{
"bbox": [
68,
403,
542,
417
],
"score": 1.0,
"content": "super-resolution models have spatial attention in addition to spatial convolutions. We found this to improve",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
415,
542,
428
],
"spans": [
{
"bbox": [
69,
415,
542,
428
],
"score": 1.0,
"content": "sample fidelity. However, as we move to higher resolutions, we switch to fully convolutional architectures, like",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
69,
426,
542,
441
],
"spans": [
{
"bbox": [
69,
426,
446,
441
],
"score": 1.0,
"content": "Saharia et al. (2022b), to minimize memory and compute costs in order to generate 1280",
"type": "text"
},
{
"bbox": [
447,
430,
455,
438
],
"score": 0.73,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
455,
426,
542,
441
],
"score": 1.0,
"content": "768 resolution data.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
71,
439,
540,
451
],
"spans": [
{
"bbox": [
71,
439,
540,
451
],
"score": 1.0,
"content": "The highest resolution SSR model in our pipeline is a fully convolutional model trained on random lower",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
70,
451,
541,
465
],
"spans": [
{
"bbox": [
70,
451,
541,
465
],
"score": 1.0,
"content": "resolution spatial crops for training time memory efficiency, and we find that the model easily generalizes to",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
462,
249,
477
],
"spans": [
{
"bbox": [
69,
462,
249,
477
],
"score": 1.0,
"content": "the full resolution during sampling time.",
"type": "text"
}
],
"index": 21
}
],
"index": 18,
"bbox_fs": [
68,
392,
542,
477
]
},
{
"type": "title",
"bbox": [
71,
489,
151,
501
],
"lines": [
{
"bbox": [
69,
488,
153,
504
],
"spans": [
{
"bbox": [
69,
488,
153,
504
],
"score": 1.0,
"content": "2.4 v-prediction",
"type": "text"
}
],
"index": 22
}
],
"index": 22
},
{
"type": "text",
"bbox": [
71,
511,
540,
594
],
"lines": [
{
"bbox": [
70,
511,
541,
524
],
"spans": [
{
"bbox": [
70,
511,
388,
524
],
"score": 1.0,
"content": "We follow Salimans & Ho (2022) and use v-prediction parameterization (",
"type": "text"
},
{
"bbox": [
388,
514,
455,
523
],
"score": 0.79,
"content": "\\mathbf { v } _ { t } \\equiv \\alpha _ { t } \\mathbf { \\epsilon } \\epsilon - \\sigma _ { t } \\mathbf { x } ,",
"type": "inline_equation"
},
{
"bbox": [
456,
511,
541,
524
],
"score": 1.0,
"content": "for all our models.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
70,
523,
542,
536
],
"spans": [
{
"bbox": [
70,
523,
92,
536
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
93,
528,
100,
533
],
"score": 0.51,
"content": "\\mathbf { v } .",
"type": "inline_equation"
},
{
"bbox": [
100,
523,
542,
536
],
"score": 1.0,
"content": "-parameterization is particularly useful for numerical stability throughout the diffusion process to",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
70,
534,
541,
548
],
"spans": [
{
"bbox": [
70,
534,
541,
548
],
"score": 1.0,
"content": "enable progressive distillation for our models. For models that operate at higher resolution in our pipeline,",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
70,
547,
541,
560
],
"spans": [
{
"bbox": [
70,
547,
541,
560
],
"score": 1.0,
"content": "we also discovered that the v-parameterization avoids color shifting artifacts that are known to affect high",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
558,
542,
572
],
"spans": [
{
"bbox": [
69,
558,
542,
572
],
"score": 1.0,
"content": "resolution diffusion models, and in the video setting it avoids temporal color shifting that sometimes appears",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
70,
570,
541,
584
],
"spans": [
{
"bbox": [
70,
570,
93,
584
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
93,
576,
99,
581
],
"score": 0.88,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
99,
570,
541,
584
],
"score": 1.0,
"content": "-prediction models. Our use of v-parameterization also has the benefit of faster convergence of sample",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
582,
212,
596
],
"spans": [
{
"bbox": [
69,
582,
212,
596
],
"score": 1.0,
"content": "quality metrics: see Section 3.3.",
"type": "text"
}
],
"index": 29
}
],
"index": 26,
"bbox_fs": [
69,
511,
542,
596
]
},
{
"type": "title",
"bbox": [
72,
608,
222,
621
],
"lines": [
{
"bbox": [
69,
606,
224,
624
],
"spans": [
{
"bbox": [
69,
606,
224,
624
],
"score": 1.0,
"content": "2.5 Conditioning Augmentation",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
70,
630,
540,
703
],
"lines": [
{
"bbox": [
70,
631,
541,
643
],
"spans": [
{
"bbox": [
70,
631,
541,
643
],
"score": 1.0,
"content": "We use noise conditioning augmentation (Ho et al., 2022a) for all our temporal and spatial super-resolution",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
70,
643,
541,
655
],
"spans": [
{
"bbox": [
70,
643,
541,
655
],
"score": 1.0,
"content": "models. Noise conditioning augmentation has been found to be critical for cascaded diffusion models for",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
70,
654,
541,
667
],
"spans": [
{
"bbox": [
70,
654,
541,
667
],
"score": 1.0,
"content": "class-conditional generation (Ho et al., 2022a) as well as text-to-image models (Saharia et al., 2022b). In",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
666,
542,
680
],
"spans": [
{
"bbox": [
70,
666,
542,
680
],
"score": 1.0,
"content": "particular, it facilitates parallel training of different models in the cascade, as it reduces the sensitivity to",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
68,
677,
541,
692
],
"spans": [
{
"bbox": [
68,
677,
541,
692
],
"score": 1.0,
"content": "domain gaps between the output of one stage of the cascade and the inputs used in training the subsequent",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
690,
99,
705
],
"spans": [
{
"bbox": [
69,
690,
99,
705
],
"score": 1.0,
"content": "stage.",
"type": "text"
}
],
"index": 36
}
],
"index": 33.5,
"bbox_fs": [
68,
631,
542,
705
]
},
{
"type": "text",
"bbox": [
70,
708,
538,
732
],
"lines": [
{
"bbox": [
69,
708,
541,
721
],
"spans": [
{
"bbox": [
69,
708,
541,
721
],
"score": 1.0,
"content": "Following Ho et al. (2022a), we apply Gaussian noise augmentation with a random signal-to-noise ratio to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
70,
720,
541,
733
],
"spans": [
{
"bbox": [
70,
720,
541,
733
],
"score": 1.0,
"content": "the conditioning input video during training, and this sampled signal-to-noise ratio is provided to the model",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
82,
541,
94
],
"spans": [
{
"bbox": [
69,
82,
541,
94
],
"score": 1.0,
"content": "as well. At sampling time we use a fixed signal-to-noise ratio such as 3 or 5, representing a small amount of",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
68,
93,
541,
108
],
"spans": [
{
"bbox": [
68,
93,
541,
108
],
"score": 1.0,
"content": "augmentation that aids in removing artifacts in the samples from the previous stage while preserving most",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
69,
105,
144,
119
],
"spans": [
{
"bbox": [
69,
105,
144,
119
],
"score": 1.0,
"content": "of the structure.",
"type": "text",
"cross_page": true
}
],
"index": 2
}
],
"index": 37.5,
"bbox_fs": [
69,
708,
541,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
118
],
"lines": [
{
"bbox": [
69,
82,
541,
94
],
"spans": [
{
"bbox": [
69,
82,
541,
94
],
"score": 1.0,
"content": "as well. At sampling time we use a fixed signal-to-noise ratio such as 3 or 5, representing a small amount of",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
68,
93,
541,
108
],
"spans": [
{
"bbox": [
68,
93,
541,
108
],
"score": 1.0,
"content": "augmentation that aids in removing artifacts in the samples from the previous stage while preserving most",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
69,
105,
144,
119
],
"spans": [
{
"bbox": [
69,
105,
144,
119
],
"score": 1.0,
"content": "of the structure.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "title",
"bbox": [
72,
131,
222,
144
],
"lines": [
{
"bbox": [
69,
129,
224,
147
],
"spans": [
{
"bbox": [
69,
129,
224,
147
],
"score": 1.0,
"content": "2.6 Video-Image Joint Training",
"type": "text"
}
],
"index": 3
}
],
"index": 3
},
{
"type": "text",
"bbox": [
70,
153,
540,
297
],
"lines": [
{
"bbox": [
69,
152,
542,
167
],
"spans": [
{
"bbox": [
69,
152,
542,
167
],
"score": 1.0,
"content": "We follow Ho et al. (2022b) in jointly training all the models in the Imagen Video pipeline on images",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
164,
541,
179
],
"spans": [
{
"bbox": [
69,
164,
541,
179
],
"score": 1.0,
"content": "and videos. During training, individual images are treated as single frame videos. We achieve this by",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
68,
176,
542,
191
],
"spans": [
{
"bbox": [
68,
176,
542,
191
],
"score": 1.0,
"content": "packing individual independent images into a sequence of the same length as a video, and bypass the",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
69,
189,
540,
202
],
"spans": [
{
"bbox": [
69,
189,
540,
202
],
"score": 1.0,
"content": "temporal convolution residual blocks by masking out their computation path. Similarly, we disable cross-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
201,
542,
215
],
"spans": [
{
"bbox": [
69,
201,
542,
215
],
"score": 1.0,
"content": "frame temporal attention by applying masking to the temporal attention maps. This strategy allows us to",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
212,
542,
227
],
"spans": [
{
"bbox": [
69,
212,
542,
227
],
"score": 1.0,
"content": "use to train our video models on image-text datasets that are significantly larger and more diverse than",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
224,
541,
238
],
"spans": [
{
"bbox": [
69,
224,
541,
238
],
"score": 1.0,
"content": "available video-text datasets. Consistent with Ho et al. (2022b), we observe that joint training with images",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
237,
541,
250
],
"spans": [
{
"bbox": [
70,
237,
541,
250
],
"score": 1.0,
"content": "significantly increases the overall quality of video samples. Another interesting artifact of joint training is the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
70,
249,
541,
262
],
"spans": [
{
"bbox": [
70,
249,
541,
262
],
"score": 1.0,
"content": "knowledge transfer from images to videos. For instance, while training on natural video data only enables",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
260,
542,
275
],
"spans": [
{
"bbox": [
69,
260,
542,
275
],
"score": 1.0,
"content": "the model to learn dynamics in natural settings, the model can learn about different image styles (such as",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
70,
273,
541,
285
],
"spans": [
{
"bbox": [
70,
273,
541,
285
],
"score": 1.0,
"content": "sketch, painting, etc.) by training on images. As a result, this joint training enables the model to generate",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
284,
400,
297
],
"spans": [
{
"bbox": [
69,
284,
400,
297
],
"score": 1.0,
"content": "interesting video dynamics in different styles. See Fig. 8 for such examples.",
"type": "text"
}
],
"index": 15
}
],
"index": 9.5
},
{
"type": "title",
"bbox": [
72,
308,
214,
321
],
"lines": [
{
"bbox": [
70,
308,
216,
322
],
"spans": [
{
"bbox": [
70,
308,
216,
322
],
"score": 1.0,
"content": "2.6.1 Classifier Free Guidance",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
71,
328,
540,
365
],
"lines": [
{
"bbox": [
70,
328,
541,
342
],
"spans": [
{
"bbox": [
70,
328,
541,
342
],
"score": 1.0,
"content": "We found classifier free guidance (Ho & Salimans, 2021) to be critical for generating high fidelity samples",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
71,
341,
540,
353
],
"spans": [
{
"bbox": [
71,
341,
540,
353
],
"score": 1.0,
"content": "which respect a given text prompt. This is consistent with earlier results on text-to-image models (Nichol",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
352,
385,
366
],
"spans": [
{
"bbox": [
69,
352,
385,
366
],
"score": 1.0,
"content": "et al., 2021; Ramesh et al., 2022; Saharia et al., 2022b; Yu et al., 2022).",
"type": "text"
}
],
"index": 19
}
],
"index": 18
},
{
"type": "text",
"bbox": [
70,
370,
540,
419
],
"lines": [
{
"bbox": [
69,
370,
542,
384
],
"spans": [
{
"bbox": [
69,
370,
272,
384
],
"score": 1.0,
"content": "In the conditional generation setting, the data",
"type": "text"
},
{
"bbox": [
273,
376,
279,
380
],
"score": 0.84,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
279,
370,
435,
384
],
"score": 1.0,
"content": "is generated conditional on a signal",
"type": "text"
},
{
"bbox": [
436,
376,
441,
380
],
"score": 0.62,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
442,
370,
542,
384
],
"score": 1.0,
"content": ", which here represents",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
67,
381,
542,
398
],
"spans": [
{
"bbox": [
67,
381,
542,
398
],
"score": 1.0,
"content": "a contextualized embedding of the text prompt, and a conditional diffusion model can be trained by using",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
393,
541,
408
],
"spans": [
{
"bbox": [
69,
393,
119,
408
],
"score": 1.0,
"content": "this signal",
"type": "text"
},
{
"bbox": [
119,
400,
124,
405
],
"score": 0.72,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
125,
393,
331,
408
],
"score": 1.0,
"content": "as an additional input to the denoising model",
"type": "text"
},
{
"bbox": [
331,
397,
368,
407
],
"score": 0.95,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
368,
393,
541,
408
],
"score": 1.0,
"content": ". After training, Ho & Salimans (2021)",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
405,
486,
421
],
"spans": [
{
"bbox": [
69,
405,
420,
421
],
"score": 1.0,
"content": "find that sample quality can be improved by adjusting the denoising prediction",
"type": "text"
},
{
"bbox": [
420,
408,
457,
419
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
457,
405,
486,
421
],
"score": 1.0,
"content": "using",
"type": "text"
}
],
"index": 23
}
],
"index": 21.5
},
{
"type": "interline_equation",
"bbox": [
221,
428,
390,
440
],
"lines": [
{
"bbox": [
221,
428,
390,
440
],
"spans": [
{
"bbox": [
221,
428,
390,
440
],
"score": 0.91,
"content": "\\widetilde { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\widehat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - w \\widehat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ,",
"type": "interline_equation",
"image_path": "06f3a1abdc1bdf6b554de868a970a4328c5556b6f081c78cb6270b02b7c614fc.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
221,
428,
390,
440
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
71,
448,
540,
522
],
"lines": [
{
"bbox": [
69,
447,
542,
463
],
"spans": [
{
"bbox": [
69,
447,
101,
463
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
101,
454,
109,
459
],
"score": 0.88,
"content": "w",
"type": "inline_equation"
},
{
"bbox": [
109,
447,
228,
463
],
"score": 1.0,
"content": "is the guidance strength,",
"type": "text"
},
{
"bbox": [
228,
451,
265,
461
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
265,
447,
407,
463
],
"score": 1.0,
"content": "is the conditional model, and",
"type": "text"
},
{
"bbox": [
407,
451,
511,
461
],
"score": 0.91,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ~ = ~ \\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ~ = ~ \\varnothing )",
"type": "inline_equation"
},
{
"bbox": [
512,
447,
542,
463
],
"score": 1.0,
"content": "is an",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
68,
459,
542,
476
],
"spans": [
{
"bbox": [
68,
459,
542,
476
],
"score": 1.0,
"content": "unconditional model. The unconditional model is jointly trained with the conditional model by dropping",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
472,
541,
487
],
"spans": [
{
"bbox": [
69,
472,
187,
487
],
"score": 1.0,
"content": "out the conditioning input",
"type": "text"
},
{
"bbox": [
187,
478,
192,
483
],
"score": 0.82,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
192,
472,
407,
487
],
"score": 1.0,
"content": ". The predictions of the adjusted denoising model",
"type": "text"
},
{
"bbox": [
408,
475,
444,
485
],
"score": 0.94,
"content": "\\tilde { \\bf x } _ { \\theta } ( { \\bf z } _ { t } , { \\bf c } )",
"type": "inline_equation"
},
{
"bbox": [
445,
472,
541,
487
],
"score": 1.0,
"content": "are clipped to respect",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
70,
485,
540,
497
],
"spans": [
{
"bbox": [
70,
485,
540,
497
],
"score": 1.0,
"content": "the range of possible pixel values, which we discuss in more detail in the next section. Note that the linear",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
496,
541,
510
],
"spans": [
{
"bbox": [
69,
496,
360,
510
],
"score": 1.0,
"content": "transformation in Equation 5 can equivalently be performed in",
"type": "text"
},
{
"bbox": [
361,
502,
367,
506
],
"score": 0.57,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
367,
496,
399,
510
],
"score": 1.0,
"content": "-space",
"type": "text"
},
{
"bbox": [
399,
498,
541,
509
],
"score": 0.91,
"content": "\\left( \\tilde { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\hat { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - \\right.",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
71,
508,
338,
522
],
"spans": [
{
"bbox": [
71,
510,
108,
521
],
"score": 0.91,
"content": "w \\hat { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
109,
508,
126,
522
],
"score": 1.0,
"content": ") or",
"type": "text"
},
{
"bbox": [
127,
514,
132,
518
],
"score": 0.86,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
132,
508,
164,
522
],
"score": 1.0,
"content": "-space",
"type": "text"
},
{
"bbox": [
164,
510,
333,
521
],
"score": 0.91,
"content": "\\tilde { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\hat { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - w \\hat { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } ) )",
"type": "inline_equation"
},
{
"bbox": [
334,
508,
338,
522
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 30
}
],
"index": 27.5
},
{
"type": "text",
"bbox": [
71,
526,
540,
586
],
"lines": [
{
"bbox": [
69,
525,
541,
541
],
"spans": [
{
"bbox": [
69,
525,
90,
541
],
"score": 1.0,
"content": "For",
"type": "text"
},
{
"bbox": [
90,
529,
119,
537
],
"score": 0.91,
"content": "w > 0",
"type": "inline_equation"
},
{
"bbox": [
120,
525,
531,
541
],
"score": 1.0,
"content": "this adjustment has the effect of over-emphasizing the effect of conditioning on the signal",
"type": "text"
},
{
"bbox": [
531,
532,
537,
537
],
"score": 0.66,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
537,
525,
541,
541
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
537,
541,
552
],
"spans": [
{
"bbox": [
69,
537,
541,
552
],
"score": 1.0,
"content": "which tends to produce samples of lower diversity but higher quality compared to sampling from the regular",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
550,
541,
564
],
"spans": [
{
"bbox": [
69,
550,
541,
564
],
"score": 1.0,
"content": "conditional model (Ho & Salimans, 2021). The method can be interpreted as a way to guide the samples",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
562,
542,
576
],
"spans": [
{
"bbox": [
70,
562,
259,
576
],
"score": 1.0,
"content": "towards areas where an implicit classifier",
"type": "text"
},
{
"bbox": [
259,
564,
289,
575
],
"score": 0.94,
"content": "p ( \\mathbf { c } | \\mathbf { z } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
289,
562,
542,
576
],
"score": 1.0,
"content": "has high likelihood; as such, it is an adaptation of the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
574,
397,
588
],
"spans": [
{
"bbox": [
70,
574,
322,
588
],
"score": 1.0,
"content": "explicit classifier guidance method proposed by Dhariwal",
"type": "text"
},
{
"bbox": [
323,
577,
331,
585
],
"score": 0.39,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
331,
574,
397,
588
],
"score": 1.0,
"content": "Nichol (2022).",
"type": "text"
}
],
"index": 35
}
],
"index": 33
},
{
"type": "title",
"bbox": [
71,
598,
217,
610
],
"lines": [
{
"bbox": [
69,
596,
218,
613
],
"spans": [
{
"bbox": [
69,
596,
218,
613
],
"score": 1.0,
"content": "2.6.2 Large Guidance Weights",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
70,
618,
540,
703
],
"lines": [
{
"bbox": [
70,
618,
542,
632
],
"spans": [
{
"bbox": [
70,
618,
290,
632
],
"score": 1.0,
"content": "When using large guidance weights, the resulting",
"type": "text"
},
{
"bbox": [
290,
621,
327,
631
],
"score": 0.94,
"content": "\\tilde { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
327,
618,
542,
632
],
"score": 1.0,
"content": "must be projected back to the possible range of",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
542,
644
],
"score": 1.0,
"content": "pixel values at every sampling step to prevent train-test mismatch. When using large guidance weights, the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
68,
642,
542,
656
],
"spans": [
{
"bbox": [
68,
642,
542,
656
],
"score": 1.0,
"content": "standard approach, i.e., clipping the values to the right range (e.g., np.clip(x, -1, 1)), leads to significant",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
69,
654,
541,
668
],
"spans": [
{
"bbox": [
69,
654,
541,
668
],
"score": 1.0,
"content": "saturation artifacts in the generated videos. A similar effect was observed in Saharia et al. (2022b) for text-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
69,
666,
541,
680
],
"spans": [
{
"bbox": [
69,
666,
541,
680
],
"score": 1.0,
"content": "to-image generation. Saharia et al. (2022b) use dynamic thresholding to alleviate this saturation issue.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
678,
541,
692
],
"spans": [
{
"bbox": [
70,
678,
480,
692
],
"score": 1.0,
"content": "Specifically, dynamic clipping involves clipping the image to a dynamically chosen threshold",
"type": "text"
},
{
"bbox": [
481,
684,
486,
688
],
"score": 0.26,
"content": "\\mathbf { s }",
"type": "inline_equation"
},
{
"bbox": [
486,
678,
541,
692
],
"score": 1.0,
"content": "followed by",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
690,
368,
704
],
"spans": [
{
"bbox": [
69,
690,
118,
704
],
"score": 1.0,
"content": "scaling by",
"type": "text"
},
{
"bbox": [
118,
695,
123,
700
],
"score": 0.62,
"content": "\\tt s",
"type": "inline_equation"
},
{
"bbox": [
123,
690,
368,
704
],
"score": 1.0,
"content": "(i.e., np.clip(x, -s, s) / s) (Saharia et al., 2022b).",
"type": "text"
}
],
"index": 43
}
],
"index": 40
},
{
"type": "text",
"bbox": [
70,
708,
537,
732
],
"lines": [
{
"bbox": [
70,
708,
538,
721
],
"spans": [
{
"bbox": [
70,
708,
538,
721
],
"score": 1.0,
"content": "Although dynamic clipping can help with over-saturation, we did not find it sufficient in initial experiments.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
70,
720,
538,
732
],
"spans": [
{
"bbox": [
70,
720,
261,
732
],
"score": 1.0,
"content": "We therefore also experiment with letting",
"type": "text"
},
{
"bbox": [
262,
725,
269,
730
],
"score": 0.88,
"content": "w",
"type": "inline_equation"
},
{
"bbox": [
270,
720,
538,
732
],
"score": 1.0,
"content": "oscillate between a high and a low guidance weight at each",
"type": "text"
}
],
"index": 45
}
],
"index": 44.5
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
70,
25,
239,
38
],
"spans": [
{
"bbox": [
70,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "10",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
118
],
"lines": [],
"index": 1,
"bbox_fs": [
68,
82,
541,
119
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
72,
131,
222,
144
],
"lines": [
{
"bbox": [
69,
129,
224,
147
],
"spans": [
{
"bbox": [
69,
129,
224,
147
],
"score": 1.0,
"content": "2.6 Video-Image Joint Training",
"type": "text"
}
],
"index": 3
}
],
"index": 3
},
{
"type": "text",
"bbox": [
70,
153,
540,
297
],
"lines": [
{
"bbox": [
69,
152,
542,
167
],
"spans": [
{
"bbox": [
69,
152,
542,
167
],
"score": 1.0,
"content": "We follow Ho et al. (2022b) in jointly training all the models in the Imagen Video pipeline on images",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
164,
541,
179
],
"spans": [
{
"bbox": [
69,
164,
541,
179
],
"score": 1.0,
"content": "and videos. During training, individual images are treated as single frame videos. We achieve this by",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
68,
176,
542,
191
],
"spans": [
{
"bbox": [
68,
176,
542,
191
],
"score": 1.0,
"content": "packing individual independent images into a sequence of the same length as a video, and bypass the",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
69,
189,
540,
202
],
"spans": [
{
"bbox": [
69,
189,
540,
202
],
"score": 1.0,
"content": "temporal convolution residual blocks by masking out their computation path. Similarly, we disable cross-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
201,
542,
215
],
"spans": [
{
"bbox": [
69,
201,
542,
215
],
"score": 1.0,
"content": "frame temporal attention by applying masking to the temporal attention maps. This strategy allows us to",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
212,
542,
227
],
"spans": [
{
"bbox": [
69,
212,
542,
227
],
"score": 1.0,
"content": "use to train our video models on image-text datasets that are significantly larger and more diverse than",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
224,
541,
238
],
"spans": [
{
"bbox": [
69,
224,
541,
238
],
"score": 1.0,
"content": "available video-text datasets. Consistent with Ho et al. (2022b), we observe that joint training with images",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
237,
541,
250
],
"spans": [
{
"bbox": [
70,
237,
541,
250
],
"score": 1.0,
"content": "significantly increases the overall quality of video samples. Another interesting artifact of joint training is the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
70,
249,
541,
262
],
"spans": [
{
"bbox": [
70,
249,
541,
262
],
"score": 1.0,
"content": "knowledge transfer from images to videos. For instance, while training on natural video data only enables",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
260,
542,
275
],
"spans": [
{
"bbox": [
69,
260,
542,
275
],
"score": 1.0,
"content": "the model to learn dynamics in natural settings, the model can learn about different image styles (such as",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
70,
273,
541,
285
],
"spans": [
{
"bbox": [
70,
273,
541,
285
],
"score": 1.0,
"content": "sketch, painting, etc.) by training on images. As a result, this joint training enables the model to generate",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
284,
400,
297
],
"spans": [
{
"bbox": [
69,
284,
400,
297
],
"score": 1.0,
"content": "interesting video dynamics in different styles. See Fig. 8 for such examples.",
"type": "text"
}
],
"index": 15
}
],
"index": 9.5,
"bbox_fs": [
68,
152,
542,
297
]
},
{
"type": "title",
"bbox": [
72,
308,
214,
321
],
"lines": [
{
"bbox": [
70,
308,
216,
322
],
"spans": [
{
"bbox": [
70,
308,
216,
322
],
"score": 1.0,
"content": "2.6.1 Classifier Free Guidance",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
71,
328,
540,
365
],
"lines": [
{
"bbox": [
70,
328,
541,
342
],
"spans": [
{
"bbox": [
70,
328,
541,
342
],
"score": 1.0,
"content": "We found classifier free guidance (Ho & Salimans, 2021) to be critical for generating high fidelity samples",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
71,
341,
540,
353
],
"spans": [
{
"bbox": [
71,
341,
540,
353
],
"score": 1.0,
"content": "which respect a given text prompt. This is consistent with earlier results on text-to-image models (Nichol",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
352,
385,
366
],
"spans": [
{
"bbox": [
69,
352,
385,
366
],
"score": 1.0,
"content": "et al., 2021; Ramesh et al., 2022; Saharia et al., 2022b; Yu et al., 2022).",
"type": "text"
}
],
"index": 19
}
],
"index": 18,
"bbox_fs": [
69,
328,
541,
366
]
},
{
"type": "text",
"bbox": [
70,
370,
540,
419
],
"lines": [
{
"bbox": [
69,
370,
542,
384
],
"spans": [
{
"bbox": [
69,
370,
272,
384
],
"score": 1.0,
"content": "In the conditional generation setting, the data",
"type": "text"
},
{
"bbox": [
273,
376,
279,
380
],
"score": 0.84,
"content": "\\mathbf { x }",
"type": "inline_equation"
},
{
"bbox": [
279,
370,
435,
384
],
"score": 1.0,
"content": "is generated conditional on a signal",
"type": "text"
},
{
"bbox": [
436,
376,
441,
380
],
"score": 0.62,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
442,
370,
542,
384
],
"score": 1.0,
"content": ", which here represents",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
67,
381,
542,
398
],
"spans": [
{
"bbox": [
67,
381,
542,
398
],
"score": 1.0,
"content": "a contextualized embedding of the text prompt, and a conditional diffusion model can be trained by using",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
393,
541,
408
],
"spans": [
{
"bbox": [
69,
393,
119,
408
],
"score": 1.0,
"content": "this signal",
"type": "text"
},
{
"bbox": [
119,
400,
124,
405
],
"score": 0.72,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
125,
393,
331,
408
],
"score": 1.0,
"content": "as an additional input to the denoising model",
"type": "text"
},
{
"bbox": [
331,
397,
368,
407
],
"score": 0.95,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
368,
393,
541,
408
],
"score": 1.0,
"content": ". After training, Ho & Salimans (2021)",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
405,
486,
421
],
"spans": [
{
"bbox": [
69,
405,
420,
421
],
"score": 1.0,
"content": "find that sample quality can be improved by adjusting the denoising prediction",
"type": "text"
},
{
"bbox": [
420,
408,
457,
419
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
457,
405,
486,
421
],
"score": 1.0,
"content": "using",
"type": "text"
}
],
"index": 23
}
],
"index": 21.5,
"bbox_fs": [
67,
370,
542,
421
]
},
{
"type": "interline_equation",
"bbox": [
221,
428,
390,
440
],
"lines": [
{
"bbox": [
221,
428,
390,
440
],
"spans": [
{
"bbox": [
221,
428,
390,
440
],
"score": 0.91,
"content": "\\widetilde { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\widehat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - w \\widehat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ,",
"type": "interline_equation",
"image_path": "06f3a1abdc1bdf6b554de868a970a4328c5556b6f081c78cb6270b02b7c614fc.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
221,
428,
390,
440
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
71,
448,
540,
522
],
"lines": [
{
"bbox": [
69,
447,
542,
463
],
"spans": [
{
"bbox": [
69,
447,
101,
463
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
101,
454,
109,
459
],
"score": 0.88,
"content": "w",
"type": "inline_equation"
},
{
"bbox": [
109,
447,
228,
463
],
"score": 1.0,
"content": "is the guidance strength,",
"type": "text"
},
{
"bbox": [
228,
451,
265,
461
],
"score": 0.94,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
265,
447,
407,
463
],
"score": 1.0,
"content": "is the conditional model, and",
"type": "text"
},
{
"bbox": [
407,
451,
511,
461
],
"score": 0.91,
"content": "\\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } ) ~ = ~ \\hat { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ~ = ~ \\varnothing )",
"type": "inline_equation"
},
{
"bbox": [
512,
447,
542,
463
],
"score": 1.0,
"content": "is an",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
68,
459,
542,
476
],
"spans": [
{
"bbox": [
68,
459,
542,
476
],
"score": 1.0,
"content": "unconditional model. The unconditional model is jointly trained with the conditional model by dropping",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
69,
472,
541,
487
],
"spans": [
{
"bbox": [
69,
472,
187,
487
],
"score": 1.0,
"content": "out the conditioning input",
"type": "text"
},
{
"bbox": [
187,
478,
192,
483
],
"score": 0.82,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
192,
472,
407,
487
],
"score": 1.0,
"content": ". The predictions of the adjusted denoising model",
"type": "text"
},
{
"bbox": [
408,
475,
444,
485
],
"score": 0.94,
"content": "\\tilde { \\bf x } _ { \\theta } ( { \\bf z } _ { t } , { \\bf c } )",
"type": "inline_equation"
},
{
"bbox": [
445,
472,
541,
487
],
"score": 1.0,
"content": "are clipped to respect",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
70,
485,
540,
497
],
"spans": [
{
"bbox": [
70,
485,
540,
497
],
"score": 1.0,
"content": "the range of possible pixel values, which we discuss in more detail in the next section. Note that the linear",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
496,
541,
510
],
"spans": [
{
"bbox": [
69,
496,
360,
510
],
"score": 1.0,
"content": "transformation in Equation 5 can equivalently be performed in",
"type": "text"
},
{
"bbox": [
361,
502,
367,
506
],
"score": 0.57,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
367,
496,
399,
510
],
"score": 1.0,
"content": "-space",
"type": "text"
},
{
"bbox": [
399,
498,
541,
509
],
"score": 0.91,
"content": "\\left( \\tilde { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\hat { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - \\right.",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
71,
508,
338,
522
],
"spans": [
{
"bbox": [
71,
510,
108,
521
],
"score": 0.91,
"content": "w \\hat { \\mathbf { v } } _ { \\theta } ( \\mathbf { z } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
109,
508,
126,
522
],
"score": 1.0,
"content": ") or",
"type": "text"
},
{
"bbox": [
127,
514,
132,
518
],
"score": 0.86,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
132,
508,
164,
522
],
"score": 1.0,
"content": "-space",
"type": "text"
},
{
"bbox": [
164,
510,
333,
521
],
"score": 0.91,
"content": "\\tilde { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) = ( 1 + w ) \\hat { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } ) - w \\hat { \\epsilon } _ { \\theta } ( \\mathbf { z } _ { t } ) )",
"type": "inline_equation"
},
{
"bbox": [
334,
508,
338,
522
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 30
}
],
"index": 27.5,
"bbox_fs": [
68,
447,
542,
522
]
},
{
"type": "text",
"bbox": [
71,
526,
540,
586
],
"lines": [
{
"bbox": [
69,
525,
541,
541
],
"spans": [
{
"bbox": [
69,
525,
90,
541
],
"score": 1.0,
"content": "For",
"type": "text"
},
{
"bbox": [
90,
529,
119,
537
],
"score": 0.91,
"content": "w > 0",
"type": "inline_equation"
},
{
"bbox": [
120,
525,
531,
541
],
"score": 1.0,
"content": "this adjustment has the effect of over-emphasizing the effect of conditioning on the signal",
"type": "text"
},
{
"bbox": [
531,
532,
537,
537
],
"score": 0.66,
"content": "\\mathbf { c }",
"type": "inline_equation"
},
{
"bbox": [
537,
525,
541,
541
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
537,
541,
552
],
"spans": [
{
"bbox": [
69,
537,
541,
552
],
"score": 1.0,
"content": "which tends to produce samples of lower diversity but higher quality compared to sampling from the regular",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
550,
541,
564
],
"spans": [
{
"bbox": [
69,
550,
541,
564
],
"score": 1.0,
"content": "conditional model (Ho & Salimans, 2021). The method can be interpreted as a way to guide the samples",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
70,
562,
542,
576
],
"spans": [
{
"bbox": [
70,
562,
259,
576
],
"score": 1.0,
"content": "towards areas where an implicit classifier",
"type": "text"
},
{
"bbox": [
259,
564,
289,
575
],
"score": 0.94,
"content": "p ( \\mathbf { c } | \\mathbf { z } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
289,
562,
542,
576
],
"score": 1.0,
"content": "has high likelihood; as such, it is an adaptation of the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
574,
397,
588
],
"spans": [
{
"bbox": [
70,
574,
322,
588
],
"score": 1.0,
"content": "explicit classifier guidance method proposed by Dhariwal",
"type": "text"
},
{
"bbox": [
323,
577,
331,
585
],
"score": 0.39,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
331,
574,
397,
588
],
"score": 1.0,
"content": "Nichol (2022).",
"type": "text"
}
],
"index": 35
}
],
"index": 33,
"bbox_fs": [
69,
525,
542,
588
]
},
{
"type": "title",
"bbox": [
71,
598,
217,
610
],
"lines": [
{
"bbox": [
69,
596,
218,
613
],
"spans": [
{
"bbox": [
69,
596,
218,
613
],
"score": 1.0,
"content": "2.6.2 Large Guidance Weights",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
70,
618,
540,
703
],
"lines": [
{
"bbox": [
70,
618,
542,
632
],
"spans": [
{
"bbox": [
70,
618,
290,
632
],
"score": 1.0,
"content": "When using large guidance weights, the resulting",
"type": "text"
},
{
"bbox": [
290,
621,
327,
631
],
"score": 0.94,
"content": "\\tilde { \\mathbf { x } } _ { \\theta } ( \\mathbf { z } _ { t } , \\mathbf { c } )",
"type": "inline_equation"
},
{
"bbox": [
327,
618,
542,
632
],
"score": 1.0,
"content": "must be projected back to the possible range of",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
68,
630,
542,
644
],
"spans": [
{
"bbox": [
68,
630,
542,
644
],
"score": 1.0,
"content": "pixel values at every sampling step to prevent train-test mismatch. When using large guidance weights, the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
68,
642,
542,
656
],
"spans": [
{
"bbox": [
68,
642,
542,
656
],
"score": 1.0,
"content": "standard approach, i.e., clipping the values to the right range (e.g., np.clip(x, -1, 1)), leads to significant",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
69,
654,
541,
668
],
"spans": [
{
"bbox": [
69,
654,
541,
668
],
"score": 1.0,
"content": "saturation artifacts in the generated videos. A similar effect was observed in Saharia et al. (2022b) for text-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
69,
666,
541,
680
],
"spans": [
{
"bbox": [
69,
666,
541,
680
],
"score": 1.0,
"content": "to-image generation. Saharia et al. (2022b) use dynamic thresholding to alleviate this saturation issue.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
70,
678,
541,
692
],
"spans": [
{
"bbox": [
70,
678,
480,
692
],
"score": 1.0,
"content": "Specifically, dynamic clipping involves clipping the image to a dynamically chosen threshold",
"type": "text"
},
{
"bbox": [
481,
684,
486,
688
],
"score": 0.26,
"content": "\\mathbf { s }",
"type": "inline_equation"
},
{
"bbox": [
486,
678,
541,
692
],
"score": 1.0,
"content": "followed by",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
69,
690,
368,
704
],
"spans": [
{
"bbox": [
69,
690,
118,
704
],
"score": 1.0,
"content": "scaling by",
"type": "text"
},
{
"bbox": [
118,
695,
123,
700
],
"score": 0.62,
"content": "\\tt s",
"type": "inline_equation"
},
{
"bbox": [
123,
690,
368,
704
],
"score": 1.0,
"content": "(i.e., np.clip(x, -s, s) / s) (Saharia et al., 2022b).",
"type": "text"
}
],
"index": 43
}
],
"index": 40,
"bbox_fs": [
68,
618,
542,
704
]
},
{
"type": "text",
"bbox": [
70,
708,
537,
732
],
"lines": [
{
"bbox": [
70,
708,
538,
721
],
"spans": [
{
"bbox": [
70,
708,
538,
721
],
"score": 1.0,
"content": "Although dynamic clipping can help with over-saturation, we did not find it sufficient in initial experiments.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
70,
720,
538,
732
],
"spans": [
{
"bbox": [
70,
720,
261,
732
],
"score": 1.0,
"content": "We therefore also experiment with letting",
"type": "text"
},
{
"bbox": [
262,
725,
269,
730
],
"score": 0.88,
"content": "w",
"type": "inline_equation"
},
{
"bbox": [
270,
720,
538,
732
],
"score": 1.0,
"content": "oscillate between a high and a low guidance weight at each",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
68,
80,
542,
97
],
"spans": [
{
"bbox": [
68,
80,
542,
97
],
"score": 1.0,
"content": "alternating sampling step, which we find significantly helps with these saturation issues. We call this sampling",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
69,
94,
541,
108
],
"spans": [
{
"bbox": [
69,
94,
541,
108
],
"score": 1.0,
"content": "technique oscillating guidance. Specifically, we use a constant high guidance weight for a certain number",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
69,
105,
541,
119
],
"spans": [
{
"bbox": [
69,
105,
541,
119
],
"score": 1.0,
"content": "of initial sampling steps, followed by oscillation between high and low guidance weights: this oscillation",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
69,
118,
541,
132
],
"spans": [
{
"bbox": [
69,
118,
541,
132
],
"score": 1.0,
"content": "is implemented simply by alternating between a large weight (such as 15) and a small weight (such as 1)",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
68,
128,
542,
145
],
"spans": [
{
"bbox": [
68,
128,
542,
145
],
"score": 1.0,
"content": "over the course of sampling. We hypothesize that a constant high guidance weight at the start of sampling",
"type": "text",
"cross_page": true
}
],
"index": 4
},
{
"bbox": [
69,
142,
541,
155
],
"spans": [
{
"bbox": [
69,
142,
541,
155
],
"score": 1.0,
"content": "helps break modes with heavy emphasis on text, while oscillating between high and low guidance weights",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
69,
154,
542,
167
],
"spans": [
{
"bbox": [
69,
154,
542,
167
],
"score": 1.0,
"content": "helps maintain a strong text alignment (via high guidance sampling step) while limiting saturation artifacts",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
70,
166,
542,
180
],
"spans": [
{
"bbox": [
70,
166,
542,
180
],
"score": 1.0,
"content": "(via low guidance sampling step). We however observed no improvement in sample fidelity and more visual",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
69,
178,
541,
192
],
"spans": [
{
"bbox": [
69,
178,
354,
192
],
"score": 1.0,
"content": "artifacts when applying oscillating guidance to models past the 80",
"type": "text",
"cross_page": true
},
{
"bbox": [
354,
182,
362,
189
],
"score": 0.29,
"content": "\\times",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
362,
178,
541,
192
],
"score": 1.0,
"content": "48 spatial resolution. Thus we only apply",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
70,
190,
337,
202
],
"spans": [
{
"bbox": [
70,
190,
337,
202
],
"score": 1.0,
"content": "oscillating guidance to the base and the first two SR models.",
"type": "text",
"cross_page": true
}
],
"index": 9
}
],
"index": 44.5,
"bbox_fs": [
70,
708,
538,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
201
],
"lines": [
{
"bbox": [
68,
80,
542,
97
],
"spans": [
{
"bbox": [
68,
80,
542,
97
],
"score": 1.0,
"content": "alternating sampling step, which we find significantly helps with these saturation issues. We call this sampling",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
69,
94,
541,
108
],
"spans": [
{
"bbox": [
69,
94,
541,
108
],
"score": 1.0,
"content": "technique oscillating guidance. Specifically, we use a constant high guidance weight for a certain number",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
69,
105,
541,
119
],
"spans": [
{
"bbox": [
69,
105,
541,
119
],
"score": 1.0,
"content": "of initial sampling steps, followed by oscillation between high and low guidance weights: this oscillation",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
69,
118,
541,
132
],
"spans": [
{
"bbox": [
69,
118,
541,
132
],
"score": 1.0,
"content": "is implemented simply by alternating between a large weight (such as 15) and a small weight (such as 1)",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
68,
128,
542,
145
],
"spans": [
{
"bbox": [
68,
128,
542,
145
],
"score": 1.0,
"content": "over the course of sampling. We hypothesize that a constant high guidance weight at the start of sampling",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
142,
541,
155
],
"spans": [
{
"bbox": [
69,
142,
541,
155
],
"score": 1.0,
"content": "helps break modes with heavy emphasis on text, while oscillating between high and low guidance weights",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
154,
542,
167
],
"spans": [
{
"bbox": [
69,
154,
542,
167
],
"score": 1.0,
"content": "helps maintain a strong text alignment (via high guidance sampling step) while limiting saturation artifacts",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
166,
542,
180
],
"spans": [
{
"bbox": [
70,
166,
542,
180
],
"score": 1.0,
"content": "(via low guidance sampling step). We however observed no improvement in sample fidelity and more visual",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
178,
541,
192
],
"spans": [
{
"bbox": [
69,
178,
354,
192
],
"score": 1.0,
"content": "artifacts when applying oscillating guidance to models past the 80",
"type": "text"
},
{
"bbox": [
354,
182,
362,
189
],
"score": 0.29,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
362,
178,
541,
192
],
"score": 1.0,
"content": "48 spatial resolution. Thus we only apply",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
70,
190,
337,
202
],
"spans": [
{
"bbox": [
70,
190,
337,
202
],
"score": 1.0,
"content": "oscillating guidance to the base and the first two SR models.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5
},
{
"type": "title",
"bbox": [
71,
215,
385,
227
],
"lines": [
{
"bbox": [
69,
214,
387,
231
],
"spans": [
{
"bbox": [
69,
214,
387,
231
],
"score": 1.0,
"content": "2.7 Progressive Distillation with Guidance and Stochastic Samplers",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
71,
236,
540,
321
],
"lines": [
{
"bbox": [
70,
237,
541,
250
],
"spans": [
{
"bbox": [
70,
237,
541,
250
],
"score": 1.0,
"content": "Salimans & Ho (2022) proposed progressive distillation to enable fast sampling of diffusion models. This",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
249,
540,
262
],
"spans": [
{
"bbox": [
69,
249,
540,
262
],
"score": 1.0,
"content": "method distills a trained deterministic DDIM sampler (Song et al., 2020) to a diffusion model that takes",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
261,
541,
274
],
"spans": [
{
"bbox": [
69,
261,
541,
274
],
"score": 1.0,
"content": "many fewer sampling steps, without losing much perceptual quality. At each iteration of the distillation",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
273,
541,
286
],
"spans": [
{
"bbox": [
69,
273,
123,
286
],
"score": 1.0,
"content": "process, an",
"type": "text"
},
{
"bbox": [
123,
276,
133,
283
],
"score": 0.91,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
133,
273,
370,
286
],
"score": 1.0,
"content": "-step DDIM sampler is distilled to a new model with",
"type": "text"
},
{
"bbox": [
370,
275,
388,
285
],
"score": 0.84,
"content": "N / 2",
"type": "inline_equation"
},
{
"bbox": [
389,
273,
541,
286
],
"score": 1.0,
"content": "-steps. This procedure is repeated",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
70,
285,
541,
298
],
"spans": [
{
"bbox": [
70,
285,
541,
298
],
"score": 1.0,
"content": "by halving the required sampling steps each iteration. Meng et al. (2022) extend this approach to samplers",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
296,
541,
309
],
"spans": [
{
"bbox": [
69,
296,
541,
309
],
"score": 1.0,
"content": "with guidance, and propose a new stochastic sampler for use with distilled models. Here we show that this",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
310,
296,
322
],
"spans": [
{
"bbox": [
69,
310,
296,
322
],
"score": 1.0,
"content": "approach also works very well for video generation.",
"type": "text"
}
],
"index": 17
}
],
"index": 14
},
{
"type": "text",
"bbox": [
71,
326,
540,
387
],
"lines": [
{
"bbox": [
71,
327,
541,
339
],
"spans": [
{
"bbox": [
71,
327,
541,
339
],
"score": 1.0,
"content": "We use a two-stage distillation approach to distill a DDIM sampler (Song et al., 2020) with classifier-free",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
339,
541,
352
],
"spans": [
{
"bbox": [
69,
339,
541,
352
],
"score": 1.0,
"content": "guidance. At the first stage, we learn a single diffusion model that matches the combined output from",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
69,
350,
542,
364
],
"spans": [
{
"bbox": [
69,
350,
542,
364
],
"score": 1.0,
"content": "the jointly trained conditional and unconditional diffusion models, where the combination coefficients are",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
70,
362,
542,
376
],
"spans": [
{
"bbox": [
70,
362,
542,
376
],
"score": 1.0,
"content": "determined by the guidance weight. Then we apply progressive distillation to that single model to produce",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
374,
328,
388
],
"spans": [
{
"bbox": [
69,
374,
328,
388
],
"score": 1.0,
"content": "models requiring fewer sampling steps at the second stage.",
"type": "text"
}
],
"index": 22
}
],
"index": 20
},
{
"type": "text",
"bbox": [
71,
393,
540,
464
],
"lines": [
{
"bbox": [
70,
392,
541,
406
],
"spans": [
{
"bbox": [
70,
392,
235,
406
],
"score": 1.0,
"content": "After distillation, we use a stochastic",
"type": "text"
},
{
"bbox": [
236,
395,
245,
402
],
"score": 0.91,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
245,
392,
541,
406
],
"score": 1.0,
"content": "-step sampler: At each step, we first apply one deterministic DDIM",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
405,
542,
418
],
"spans": [
{
"bbox": [
69,
405,
363,
418
],
"score": 1.0,
"content": "update with twice the original step size (i.e., the same step size as a",
"type": "text"
},
{
"bbox": [
364,
406,
383,
417
],
"score": 0.91,
"content": "N / 2",
"type": "inline_equation"
},
{
"bbox": [
383,
405,
542,
418
],
"score": 1.0,
"content": "-step sampler), and then we perform",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
416,
542,
430
],
"spans": [
{
"bbox": [
69,
416,
542,
430
],
"score": 1.0,
"content": "one stochastic step backward (i.e., perturbed with noise following the forward diffusion process) with the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
428,
542,
442
],
"spans": [
{
"bbox": [
69,
428,
542,
442
],
"score": 1.0,
"content": "original step size, inspired by Karras et al. (2022). See Meng et al. (2022) for more details. Using this",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
440,
542,
454
],
"spans": [
{
"bbox": [
68,
440,
542,
454
],
"score": 1.0,
"content": "approach, we are able to distill all 7 video diffusion models down to just 8 sampling steps per model without",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
452,
250,
466
],
"spans": [
{
"bbox": [
69,
452,
250,
466
],
"score": 1.0,
"content": "any noticeable loss in perceptual quality.",
"type": "text"
}
],
"index": 28
}
],
"index": 25.5
},
{
"type": "title",
"bbox": [
71,
479,
159,
493
],
"lines": [
{
"bbox": [
68,
477,
161,
497
],
"spans": [
{
"bbox": [
68,
477,
161,
497
],
"score": 1.0,
"content": "3 Experiments",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
70,
505,
540,
625
],
"lines": [
{
"bbox": [
69,
505,
541,
519
],
"spans": [
{
"bbox": [
69,
505,
541,
519
],
"score": 1.0,
"content": "We train our models on a combination of an internal dataset consisting of 14 million video-text pairs and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
516,
541,
531
],
"spans": [
{
"bbox": [
69,
516,
541,
531
],
"score": 1.0,
"content": "60 million image-text pairs, and the publicly available LAION-400M image-text dataset (Schuhmann et al.,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
68,
528,
542,
543
],
"spans": [
{
"bbox": [
68,
528,
542,
543
],
"score": 1.0,
"content": "2021). To process the data into a form suitable for training our cascading pipeline, we spatially resize",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
541,
541,
555
],
"spans": [
{
"bbox": [
69,
541,
541,
555
],
"score": 1.0,
"content": "images and videos using antialiased bilinear resizing, and we temporally resize videos by skipping frames.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
69,
553,
541,
567
],
"spans": [
{
"bbox": [
69,
553,
541,
567
],
"score": 1.0,
"content": "Throughout our model development process, we evaluated Imagen Video on several different metrics, such",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
68,
564,
542,
580
],
"spans": [
{
"bbox": [
68,
564,
542,
580
],
"score": 1.0,
"content": "as FID on individual frames (Heusel et al., 2017), FVD (Unterthiner et al., 2019) for temporal consistency,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
577,
542,
591
],
"spans": [
{
"bbox": [
69,
577,
542,
591
],
"score": 1.0,
"content": "and frame-wise CLIP scores (Hessel et al., 2021; Park et al., 2021) for video-text alignment. Below, we",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
70,
589,
542,
603
],
"spans": [
{
"bbox": [
70,
589,
542,
603
],
"score": 1.0,
"content": "explore the capabilities of our model and investigate its performance in regards to 1) scaling up the number",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
601,
542,
615
],
"spans": [
{
"bbox": [
69,
601,
542,
615
],
"score": 1.0,
"content": "of parameters in our model, 2) changing the parameterization of our model, and 3) distilling our models so",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
612,
221,
627
],
"spans": [
{
"bbox": [
69,
612,
221,
627
],
"score": 1.0,
"content": "that they are fast to sample from.",
"type": "text"
}
],
"index": 39
}
],
"index": 34.5
},
{
"type": "title",
"bbox": [
72,
639,
262,
651
],
"lines": [
{
"bbox": [
69,
638,
263,
653
],
"spans": [
{
"bbox": [
69,
638,
263,
653
],
"score": 1.0,
"content": "3.1 Unique video generation capabilities",
"type": "text"
}
],
"index": 40
}
],
"index": 40
},
{
"type": "text",
"bbox": [
71,
660,
540,
732
],
"lines": [
{
"bbox": [
70,
659,
541,
675
],
"spans": [
{
"bbox": [
70,
659,
541,
675
],
"score": 1.0,
"content": "We find that Imagen Video is capable of generating high fidelity video, and that it possesses several unique",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "capabilities that are not traditionally found in unstructured generative models learned purely from data. For",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
68,
682,
542,
700
],
"spans": [
{
"bbox": [
68,
682,
542,
700
],
"score": 1.0,
"content": "example, Fig. 8 shows that our model is capable of generating videos with artistic styles learned from image",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
69,
695,
541,
710
],
"spans": [
{
"bbox": [
69,
695,
541,
710
],
"score": 1.0,
"content": "information, such as videos in the style of van Gogh paintings or watercolor paintings. Fig. 9 shows that",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
70,
709,
541,
721
],
"spans": [
{
"bbox": [
70,
709,
541,
721
],
"score": 1.0,
"content": "Imagen Video possesses an understanding of 3D structure, as it is capable of generating videos of objects",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
69,
720,
541,
734
],
"spans": [
{
"bbox": [
69,
720,
541,
734
],
"score": 1.0,
"content": "rotating while roughly preserving structure. While the 3D consistency over the course of rotation is not",
"type": "text"
}
],
"index": 46
}
],
"index": 43.5
}
],
"page_idx": 10,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
36
],
"lines": [
{
"bbox": [
70,
25,
238,
38
],
"spans": [
{
"bbox": [
70,
25,
238,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
70,
82,
540,
201
],
"lines": [],
"index": 4.5,
"bbox_fs": [
68,
80,
542,
202
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
71,
215,
385,
227
],
"lines": [
{
"bbox": [
69,
214,
387,
231
],
"spans": [
{
"bbox": [
69,
214,
387,
231
],
"score": 1.0,
"content": "2.7 Progressive Distillation with Guidance and Stochastic Samplers",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
71,
236,
540,
321
],
"lines": [
{
"bbox": [
70,
237,
541,
250
],
"spans": [
{
"bbox": [
70,
237,
541,
250
],
"score": 1.0,
"content": "Salimans & Ho (2022) proposed progressive distillation to enable fast sampling of diffusion models. This",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
249,
540,
262
],
"spans": [
{
"bbox": [
69,
249,
540,
262
],
"score": 1.0,
"content": "method distills a trained deterministic DDIM sampler (Song et al., 2020) to a diffusion model that takes",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
261,
541,
274
],
"spans": [
{
"bbox": [
69,
261,
541,
274
],
"score": 1.0,
"content": "many fewer sampling steps, without losing much perceptual quality. At each iteration of the distillation",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
273,
541,
286
],
"spans": [
{
"bbox": [
69,
273,
123,
286
],
"score": 1.0,
"content": "process, an",
"type": "text"
},
{
"bbox": [
123,
276,
133,
283
],
"score": 0.91,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
133,
273,
370,
286
],
"score": 1.0,
"content": "-step DDIM sampler is distilled to a new model with",
"type": "text"
},
{
"bbox": [
370,
275,
388,
285
],
"score": 0.84,
"content": "N / 2",
"type": "inline_equation"
},
{
"bbox": [
389,
273,
541,
286
],
"score": 1.0,
"content": "-steps. This procedure is repeated",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
70,
285,
541,
298
],
"spans": [
{
"bbox": [
70,
285,
541,
298
],
"score": 1.0,
"content": "by halving the required sampling steps each iteration. Meng et al. (2022) extend this approach to samplers",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
296,
541,
309
],
"spans": [
{
"bbox": [
69,
296,
541,
309
],
"score": 1.0,
"content": "with guidance, and propose a new stochastic sampler for use with distilled models. Here we show that this",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
310,
296,
322
],
"spans": [
{
"bbox": [
69,
310,
296,
322
],
"score": 1.0,
"content": "approach also works very well for video generation.",
"type": "text"
}
],
"index": 17
}
],
"index": 14,
"bbox_fs": [
69,
237,
541,
322
]
},
{
"type": "text",
"bbox": [
71,
326,
540,
387
],
"lines": [
{
"bbox": [
71,
327,
541,
339
],
"spans": [
{
"bbox": [
71,
327,
541,
339
],
"score": 1.0,
"content": "We use a two-stage distillation approach to distill a DDIM sampler (Song et al., 2020) with classifier-free",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
339,
541,
352
],
"spans": [
{
"bbox": [
69,
339,
541,
352
],
"score": 1.0,
"content": "guidance. At the first stage, we learn a single diffusion model that matches the combined output from",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
69,
350,
542,
364
],
"spans": [
{
"bbox": [
69,
350,
542,
364
],
"score": 1.0,
"content": "the jointly trained conditional and unconditional diffusion models, where the combination coefficients are",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
70,
362,
542,
376
],
"spans": [
{
"bbox": [
70,
362,
542,
376
],
"score": 1.0,
"content": "determined by the guidance weight. Then we apply progressive distillation to that single model to produce",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
374,
328,
388
],
"spans": [
{
"bbox": [
69,
374,
328,
388
],
"score": 1.0,
"content": "models requiring fewer sampling steps at the second stage.",
"type": "text"
}
],
"index": 22
}
],
"index": 20,
"bbox_fs": [
69,
327,
542,
388
]
},
{
"type": "text",
"bbox": [
71,
393,
540,
464
],
"lines": [
{
"bbox": [
70,
392,
541,
406
],
"spans": [
{
"bbox": [
70,
392,
235,
406
],
"score": 1.0,
"content": "After distillation, we use a stochastic",
"type": "text"
},
{
"bbox": [
236,
395,
245,
402
],
"score": 0.91,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
245,
392,
541,
406
],
"score": 1.0,
"content": "-step sampler: At each step, we first apply one deterministic DDIM",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
405,
542,
418
],
"spans": [
{
"bbox": [
69,
405,
363,
418
],
"score": 1.0,
"content": "update with twice the original step size (i.e., the same step size as a",
"type": "text"
},
{
"bbox": [
364,
406,
383,
417
],
"score": 0.91,
"content": "N / 2",
"type": "inline_equation"
},
{
"bbox": [
383,
405,
542,
418
],
"score": 1.0,
"content": "-step sampler), and then we perform",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
416,
542,
430
],
"spans": [
{
"bbox": [
69,
416,
542,
430
],
"score": 1.0,
"content": "one stochastic step backward (i.e., perturbed with noise following the forward diffusion process) with the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
428,
542,
442
],
"spans": [
{
"bbox": [
69,
428,
542,
442
],
"score": 1.0,
"content": "original step size, inspired by Karras et al. (2022). See Meng et al. (2022) for more details. Using this",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
440,
542,
454
],
"spans": [
{
"bbox": [
68,
440,
542,
454
],
"score": 1.0,
"content": "approach, we are able to distill all 7 video diffusion models down to just 8 sampling steps per model without",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
69,
452,
250,
466
],
"spans": [
{
"bbox": [
69,
452,
250,
466
],
"score": 1.0,
"content": "any noticeable loss in perceptual quality.",
"type": "text"
}
],
"index": 28
}
],
"index": 25.5,
"bbox_fs": [
68,
392,
542,
466
]
},
{
"type": "title",
"bbox": [
71,
479,
159,
493
],
"lines": [
{
"bbox": [
68,
477,
161,
497
],
"spans": [
{
"bbox": [
68,
477,
161,
497
],
"score": 1.0,
"content": "3 Experiments",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
70,
505,
540,
625
],
"lines": [
{
"bbox": [
69,
505,
541,
519
],
"spans": [
{
"bbox": [
69,
505,
541,
519
],
"score": 1.0,
"content": "We train our models on a combination of an internal dataset consisting of 14 million video-text pairs and",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
516,
541,
531
],
"spans": [
{
"bbox": [
69,
516,
541,
531
],
"score": 1.0,
"content": "60 million image-text pairs, and the publicly available LAION-400M image-text dataset (Schuhmann et al.,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
68,
528,
542,
543
],
"spans": [
{
"bbox": [
68,
528,
542,
543
],
"score": 1.0,
"content": "2021). To process the data into a form suitable for training our cascading pipeline, we spatially resize",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
541,
541,
555
],
"spans": [
{
"bbox": [
69,
541,
541,
555
],
"score": 1.0,
"content": "images and videos using antialiased bilinear resizing, and we temporally resize videos by skipping frames.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
69,
553,
541,
567
],
"spans": [
{
"bbox": [
69,
553,
541,
567
],
"score": 1.0,
"content": "Throughout our model development process, we evaluated Imagen Video on several different metrics, such",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
68,
564,
542,
580
],
"spans": [
{
"bbox": [
68,
564,
542,
580
],
"score": 1.0,
"content": "as FID on individual frames (Heusel et al., 2017), FVD (Unterthiner et al., 2019) for temporal consistency,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
577,
542,
591
],
"spans": [
{
"bbox": [
69,
577,
542,
591
],
"score": 1.0,
"content": "and frame-wise CLIP scores (Hessel et al., 2021; Park et al., 2021) for video-text alignment. Below, we",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
70,
589,
542,
603
],
"spans": [
{
"bbox": [
70,
589,
542,
603
],
"score": 1.0,
"content": "explore the capabilities of our model and investigate its performance in regards to 1) scaling up the number",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
69,
601,
542,
615
],
"spans": [
{
"bbox": [
69,
601,
542,
615
],
"score": 1.0,
"content": "of parameters in our model, 2) changing the parameterization of our model, and 3) distilling our models so",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
612,
221,
627
],
"spans": [
{
"bbox": [
69,
612,
221,
627
],
"score": 1.0,
"content": "that they are fast to sample from.",
"type": "text"
}
],
"index": 39
}
],
"index": 34.5,
"bbox_fs": [
68,
505,
542,
627
]
},
{
"type": "title",
"bbox": [
72,
639,
262,
651
],
"lines": [
{
"bbox": [
69,
638,
263,
653
],
"spans": [
{
"bbox": [
69,
638,
263,
653
],
"score": 1.0,
"content": "3.1 Unique video generation capabilities",
"type": "text"
}
],
"index": 40
}
],
"index": 40
},
{
"type": "text",
"bbox": [
71,
660,
540,
732
],
"lines": [
{
"bbox": [
70,
659,
541,
675
],
"spans": [
{
"bbox": [
70,
659,
541,
675
],
"score": 1.0,
"content": "We find that Imagen Video is capable of generating high fidelity video, and that it possesses several unique",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
69,
672,
541,
686
],
"spans": [
{
"bbox": [
69,
672,
541,
686
],
"score": 1.0,
"content": "capabilities that are not traditionally found in unstructured generative models learned purely from data. For",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
68,
682,
542,
700
],
"spans": [
{
"bbox": [
68,
682,
542,
700
],
"score": 1.0,
"content": "example, Fig. 8 shows that our model is capable of generating videos with artistic styles learned from image",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
69,
695,
541,
710
],
"spans": [
{
"bbox": [
69,
695,
541,
710
],
"score": 1.0,
"content": "information, such as videos in the style of van Gogh paintings or watercolor paintings. Fig. 9 shows that",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
70,
709,
541,
721
],
"spans": [
{
"bbox": [
70,
709,
541,
721
],
"score": 1.0,
"content": "Imagen Video possesses an understanding of 3D structure, as it is capable of generating videos of objects",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
69,
720,
541,
734
],
"spans": [
{
"bbox": [
69,
720,
541,
734
],
"score": 1.0,
"content": "rotating while roughly preserving structure. While the 3D consistency over the course of rotation is not",
"type": "text"
}
],
"index": 46
}
],
"index": 43.5,
"bbox_fs": [
68,
659,
542,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
84,
52,
528,
241
],
"blocks": [
{
"type": "image_body",
"bbox": [
84,
52,
528,
241
],
"group_id": 1,
"lines": [
{
"bbox": [
84,
52,
528,
241
],
"spans": [
{
"bbox": [
84,
52,
528,
241
],
"score": 0.895,
"type": "image",
"image_path": "b004e146b69df389b282868202d21643cdcce59a70e00811de230efb4e553a71.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
84,
52,
528,
115.0
],
"spans": [],
"index": 0
},
{
"bbox": [
84,
115.0,
528,
178.0
],
"spans": [],
"index": 1
},
{
"bbox": [
84,
178.0,
528,
241.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
214,
241,
396,
251
],
"group_id": 1,
"lines": [
{
"bbox": [
213,
240,
394,
252
],
"spans": [
{
"bbox": [
213,
240,
394,
252
],
"score": 1.0,
"content": "Drone flythrough of a pixel art of futuristic city.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "text",
"bbox": [
71,
252,
535,
275
],
"lines": [
{
"bbox": [
69,
252,
537,
264
],
"spans": [
{
"bbox": [
69,
252,
537,
264
],
"score": 1.0,
"content": "Figure 8: Snapshots of frames from videos generated by Imagen Video demonstrating the ability of the mode",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
264,
280,
277
],
"spans": [
{
"bbox": [
69,
264,
280,
277
],
"score": 1.0,
"content": "to generate dynamics in different artistic styles.",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5
},
{
"type": "image",
"bbox": [
84,
279,
524,
462
],
"blocks": [
{
"type": "image_body",
"bbox": [
84,
279,
524,
462
],
"group_id": 0,
"lines": [
{
"bbox": [
84,
279,
524,
462
],
"spans": [
{
"bbox": [
84,
279,
524,
462
],
"score": 0.893,
"type": "image",
"image_path": "1221ffbdcd6a012e86f0f7ae55bf9572a1a4cc73e17dbc2816c7853da737fc93.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
84,
279,
524,
340.0
],
"spans": [],
"index": 6
},
{
"bbox": [
84,
340.0,
524,
401.0
],
"spans": [],
"index": 7
},
{
"bbox": [
84,
401.0,
524,
462.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
203,
467,
403,
477
],
"group_id": 0,
"lines": [
{
"bbox": [
205,
465,
402,
478
],
"spans": [
{
"bbox": [
205,
465,
402,
478
],
"score": 1.0,
"content": "A 3D model of an elephant origami. Studio lighting.",
"type": "text"
}
],
"index": 9
}
],
"index": 9
}
],
"index": 8.0
},
{
"type": "image",
"bbox": [
85,
504,
523,
556
],
"blocks": [
{
"type": "image_caption",
"bbox": [
73,
477,
533,
500
],
"group_id": 2,
"lines": [
{
"bbox": [
70,
477,
535,
489
],
"spans": [
{
"bbox": [
70,
477,
535,
489
],
"score": 1.0,
"content": "Figure 9: Snapshots of frames from videos generated by Imagen Video demonstrating the model’s under",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
71,
489,
186,
501
],
"spans": [
{
"bbox": [
71,
489,
186,
501
],
"score": 1.0,
"content": "standing of 3D structures.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
},
{
"type": "image_body",
"bbox": [
85,
504,
523,
556
],
"group_id": 2,
"lines": [
{
"bbox": [
85,
504,
523,
556
],
"spans": [
{
"bbox": [
85,
504,
523,
556
],
"score": 0.841,
"type": "image",
"image_path": "a9874e80b01e0778d489f71b11bd54576424d9316e72970675c0188dd6cafc84.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
85,
504,
523,
521.3333333333334
],
"spans": [],
"index": 12
},
{
"bbox": [
85,
521.3333333333334,
523,
538.6666666666667
],
"spans": [],
"index": 13
},
{
"bbox": [
85,
538.6666666666667,
523,
556.0000000000001
],
"spans": [],
"index": 14
}
]
}
],
"index": 11.75
},
{
"type": "image",
"bbox": [
85,
570,
523,
622
],
"blocks": [
{
"type": "image_caption",
"bbox": [
96,
560,
508,
569
],
"group_id": 3,
"lines": [
{
"bbox": [
101,
558,
506,
570
],
"spans": [
{
"bbox": [
101,
558,
506,
570
],
"score": 1.0,
"content": "A colorful professional animated logo for ’Diffusion’ written using paint brush in cursive. Smooth animation.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "image_body",
"bbox": [
85,
570,
523,
622
],
"group_id": 3,
"lines": [
{
"bbox": [
85,
570,
523,
622
],
"spans": [
{
"bbox": [
85,
570,
523,
622
],
"score": 0.589,
"type": "image",
"image_path": "97dea5c09c0516c0d2a02182394557275c599af5ef20c8ebd86e9614f41c37ad.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
85,
570,
523,
587.3333333333334
],
"spans": [],
"index": 16
},
{
"bbox": [
85,
587.3333333333334,
523,
604.6666666666667
],
"spans": [],
"index": 17
},
{
"bbox": [
85,
604.6666666666667,
523,
622.0000000000001
],
"spans": [],
"index": 18
}
]
}
],
"index": 16.0
},
{
"type": "text",
"bbox": [
175,
626,
430,
635
],
"lines": [
{
"bbox": [
174,
624,
432,
637
],
"spans": [
{
"bbox": [
174,
624,
432,
637
],
"score": 1.0,
"content": "Sprouts in the shape of text ’Imagen’ coming out of a fairytale book.",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "image",
"bbox": [
85,
637,
524,
687
],
"blocks": [
{
"type": "image_body",
"bbox": [
85,
637,
524,
687
],
"group_id": 4,
"lines": [
{
"bbox": [
85,
637,
524,
687
],
"spans": [
{
"bbox": [
85,
637,
524,
687
],
"score": 0.423,
"type": "image",
"image_path": "342234d4e502eb9100353aedd56b6b2178aaac41c69626055fd6e62b0b83d2bc.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
85,
637,
524,
653.6666666666666
],
"spans": [],
"index": 20
},
{
"bbox": [
85,
653.6666666666666,
524,
670.3333333333333
],
"spans": [],
"index": 21
},
{
"bbox": [
85,
670.3333333333333,
524,
686.9999999999999
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
72,
691,
538,
726
],
"group_id": 4,
"lines": [
{
"bbox": [
87,
690,
520,
703
],
"spans": [
{
"bbox": [
87,
690,
520,
703
],
"score": 1.0,
"content": "Thousands of fast brush strokes slowly forming the text ’Imagen Video’ on a light beige canvas. Smooth animation.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
702,
540,
715
],
"spans": [
{
"bbox": [
69,
702,
540,
715
],
"score": 1.0,
"content": "Figure 10: Snapshots of frames from videos generated by Imagen Video demonstrating the ability of the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
713,
367,
727
],
"spans": [
{
"bbox": [
69,
713,
367,
727
],
"score": 1.0,
"content": "model to render a variety of text with different style and dynamics.",
"type": "text"
}
],
"index": 25
}
],
"index": 24
}
],
"index": 22.5
}
],
"page_idx": 11,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "image",
"bbox": [
84,
52,
528,
241
],
"blocks": [
{
"type": "image_body",
"bbox": [
84,
52,
528,
241
],
"group_id": 1,
"lines": [
{
"bbox": [
84,
52,
528,
241
],
"spans": [
{
"bbox": [
84,
52,
528,
241
],
"score": 0.895,
"type": "image",
"image_path": "b004e146b69df389b282868202d21643cdcce59a70e00811de230efb4e553a71.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
84,
52,
528,
115.0
],
"spans": [],
"index": 0
},
{
"bbox": [
84,
115.0,
528,
178.0
],
"spans": [],
"index": 1
},
{
"bbox": [
84,
178.0,
528,
241.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
214,
241,
396,
251
],
"group_id": 1,
"lines": [
{
"bbox": [
213,
240,
394,
252
],
"spans": [
{
"bbox": [
213,
240,
394,
252
],
"score": 1.0,
"content": "Drone flythrough of a pixel art of futuristic city.",
"type": "text"
}
],
"index": 3
}
],
"index": 3
}
],
"index": 2.0
},
{
"type": "text",
"bbox": [
71,
252,
535,
275
],
"lines": [
{
"bbox": [
69,
252,
537,
264
],
"spans": [
{
"bbox": [
69,
252,
537,
264
],
"score": 1.0,
"content": "Figure 8: Snapshots of frames from videos generated by Imagen Video demonstrating the ability of the mode",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
264,
280,
277
],
"spans": [
{
"bbox": [
69,
264,
280,
277
],
"score": 1.0,
"content": "to generate dynamics in different artistic styles.",
"type": "text"
}
],
"index": 5
}
],
"index": 4.5,
"bbox_fs": [
69,
252,
537,
277
]
},
{
"type": "image",
"bbox": [
84,
279,
524,
462
],
"blocks": [
{
"type": "image_body",
"bbox": [
84,
279,
524,
462
],
"group_id": 0,
"lines": [
{
"bbox": [
84,
279,
524,
462
],
"spans": [
{
"bbox": [
84,
279,
524,
462
],
"score": 0.893,
"type": "image",
"image_path": "1221ffbdcd6a012e86f0f7ae55bf9572a1a4cc73e17dbc2816c7853da737fc93.jpg"
}
]
}
],
"index": 7,
"virtual_lines": [
{
"bbox": [
84,
279,
524,
340.0
],
"spans": [],
"index": 6
},
{
"bbox": [
84,
340.0,
524,
401.0
],
"spans": [],
"index": 7
},
{
"bbox": [
84,
401.0,
524,
462.0
],
"spans": [],
"index": 8
}
]
},
{
"type": "image_caption",
"bbox": [
203,
467,
403,
477
],
"group_id": 0,
"lines": [
{
"bbox": [
205,
465,
402,
478
],
"spans": [
{
"bbox": [
205,
465,
402,
478
],
"score": 1.0,
"content": "A 3D model of an elephant origami. Studio lighting.",
"type": "text"
}
],
"index": 9
}
],
"index": 9
}
],
"index": 8.0
},
{
"type": "image",
"bbox": [
85,
504,
523,
556
],
"blocks": [
{
"type": "image_caption",
"bbox": [
73,
477,
533,
500
],
"group_id": 2,
"lines": [
{
"bbox": [
70,
477,
535,
489
],
"spans": [
{
"bbox": [
70,
477,
535,
489
],
"score": 1.0,
"content": "Figure 9: Snapshots of frames from videos generated by Imagen Video demonstrating the model’s under",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
71,
489,
186,
501
],
"spans": [
{
"bbox": [
71,
489,
186,
501
],
"score": 1.0,
"content": "standing of 3D structures.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
},
{
"type": "image_body",
"bbox": [
85,
504,
523,
556
],
"group_id": 2,
"lines": [
{
"bbox": [
85,
504,
523,
556
],
"spans": [
{
"bbox": [
85,
504,
523,
556
],
"score": 0.841,
"type": "image",
"image_path": "a9874e80b01e0778d489f71b11bd54576424d9316e72970675c0188dd6cafc84.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
85,
504,
523,
521.3333333333334
],
"spans": [],
"index": 12
},
{
"bbox": [
85,
521.3333333333334,
523,
538.6666666666667
],
"spans": [],
"index": 13
},
{
"bbox": [
85,
538.6666666666667,
523,
556.0000000000001
],
"spans": [],
"index": 14
}
]
}
],
"index": 11.75
},
{
"type": "image",
"bbox": [
85,
570,
523,
622
],
"blocks": [
{
"type": "image_caption",
"bbox": [
96,
560,
508,
569
],
"group_id": 3,
"lines": [
{
"bbox": [
101,
558,
506,
570
],
"spans": [
{
"bbox": [
101,
558,
506,
570
],
"score": 1.0,
"content": "A colorful professional animated logo for ’Diffusion’ written using paint brush in cursive. Smooth animation.",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "image_body",
"bbox": [
85,
570,
523,
622
],
"group_id": 3,
"lines": [
{
"bbox": [
85,
570,
523,
622
],
"spans": [
{
"bbox": [
85,
570,
523,
622
],
"score": 0.589,
"type": "image",
"image_path": "97dea5c09c0516c0d2a02182394557275c599af5ef20c8ebd86e9614f41c37ad.jpg"
}
]
}
],
"index": 17,
"virtual_lines": [
{
"bbox": [
85,
570,
523,
587.3333333333334
],
"spans": [],
"index": 16
},
{
"bbox": [
85,
587.3333333333334,
523,
604.6666666666667
],
"spans": [],
"index": 17
},
{
"bbox": [
85,
604.6666666666667,
523,
622.0000000000001
],
"spans": [],
"index": 18
}
]
}
],
"index": 16.0
},
{
"type": "text",
"bbox": [
175,
626,
430,
635
],
"lines": [
{
"bbox": [
174,
624,
432,
637
],
"spans": [
{
"bbox": [
174,
624,
432,
637
],
"score": 1.0,
"content": "Sprouts in the shape of text ’Imagen’ coming out of a fairytale book.",
"type": "text"
}
],
"index": 19
}
],
"index": 19,
"bbox_fs": [
174,
624,
432,
637
]
},
{
"type": "image",
"bbox": [
85,
637,
524,
687
],
"blocks": [
{
"type": "image_body",
"bbox": [
85,
637,
524,
687
],
"group_id": 4,
"lines": [
{
"bbox": [
85,
637,
524,
687
],
"spans": [
{
"bbox": [
85,
637,
524,
687
],
"score": 0.423,
"type": "image",
"image_path": "342234d4e502eb9100353aedd56b6b2178aaac41c69626055fd6e62b0b83d2bc.jpg"
}
]
}
],
"index": 21,
"virtual_lines": [
{
"bbox": [
85,
637,
524,
653.6666666666666
],
"spans": [],
"index": 20
},
{
"bbox": [
85,
653.6666666666666,
524,
670.3333333333333
],
"spans": [],
"index": 21
},
{
"bbox": [
85,
670.3333333333333,
524,
686.9999999999999
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
72,
691,
538,
726
],
"group_id": 4,
"lines": [
{
"bbox": [
87,
690,
520,
703
],
"spans": [
{
"bbox": [
87,
690,
520,
703
],
"score": 1.0,
"content": "Thousands of fast brush strokes slowly forming the text ’Imagen Video’ on a light beige canvas. Smooth animation.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
702,
540,
715
],
"spans": [
{
"bbox": [
69,
702,
540,
715
],
"score": 1.0,
"content": "Figure 10: Snapshots of frames from videos generated by Imagen Video demonstrating the ability of the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
713,
367,
727
],
"spans": [
{
"bbox": [
69,
713,
367,
727
],
"score": 1.0,
"content": "model to render a variety of text with different style and dynamics.",
"type": "text"
}
],
"index": 25
}
],
"index": 24
}
],
"index": 22.5
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
142
],
"lines": [
{
"bbox": [
70,
83,
540,
94
],
"spans": [
{
"bbox": [
70,
83,
540,
94
],
"score": 1.0,
"content": "exact, we believe Imagen Video shows that video models can serve as effective priors for methods that do",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
69,
94,
541,
107
],
"spans": [
{
"bbox": [
69,
94,
541,
107
],
"score": 1.0,
"content": "force 3D consistency. Fig. 10 shows that Imagen Video is also reliably capable of generating text in a wide",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
70,
106,
541,
119
],
"spans": [
{
"bbox": [
70,
106,
541,
119
],
"score": 1.0,
"content": "variety of animation styles, some of which would be difficult to animate using traditional tools. We see",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
68,
117,
541,
132
],
"spans": [
{
"bbox": [
68,
117,
541,
132
],
"score": 1.0,
"content": "results such as these as an exciting indication of how general purpose generative models such as Imagen",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
129,
423,
144
],
"spans": [
{
"bbox": [
70,
129,
423,
144
],
"score": 1.0,
"content": "Video can significantly decrease the difficulty of high quality content generation.",
"type": "text"
}
],
"index": 4
}
],
"index": 2
},
{
"type": "title",
"bbox": [
71,
154,
129,
167
],
"lines": [
{
"bbox": [
68,
151,
132,
171
],
"spans": [
{
"bbox": [
68,
151,
132,
171
],
"score": 1.0,
"content": "3.2 Scaling",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "text",
"bbox": [
71,
176,
540,
248
],
"lines": [
{
"bbox": [
69,
175,
541,
190
],
"spans": [
{
"bbox": [
69,
175,
541,
190
],
"score": 1.0,
"content": "In Figure 11 we show that our base video model strongly benefits from scaling up the parameter count of the",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
187,
541,
201
],
"spans": [
{
"bbox": [
70,
187,
541,
201
],
"score": 1.0,
"content": "video U-Net. We performed this scaling by increasing the base channel count and depth of the network. This",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
70,
200,
540,
212
],
"spans": [
{
"bbox": [
70,
200,
540,
212
],
"score": 1.0,
"content": "result is contrary to the text-to-image U-Net scaling results by Saharia et al. (2022b), which found limited",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
212,
541,
226
],
"spans": [
{
"bbox": [
69,
212,
541,
226
],
"score": 1.0,
"content": "benefit from diffusion model scaling when measured by image-text sample quality scores. We conclude that",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
223,
541,
238
],
"spans": [
{
"bbox": [
69,
223,
541,
238
],
"score": 1.0,
"content": "video modeling is a harder task for which performance is not yet saturated at current model sizes, implying",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
235,
337,
250
],
"spans": [
{
"bbox": [
70,
235,
337,
250
],
"score": 1.0,
"content": "future benefits to further model scaling for video generation.",
"type": "text"
}
],
"index": 11
}
],
"index": 8.5
},
{
"type": "image",
"bbox": [
73,
257,
523,
368
],
"blocks": [
{
"type": "image_body",
"bbox": [
73,
257,
523,
368
],
"group_id": 0,
"lines": [
{
"bbox": [
73,
257,
523,
368
],
"spans": [
{
"bbox": [
73,
257,
523,
368
],
"score": 0.838,
"type": "image",
"image_path": "789e8cc4f7559686e6db1efe096b011df90ad9bd57a151db67d7f5d533171a51.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
73,
257,
523,
294.0
],
"spans": [],
"index": 12
},
{
"bbox": [
73,
294.0,
523,
331.0
],
"spans": [],
"index": 13
},
{
"bbox": [
73,
331.0,
523,
368.0
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
70,
377,
541,
413
],
"group_id": 0,
"lines": [
{
"bbox": [
69,
376,
541,
390
],
"spans": [
{
"bbox": [
69,
376,
270,
390
],
"score": 1.0,
"content": "Figure 11: Scaling Comparison for the base",
"type": "text"
},
{
"bbox": [
271,
378,
318,
388
],
"score": 0.77,
"content": "1 6 \\times 4 0 \\times 2 4",
"type": "inline_equation"
},
{
"bbox": [
318,
376,
541,
390
],
"score": 1.0,
"content": "video model on FVD and CLIP scores (on 0-100",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
388,
541,
402
],
"spans": [
{
"bbox": [
69,
388,
541,
402
],
"score": 1.0,
"content": "scale). Both FVD and CLIP scores are computed on 4096 video samples. We see clear signs of improvement",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
399,
376,
414
],
"spans": [
{
"bbox": [
69,
399,
376,
414
],
"score": 1.0,
"content": "on both metrics when scaling from 500M to 1.6B to 5.6B parameters.",
"type": "text"
}
],
"index": 17
}
],
"index": 16
}
],
"index": 14.5
},
{
"type": "title",
"bbox": [
71,
431,
279,
443
],
"lines": [
{
"bbox": [
69,
430,
280,
446
],
"spans": [
{
"bbox": [
69,
430,
280,
446
],
"score": 1.0,
"content": "3.3 Comparing prediction parameterizations",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
70,
452,
540,
549
],
"lines": [
{
"bbox": [
69,
452,
542,
467
],
"spans": [
{
"bbox": [
69,
452,
272,
467
],
"score": 1.0,
"content": "In early experiments we found that training",
"type": "text"
},
{
"bbox": [
272,
458,
278,
463
],
"score": 0.52,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
278,
452,
542,
467
],
"score": 1.0,
"content": "-prediction models (Ho et al., 2020) performed worse than",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
71,
464,
541,
478
],
"spans": [
{
"bbox": [
71,
470,
78,
475
],
"score": 0.53,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
78,
464,
541,
478
],
"score": 1.0,
"content": "-prediction (Salimans & Ho, 2022) especially at high resolutions. Specifically, for high resolution SSR",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
68,
476,
541,
490
],
"spans": [
{
"bbox": [
68,
476,
188,
490
],
"score": 1.0,
"content": "models, we observed that",
"type": "text"
},
{
"bbox": [
189,
482,
194,
487
],
"score": 0.48,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
194,
476,
541,
490
],
"score": 1.0,
"content": "-prediction converges relatively slowly in terms of sample quality metrics and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
68,
488,
541,
502
],
"spans": [
{
"bbox": [
68,
488,
541,
502
],
"score": 1.0,
"content": "suffers from color shift and color inconsistency across frames in the generated videos. Fig. 12 shows the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
501,
541,
514
],
"spans": [
{
"bbox": [
69,
501,
164,
514
],
"score": 1.0,
"content": "comparison between",
"type": "text"
},
{
"bbox": [
164,
506,
169,
511
],
"score": 0.84,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
169,
501,
240,
514
],
"score": 1.0,
"content": "-prediction and",
"type": "text"
},
{
"bbox": [
240,
506,
246,
511
],
"score": 0.65,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
247,
501,
331,
514
],
"score": 1.0,
"content": "-prediction on a 80",
"type": "text"
},
{
"bbox": [
332,
505,
339,
511
],
"score": 0.61,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
340,
501,
382,
514
],
"score": 1.0,
"content": "48 → 320",
"type": "text"
},
{
"bbox": [
383,
505,
390,
511
],
"score": 0.67,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
390,
501,
541,
514
],
"score": 1.0,
"content": "192 video spatial super-resolution",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
512,
541,
526
],
"spans": [
{
"bbox": [
69,
512,
161,
526
],
"score": 1.0,
"content": "task. It is clear that",
"type": "text"
},
{
"bbox": [
162,
518,
167,
523
],
"score": 0.85,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
167,
512,
390,
526
],
"score": 1.0,
"content": "-parameterization produces worse generations than",
"type": "text"
},
{
"bbox": [
391,
518,
397,
523
],
"score": 0.65,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
397,
512,
541,
526
],
"score": 1.0,
"content": "-parameterization. Fig. 13 shows",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
524,
541,
539
],
"spans": [
{
"bbox": [
69,
524,
541,
539
],
"score": 1.0,
"content": "the quantitative comparison between the two parameterizations as a function of training steps. We observe",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
536,
412,
550
],
"spans": [
{
"bbox": [
69,
536,
92,
550
],
"score": 1.0,
"content": "that",
"type": "text"
},
{
"bbox": [
93,
542,
99,
546
],
"score": 0.83,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
100,
536,
326,
550
],
"score": 1.0,
"content": "parameterization converges much more faster than",
"type": "text"
},
{
"bbox": [
326,
542,
331,
546
],
"score": 0.85,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
331,
536,
412,
550
],
"score": 1.0,
"content": "parameterization.",
"type": "text"
}
],
"index": 26
}
],
"index": 22.5
},
{
"type": "title",
"bbox": [
71,
561,
250,
573
],
"lines": [
{
"bbox": [
69,
559,
251,
576
],
"spans": [
{
"bbox": [
69,
559,
251,
576
],
"score": 1.0,
"content": "3.4 Perceptual quality and distillation",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
70,
582,
540,
654
],
"lines": [
{
"bbox": [
69,
581,
541,
596
],
"spans": [
{
"bbox": [
69,
581,
541,
596
],
"score": 1.0,
"content": "In Table 1 we report perceptual quality metrics (CLIP score and CLIP R-Precision) for our model samples,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
595,
542,
609
],
"spans": [
{
"bbox": [
69,
595,
434,
609
],
"score": 1.0,
"content": "as well as for their distilled version. Samples are generated and evaluated at 192",
"type": "text"
},
{
"bbox": [
435,
598,
443,
605
],
"score": 0.26,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
443,
595,
542,
609
],
"score": 1.0,
"content": "320 resolution for 128",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
606,
541,
621
],
"spans": [
{
"bbox": [
69,
606,
541,
621
],
"score": 1.0,
"content": "frames at 24 frames per second. For CLIP score, we take the average score over all frames. For CLIP",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
618,
542,
632
],
"spans": [
{
"bbox": [
69,
618,
371,
632
],
"score": 1.0,
"content": "R-Precision (Park et al., 2021) we compute the top-1 accuracy (i.e.",
"type": "text"
},
{
"bbox": [
371,
621,
399,
628
],
"score": 0.9,
"content": "R = 1",
"type": "inline_equation"
},
{
"bbox": [
399,
618,
542,
632
],
"score": 1.0,
"content": "), treating the frames of a video",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
630,
541,
644
],
"spans": [
{
"bbox": [
69,
630,
541,
644
],
"score": 1.0,
"content": "sample as images sharing the same text label (the prompt). We repeat these over four different runs and",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
68,
642,
233,
657
],
"spans": [
{
"bbox": [
68,
642,
233,
657
],
"score": 1.0,
"content": "report the mean and standard error.",
"type": "text"
}
],
"index": 33
}
],
"index": 30.5
},
{
"type": "text",
"bbox": [
70,
660,
540,
732
],
"lines": [
{
"bbox": [
69,
658,
541,
675
],
"spans": [
{
"bbox": [
69,
658,
541,
675
],
"score": 1.0,
"content": "We find that distillation provides a very favorable trade-off between sampling time and perceptual quality:",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
672,
542,
686
],
"spans": [
{
"bbox": [
70,
672,
212,
686
],
"score": 1.0,
"content": "the distilled cascade is about 18",
"type": "text"
},
{
"bbox": [
212,
676,
219,
683
],
"score": 0.82,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
219,
672,
542,
686
],
"score": 1.0,
"content": "faster, while producing videos of similar quality to the samples from the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
685,
542,
697
],
"spans": [
{
"bbox": [
69,
685,
374,
697
],
"score": 1.0,
"content": "original models. In terms of FLOPs, the distilled models are about 36",
"type": "text"
},
{
"bbox": [
375,
688,
382,
695
],
"score": 0.68,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
382,
685,
542,
697
],
"score": 1.0,
"content": "more efficient: The original cascade",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
69,
696,
541,
710
],
"spans": [
{
"bbox": [
69,
696,
541,
710
],
"score": 1.0,
"content": "evaluates each model twice (in parallel) to apply classifier-free guidance, while our distilled models do not,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
68,
707,
541,
722
],
"spans": [
{
"bbox": [
68,
707,
541,
722
],
"score": 1.0,
"content": "since they distilled the effect of guidance into a single model. We provide samples from our original and",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
720,
268,
733
],
"spans": [
{
"bbox": [
69,
720,
268,
733
],
"score": 1.0,
"content": "distilled cascade in Figure 14 for illustration.",
"type": "text"
}
],
"index": 39
}
],
"index": 36.5
}
],
"page_idx": 12,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
70,
26,
239,
38
],
"spans": [
{
"bbox": [
70,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "13",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
142
],
"lines": [
{
"bbox": [
70,
83,
540,
94
],
"spans": [
{
"bbox": [
70,
83,
540,
94
],
"score": 1.0,
"content": "exact, we believe Imagen Video shows that video models can serve as effective priors for methods that do",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
69,
94,
541,
107
],
"spans": [
{
"bbox": [
69,
94,
541,
107
],
"score": 1.0,
"content": "force 3D consistency. Fig. 10 shows that Imagen Video is also reliably capable of generating text in a wide",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
70,
106,
541,
119
],
"spans": [
{
"bbox": [
70,
106,
541,
119
],
"score": 1.0,
"content": "variety of animation styles, some of which would be difficult to animate using traditional tools. We see",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
68,
117,
541,
132
],
"spans": [
{
"bbox": [
68,
117,
541,
132
],
"score": 1.0,
"content": "results such as these as an exciting indication of how general purpose generative models such as Imagen",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
129,
423,
144
],
"spans": [
{
"bbox": [
70,
129,
423,
144
],
"score": 1.0,
"content": "Video can significantly decrease the difficulty of high quality content generation.",
"type": "text"
}
],
"index": 4
}
],
"index": 2,
"bbox_fs": [
68,
83,
541,
144
]
},
{
"type": "title",
"bbox": [
71,
154,
129,
167
],
"lines": [
{
"bbox": [
68,
151,
132,
171
],
"spans": [
{
"bbox": [
68,
151,
132,
171
],
"score": 1.0,
"content": "3.2 Scaling",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "text",
"bbox": [
71,
176,
540,
248
],
"lines": [
{
"bbox": [
69,
175,
541,
190
],
"spans": [
{
"bbox": [
69,
175,
541,
190
],
"score": 1.0,
"content": "In Figure 11 we show that our base video model strongly benefits from scaling up the parameter count of the",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
187,
541,
201
],
"spans": [
{
"bbox": [
70,
187,
541,
201
],
"score": 1.0,
"content": "video U-Net. We performed this scaling by increasing the base channel count and depth of the network. This",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
70,
200,
540,
212
],
"spans": [
{
"bbox": [
70,
200,
540,
212
],
"score": 1.0,
"content": "result is contrary to the text-to-image U-Net scaling results by Saharia et al. (2022b), which found limited",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
212,
541,
226
],
"spans": [
{
"bbox": [
69,
212,
541,
226
],
"score": 1.0,
"content": "benefit from diffusion model scaling when measured by image-text sample quality scores. We conclude that",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
223,
541,
238
],
"spans": [
{
"bbox": [
69,
223,
541,
238
],
"score": 1.0,
"content": "video modeling is a harder task for which performance is not yet saturated at current model sizes, implying",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
70,
235,
337,
250
],
"spans": [
{
"bbox": [
70,
235,
337,
250
],
"score": 1.0,
"content": "future benefits to further model scaling for video generation.",
"type": "text"
}
],
"index": 11
}
],
"index": 8.5,
"bbox_fs": [
69,
175,
541,
250
]
},
{
"type": "image",
"bbox": [
73,
257,
523,
368
],
"blocks": [
{
"type": "image_body",
"bbox": [
73,
257,
523,
368
],
"group_id": 0,
"lines": [
{
"bbox": [
73,
257,
523,
368
],
"spans": [
{
"bbox": [
73,
257,
523,
368
],
"score": 0.838,
"type": "image",
"image_path": "789e8cc4f7559686e6db1efe096b011df90ad9bd57a151db67d7f5d533171a51.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
73,
257,
523,
294.0
],
"spans": [],
"index": 12
},
{
"bbox": [
73,
294.0,
523,
331.0
],
"spans": [],
"index": 13
},
{
"bbox": [
73,
331.0,
523,
368.0
],
"spans": [],
"index": 14
}
]
},
{
"type": "image_caption",
"bbox": [
70,
377,
541,
413
],
"group_id": 0,
"lines": [
{
"bbox": [
69,
376,
541,
390
],
"spans": [
{
"bbox": [
69,
376,
270,
390
],
"score": 1.0,
"content": "Figure 11: Scaling Comparison for the base",
"type": "text"
},
{
"bbox": [
271,
378,
318,
388
],
"score": 0.77,
"content": "1 6 \\times 4 0 \\times 2 4",
"type": "inline_equation"
},
{
"bbox": [
318,
376,
541,
390
],
"score": 1.0,
"content": "video model on FVD and CLIP scores (on 0-100",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
388,
541,
402
],
"spans": [
{
"bbox": [
69,
388,
541,
402
],
"score": 1.0,
"content": "scale). Both FVD and CLIP scores are computed on 4096 video samples. We see clear signs of improvement",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
399,
376,
414
],
"spans": [
{
"bbox": [
69,
399,
376,
414
],
"score": 1.0,
"content": "on both metrics when scaling from 500M to 1.6B to 5.6B parameters.",
"type": "text"
}
],
"index": 17
}
],
"index": 16
}
],
"index": 14.5
},
{
"type": "title",
"bbox": [
71,
431,
279,
443
],
"lines": [
{
"bbox": [
69,
430,
280,
446
],
"spans": [
{
"bbox": [
69,
430,
280,
446
],
"score": 1.0,
"content": "3.3 Comparing prediction parameterizations",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
70,
452,
540,
549
],
"lines": [
{
"bbox": [
69,
452,
542,
467
],
"spans": [
{
"bbox": [
69,
452,
272,
467
],
"score": 1.0,
"content": "In early experiments we found that training",
"type": "text"
},
{
"bbox": [
272,
458,
278,
463
],
"score": 0.52,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
278,
452,
542,
467
],
"score": 1.0,
"content": "-prediction models (Ho et al., 2020) performed worse than",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
71,
464,
541,
478
],
"spans": [
{
"bbox": [
71,
470,
78,
475
],
"score": 0.53,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
78,
464,
541,
478
],
"score": 1.0,
"content": "-prediction (Salimans & Ho, 2022) especially at high resolutions. Specifically, for high resolution SSR",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
68,
476,
541,
490
],
"spans": [
{
"bbox": [
68,
476,
188,
490
],
"score": 1.0,
"content": "models, we observed that",
"type": "text"
},
{
"bbox": [
189,
482,
194,
487
],
"score": 0.48,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
194,
476,
541,
490
],
"score": 1.0,
"content": "-prediction converges relatively slowly in terms of sample quality metrics and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
68,
488,
541,
502
],
"spans": [
{
"bbox": [
68,
488,
541,
502
],
"score": 1.0,
"content": "suffers from color shift and color inconsistency across frames in the generated videos. Fig. 12 shows the",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
501,
541,
514
],
"spans": [
{
"bbox": [
69,
501,
164,
514
],
"score": 1.0,
"content": "comparison between",
"type": "text"
},
{
"bbox": [
164,
506,
169,
511
],
"score": 0.84,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
169,
501,
240,
514
],
"score": 1.0,
"content": "-prediction and",
"type": "text"
},
{
"bbox": [
240,
506,
246,
511
],
"score": 0.65,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
247,
501,
331,
514
],
"score": 1.0,
"content": "-prediction on a 80",
"type": "text"
},
{
"bbox": [
332,
505,
339,
511
],
"score": 0.61,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
340,
501,
382,
514
],
"score": 1.0,
"content": "48 → 320",
"type": "text"
},
{
"bbox": [
383,
505,
390,
511
],
"score": 0.67,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
390,
501,
541,
514
],
"score": 1.0,
"content": "192 video spatial super-resolution",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
512,
541,
526
],
"spans": [
{
"bbox": [
69,
512,
161,
526
],
"score": 1.0,
"content": "task. It is clear that",
"type": "text"
},
{
"bbox": [
162,
518,
167,
523
],
"score": 0.85,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
167,
512,
390,
526
],
"score": 1.0,
"content": "-parameterization produces worse generations than",
"type": "text"
},
{
"bbox": [
391,
518,
397,
523
],
"score": 0.65,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
397,
512,
541,
526
],
"score": 1.0,
"content": "-parameterization. Fig. 13 shows",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
524,
541,
539
],
"spans": [
{
"bbox": [
69,
524,
541,
539
],
"score": 1.0,
"content": "the quantitative comparison between the two parameterizations as a function of training steps. We observe",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
536,
412,
550
],
"spans": [
{
"bbox": [
69,
536,
92,
550
],
"score": 1.0,
"content": "that",
"type": "text"
},
{
"bbox": [
93,
542,
99,
546
],
"score": 0.83,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
100,
536,
326,
550
],
"score": 1.0,
"content": "parameterization converges much more faster than",
"type": "text"
},
{
"bbox": [
326,
542,
331,
546
],
"score": 0.85,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
331,
536,
412,
550
],
"score": 1.0,
"content": "parameterization.",
"type": "text"
}
],
"index": 26
}
],
"index": 22.5,
"bbox_fs": [
68,
452,
542,
550
]
},
{
"type": "title",
"bbox": [
71,
561,
250,
573
],
"lines": [
{
"bbox": [
69,
559,
251,
576
],
"spans": [
{
"bbox": [
69,
559,
251,
576
],
"score": 1.0,
"content": "3.4 Perceptual quality and distillation",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
70,
582,
540,
654
],
"lines": [
{
"bbox": [
69,
581,
541,
596
],
"spans": [
{
"bbox": [
69,
581,
541,
596
],
"score": 1.0,
"content": "In Table 1 we report perceptual quality metrics (CLIP score and CLIP R-Precision) for our model samples,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
595,
542,
609
],
"spans": [
{
"bbox": [
69,
595,
434,
609
],
"score": 1.0,
"content": "as well as for their distilled version. Samples are generated and evaluated at 192",
"type": "text"
},
{
"bbox": [
435,
598,
443,
605
],
"score": 0.26,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
443,
595,
542,
609
],
"score": 1.0,
"content": "320 resolution for 128",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
606,
541,
621
],
"spans": [
{
"bbox": [
69,
606,
541,
621
],
"score": 1.0,
"content": "frames at 24 frames per second. For CLIP score, we take the average score over all frames. For CLIP",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
618,
542,
632
],
"spans": [
{
"bbox": [
69,
618,
371,
632
],
"score": 1.0,
"content": "R-Precision (Park et al., 2021) we compute the top-1 accuracy (i.e.",
"type": "text"
},
{
"bbox": [
371,
621,
399,
628
],
"score": 0.9,
"content": "R = 1",
"type": "inline_equation"
},
{
"bbox": [
399,
618,
542,
632
],
"score": 1.0,
"content": "), treating the frames of a video",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
630,
541,
644
],
"spans": [
{
"bbox": [
69,
630,
541,
644
],
"score": 1.0,
"content": "sample as images sharing the same text label (the prompt). We repeat these over four different runs and",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
68,
642,
233,
657
],
"spans": [
{
"bbox": [
68,
642,
233,
657
],
"score": 1.0,
"content": "report the mean and standard error.",
"type": "text"
}
],
"index": 33
}
],
"index": 30.5,
"bbox_fs": [
68,
581,
542,
657
]
},
{
"type": "text",
"bbox": [
70,
660,
540,
732
],
"lines": [
{
"bbox": [
69,
658,
541,
675
],
"spans": [
{
"bbox": [
69,
658,
541,
675
],
"score": 1.0,
"content": "We find that distillation provides a very favorable trade-off between sampling time and perceptual quality:",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
70,
672,
542,
686
],
"spans": [
{
"bbox": [
70,
672,
212,
686
],
"score": 1.0,
"content": "the distilled cascade is about 18",
"type": "text"
},
{
"bbox": [
212,
676,
219,
683
],
"score": 0.82,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
219,
672,
542,
686
],
"score": 1.0,
"content": "faster, while producing videos of similar quality to the samples from the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
685,
542,
697
],
"spans": [
{
"bbox": [
69,
685,
374,
697
],
"score": 1.0,
"content": "original models. In terms of FLOPs, the distilled models are about 36",
"type": "text"
},
{
"bbox": [
375,
688,
382,
695
],
"score": 0.68,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
382,
685,
542,
697
],
"score": 1.0,
"content": "more efficient: The original cascade",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
69,
696,
541,
710
],
"spans": [
{
"bbox": [
69,
696,
541,
710
],
"score": 1.0,
"content": "evaluates each model twice (in parallel) to apply classifier-free guidance, while our distilled models do not,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
68,
707,
541,
722
],
"spans": [
{
"bbox": [
68,
707,
541,
722
],
"score": 1.0,
"content": "since they distilled the effect of guidance into a single model. We provide samples from our original and",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
69,
720,
268,
733
],
"spans": [
{
"bbox": [
69,
720,
268,
733
],
"score": 1.0,
"content": "distilled cascade in Figure 14 for illustration.",
"type": "text"
}
],
"index": 39
}
],
"index": 36.5,
"bbox_fs": [
68,
658,
542,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
66,
92,
550,
335
],
"blocks": [
{
"type": "image_caption",
"bbox": [
258,
83,
356,
92
],
"group_id": 0,
"lines": [
{
"bbox": [
258,
81,
357,
93
],
"spans": [
{
"bbox": [
258,
82,
284,
91
],
"score": 0.6,
"content": "8 0 \\times 4 8",
"type": "inline_equation"
},
{
"bbox": [
284,
81,
357,
93
],
"score": 1.0,
"content": "input video frames",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image_body",
"bbox": [
66,
92,
550,
335
],
"group_id": 0,
"lines": [
{
"bbox": [
66,
92,
550,
335
],
"spans": [
{
"bbox": [
66,
92,
550,
335
],
"score": 0.974,
"type": "image",
"image_path": "a6da3405e211d7c3d0b48d63a33d2636490901a003f68bd8bdb179bf98bb6721.jpg"
}
]
}
],
"index": 2,
"virtual_lines": [
{
"bbox": [
66,
92,
550,
173.0
],
"spans": [],
"index": 1
},
{
"bbox": [
66,
173.0,
550,
254.0
],
"spans": [],
"index": 2
},
{
"bbox": [
66,
254.0,
550,
335.0
],
"spans": [],
"index": 3
}
]
},
{
"type": "image_caption",
"bbox": [
70,
347,
541,
395
],
"group_id": 0,
"lines": [
{
"bbox": [
68,
346,
542,
362
],
"spans": [
{
"bbox": [
68,
346,
234,
362
],
"score": 1.0,
"content": "Figure 12: Comparison between",
"type": "text"
},
{
"bbox": [
234,
353,
239,
358
],
"score": 0.26,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
239,
346,
542,
362
],
"score": 1.0,
"content": "-prediction (middle row) and v-prediction (bottom row) for a",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
359,
541,
373
],
"spans": [
{
"bbox": [
69,
359,
76,
373
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
76,
364,
84,
370
],
"score": 0.56,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
84,
359,
93,
373
],
"score": 1.0,
"content": "80",
"type": "text"
},
{
"bbox": [
94,
363,
102,
370
],
"score": 0.6,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
102,
359,
111,
373
],
"score": 1.0,
"content": "48",
"type": "text"
},
{
"bbox": [
112,
363,
122,
370
],
"score": 0.35,
"content": "",
"type": "inline_equation"
},
{
"bbox": [
122,
359,
126,
373
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
127,
363,
135,
370
],
"score": 0.66,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
135,
359,
149,
373
],
"score": 1.0,
"content": "320",
"type": "text"
},
{
"bbox": [
149,
363,
157,
370
],
"score": 0.48,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
158,
359,
541,
373
],
"score": 1.0,
"content": "192 spatial super-resolution architecture at 200k training steps. The frames from the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
71,
371,
542,
384
],
"spans": [
{
"bbox": [
71,
377,
76,
381
],
"score": 0.64,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
77,
371,
542,
384
],
"score": 1.0,
"content": "-prediction model are generally worse, suffering from unnatural global color shifts across frames. The",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
383,
370,
396
],
"spans": [
{
"bbox": [
70,
383,
370,
396
],
"score": 1.0,
"content": "frames from the v-prediction model do not and are more consistent.",
"type": "text"
}
],
"index": 7
}
],
"index": 5.5
}
],
"index": 2
},
{
"type": "image",
"bbox": [
205,
411,
403,
501
],
"blocks": [
{
"type": "image_body",
"bbox": [
205,
411,
403,
501
],
"group_id": 1,
"lines": [
{
"bbox": [
205,
411,
403,
501
],
"spans": [
{
"bbox": [
205,
411,
403,
501
],
"score": 0.967,
"type": "image",
"image_path": "afec2a06f98752ec7650e99490a85f9122e529bab0d49bc4763af3f3a487a8eb.jpg"
}
]
}
],
"index": 10.5,
"virtual_lines": [
{
"bbox": [
205,
411,
403,
426.0
],
"spans": [],
"index": 8
},
{
"bbox": [
205,
426.0,
403,
441.0
],
"spans": [],
"index": 9
},
{
"bbox": [
205,
441.0,
403,
456.0
],
"spans": [],
"index": 10
},
{
"bbox": [
205,
456.0,
403,
471.0
],
"spans": [],
"index": 11
},
{
"bbox": [
205,
471.0,
403,
486.0
],
"spans": [],
"index": 12
},
{
"bbox": [
205,
486.0,
403,
501.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "image_caption",
"bbox": [
70,
511,
541,
559
],
"group_id": 1,
"lines": [
{
"bbox": [
69,
510,
541,
524
],
"spans": [
{
"bbox": [
69,
510,
214,
524
],
"score": 1.0,
"content": "Figure 13: Comparison between",
"type": "text"
},
{
"bbox": [
215,
513,
299,
522
],
"score": 0.76,
"content": "8 0 \\times 4 8 3 2 0 \\times 1 9 2",
"type": "inline_equation"
},
{
"bbox": [
299,
510,
415,
524
],
"score": 1.0,
"content": "SSR models trained with",
"type": "text"
},
{
"bbox": [
415,
516,
420,
521
],
"score": 0.7,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
421,
510,
447,
524
],
"score": 1.0,
"content": "- and",
"type": "text"
},
{
"bbox": [
447,
516,
453,
522
],
"score": 0.26,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
454,
510,
541,
524
],
"score": 1.0,
"content": "-prediction parame-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
523,
541,
537
],
"spans": [
{
"bbox": [
69,
523,
541,
537
],
"score": 1.0,
"content": "terizations. We report FID evaluated on the first upsampled frame; FVD score is excessively noisy for the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
71,
534,
542,
549
],
"spans": [
{
"bbox": [
71,
541,
76,
545
],
"score": 0.75,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
77,
534,
357,
549
],
"score": 1.0,
"content": "-prediction model. We observe that the sample quality of the",
"type": "text"
},
{
"bbox": [
358,
541,
363,
545
],
"score": 0.84,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
363,
534,
542,
549
],
"score": 1.0,
"content": "-prediction model converges much more",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
70,
548,
261,
559
],
"spans": [
{
"bbox": [
70,
548,
261,
559
],
"score": 1.0,
"content": "slowly than that of the v-prediction model.",
"type": "text"
}
],
"index": 17
}
],
"index": 15.5
}
],
"index": 13.0
},
{
"type": "title",
"bbox": [
70,
583,
267,
598
],
"lines": [
{
"bbox": [
68,
582,
268,
600
],
"spans": [
{
"bbox": [
68,
582,
268,
600
],
"score": 1.0,
"content": "4 Limitations and Societal Impact",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
70,
612,
540,
732
],
"lines": [
{
"bbox": [
68,
612,
541,
627
],
"spans": [
{
"bbox": [
68,
612,
541,
627
],
"score": 1.0,
"content": "Generative modeling has made tremendous progress, especially in recent text-to-image models (Saharia",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
69,
624,
541,
638
],
"spans": [
{
"bbox": [
69,
624,
541,
638
],
"score": 1.0,
"content": "et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022). Imagen Video is another step forward in generative",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
636,
541,
650
],
"spans": [
{
"bbox": [
69,
636,
541,
650
],
"score": 1.0,
"content": "modelling capabilities, advancing text-to-video AI systems. Video generative models can be used to positively",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
649,
540,
661
],
"spans": [
{
"bbox": [
69,
649,
540,
661
],
"score": 1.0,
"content": "impact society, for example by amplifying and augmenting human creativity. However, these generative",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
660,
540,
673
],
"spans": [
{
"bbox": [
69,
660,
540,
673
],
"score": 1.0,
"content": "models may also be misused, for example to generate fake, hateful, explicit or harmful content. We have taken",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
68,
672,
541,
687
],
"spans": [
{
"bbox": [
68,
672,
541,
687
],
"score": 1.0,
"content": "multiple steps to minimize these concerns, for example in internal trials, we apply input text prompt filtering,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
685,
541,
698
],
"spans": [
{
"bbox": [
69,
685,
541,
698
],
"score": 1.0,
"content": "and output video content filtering. However, there are several important safety and ethical challenges",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
696,
539,
709
],
"spans": [
{
"bbox": [
69,
696,
531,
709
],
"score": 1.0,
"content": "remaining. Imagen Video and its frozen T5-XXL text encoder were trained on problematic data (Bordia",
"type": "text"
},
{
"bbox": [
531,
699,
539,
707
],
"score": 0.27,
"content": "\\&",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
69,
708,
540,
722
],
"spans": [
{
"bbox": [
69,
708,
540,
722
],
"score": 1.0,
"content": "Bowman, 2017; Birhane et al., 2021; Bender et al., 2021). While our internal testing suggests much of explicit",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
68,
718,
541,
735
],
"spans": [
{
"bbox": [
68,
718,
541,
735
],
"score": 1.0,
"content": "and violent content can be filtered out, there still exists social biases and stereotypes which are challenging",
"type": "text"
}
],
"index": 28
}
],
"index": 23.5
}
],
"page_idx": 13,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "14",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
66,
92,
550,
335
],
"blocks": [
{
"type": "image_caption",
"bbox": [
258,
83,
356,
92
],
"group_id": 0,
"lines": [
{
"bbox": [
258,
81,
357,
93
],
"spans": [
{
"bbox": [
258,
82,
284,
91
],
"score": 0.6,
"content": "8 0 \\times 4 8",
"type": "inline_equation"
},
{
"bbox": [
284,
81,
357,
93
],
"score": 1.0,
"content": "input video frames",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image_body",
"bbox": [
66,
92,
550,
335
],
"group_id": 0,
"lines": [
{
"bbox": [
66,
92,
550,
335
],
"spans": [
{
"bbox": [
66,
92,
550,
335
],
"score": 0.974,
"type": "image",
"image_path": "a6da3405e211d7c3d0b48d63a33d2636490901a003f68bd8bdb179bf98bb6721.jpg"
}
]
}
],
"index": 2,
"virtual_lines": [
{
"bbox": [
66,
92,
550,
173.0
],
"spans": [],
"index": 1
},
{
"bbox": [
66,
173.0,
550,
254.0
],
"spans": [],
"index": 2
},
{
"bbox": [
66,
254.0,
550,
335.0
],
"spans": [],
"index": 3
}
]
},
{
"type": "image_caption",
"bbox": [
70,
347,
541,
395
],
"group_id": 0,
"lines": [
{
"bbox": [
68,
346,
542,
362
],
"spans": [
{
"bbox": [
68,
346,
234,
362
],
"score": 1.0,
"content": "Figure 12: Comparison between",
"type": "text"
},
{
"bbox": [
234,
353,
239,
358
],
"score": 0.26,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
239,
346,
542,
362
],
"score": 1.0,
"content": "-prediction (middle row) and v-prediction (bottom row) for a",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
359,
541,
373
],
"spans": [
{
"bbox": [
69,
359,
76,
373
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
76,
364,
84,
370
],
"score": 0.56,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
84,
359,
93,
373
],
"score": 1.0,
"content": "80",
"type": "text"
},
{
"bbox": [
94,
363,
102,
370
],
"score": 0.6,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
102,
359,
111,
373
],
"score": 1.0,
"content": "48",
"type": "text"
},
{
"bbox": [
112,
363,
122,
370
],
"score": 0.35,
"content": "",
"type": "inline_equation"
},
{
"bbox": [
122,
359,
126,
373
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
127,
363,
135,
370
],
"score": 0.66,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
135,
359,
149,
373
],
"score": 1.0,
"content": "320",
"type": "text"
},
{
"bbox": [
149,
363,
157,
370
],
"score": 0.48,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
158,
359,
541,
373
],
"score": 1.0,
"content": "192 spatial super-resolution architecture at 200k training steps. The frames from the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
71,
371,
542,
384
],
"spans": [
{
"bbox": [
71,
377,
76,
381
],
"score": 0.64,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
77,
371,
542,
384
],
"score": 1.0,
"content": "-prediction model are generally worse, suffering from unnatural global color shifts across frames. The",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
383,
370,
396
],
"spans": [
{
"bbox": [
70,
383,
370,
396
],
"score": 1.0,
"content": "frames from the v-prediction model do not and are more consistent.",
"type": "text"
}
],
"index": 7
}
],
"index": 5.5
}
],
"index": 2
},
{
"type": "image",
"bbox": [
205,
411,
403,
501
],
"blocks": [
{
"type": "image_body",
"bbox": [
205,
411,
403,
501
],
"group_id": 1,
"lines": [
{
"bbox": [
205,
411,
403,
501
],
"spans": [
{
"bbox": [
205,
411,
403,
501
],
"score": 0.967,
"type": "image",
"image_path": "afec2a06f98752ec7650e99490a85f9122e529bab0d49bc4763af3f3a487a8eb.jpg"
}
]
}
],
"index": 10.5,
"virtual_lines": [
{
"bbox": [
205,
411,
403,
426.0
],
"spans": [],
"index": 8
},
{
"bbox": [
205,
426.0,
403,
441.0
],
"spans": [],
"index": 9
},
{
"bbox": [
205,
441.0,
403,
456.0
],
"spans": [],
"index": 10
},
{
"bbox": [
205,
456.0,
403,
471.0
],
"spans": [],
"index": 11
},
{
"bbox": [
205,
471.0,
403,
486.0
],
"spans": [],
"index": 12
},
{
"bbox": [
205,
486.0,
403,
501.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "image_caption",
"bbox": [
70,
511,
541,
559
],
"group_id": 1,
"lines": [
{
"bbox": [
69,
510,
541,
524
],
"spans": [
{
"bbox": [
69,
510,
214,
524
],
"score": 1.0,
"content": "Figure 13: Comparison between",
"type": "text"
},
{
"bbox": [
215,
513,
299,
522
],
"score": 0.76,
"content": "8 0 \\times 4 8 3 2 0 \\times 1 9 2",
"type": "inline_equation"
},
{
"bbox": [
299,
510,
415,
524
],
"score": 1.0,
"content": "SSR models trained with",
"type": "text"
},
{
"bbox": [
415,
516,
420,
521
],
"score": 0.7,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
421,
510,
447,
524
],
"score": 1.0,
"content": "- and",
"type": "text"
},
{
"bbox": [
447,
516,
453,
522
],
"score": 0.26,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
454,
510,
541,
524
],
"score": 1.0,
"content": "-prediction parame-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
523,
541,
537
],
"spans": [
{
"bbox": [
69,
523,
541,
537
],
"score": 1.0,
"content": "terizations. We report FID evaluated on the first upsampled frame; FVD score is excessively noisy for the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
71,
534,
542,
549
],
"spans": [
{
"bbox": [
71,
541,
76,
545
],
"score": 0.75,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
77,
534,
357,
549
],
"score": 1.0,
"content": "-prediction model. We observe that the sample quality of the",
"type": "text"
},
{
"bbox": [
358,
541,
363,
545
],
"score": 0.84,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
363,
534,
542,
549
],
"score": 1.0,
"content": "-prediction model converges much more",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
70,
548,
261,
559
],
"spans": [
{
"bbox": [
70,
548,
261,
559
],
"score": 1.0,
"content": "slowly than that of the v-prediction model.",
"type": "text"
}
],
"index": 17
}
],
"index": 15.5
}
],
"index": 13.0
},
{
"type": "title",
"bbox": [
70,
583,
267,
598
],
"lines": [
{
"bbox": [
68,
582,
268,
600
],
"spans": [
{
"bbox": [
68,
582,
268,
600
],
"score": 1.0,
"content": "4 Limitations and Societal Impact",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
70,
612,
540,
732
],
"lines": [
{
"bbox": [
68,
612,
541,
627
],
"spans": [
{
"bbox": [
68,
612,
541,
627
],
"score": 1.0,
"content": "Generative modeling has made tremendous progress, especially in recent text-to-image models (Saharia",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
69,
624,
541,
638
],
"spans": [
{
"bbox": [
69,
624,
541,
638
],
"score": 1.0,
"content": "et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022). Imagen Video is another step forward in generative",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
69,
636,
541,
650
],
"spans": [
{
"bbox": [
69,
636,
541,
650
],
"score": 1.0,
"content": "modelling capabilities, advancing text-to-video AI systems. Video generative models can be used to positively",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
649,
540,
661
],
"spans": [
{
"bbox": [
69,
649,
540,
661
],
"score": 1.0,
"content": "impact society, for example by amplifying and augmenting human creativity. However, these generative",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
660,
540,
673
],
"spans": [
{
"bbox": [
69,
660,
540,
673
],
"score": 1.0,
"content": "models may also be misused, for example to generate fake, hateful, explicit or harmful content. We have taken",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
68,
672,
541,
687
],
"spans": [
{
"bbox": [
68,
672,
541,
687
],
"score": 1.0,
"content": "multiple steps to minimize these concerns, for example in internal trials, we apply input text prompt filtering,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
69,
685,
541,
698
],
"spans": [
{
"bbox": [
69,
685,
541,
698
],
"score": 1.0,
"content": "and output video content filtering. However, there are several important safety and ethical challenges",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
696,
539,
709
],
"spans": [
{
"bbox": [
69,
696,
531,
709
],
"score": 1.0,
"content": "remaining. Imagen Video and its frozen T5-XXL text encoder were trained on problematic data (Bordia",
"type": "text"
},
{
"bbox": [
531,
699,
539,
707
],
"score": 0.27,
"content": "\\&",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
69,
708,
540,
722
],
"spans": [
{
"bbox": [
69,
708,
540,
722
],
"score": 1.0,
"content": "Bowman, 2017; Birhane et al., 2021; Bender et al., 2021). While our internal testing suggests much of explicit",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
68,
718,
541,
735
],
"spans": [
{
"bbox": [
68,
718,
541,
735
],
"score": 1.0,
"content": "and violent content can be filtered out, there still exists social biases and stereotypes which are challenging",
"type": "text"
}
],
"index": 28
}
],
"index": 23.5,
"bbox_fs": [
68,
612,
541,
735
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
110,
80,
500,
190
],
"blocks": [
{
"type": "table_body",
"bbox": [
110,
80,
500,
190
],
"group_id": 0,
"lines": [
{
"bbox": [
110,
80,
500,
190
],
"spans": [
{
"bbox": [
110,
80,
500,
190
],
"score": 0.981,
"html": "| Guidance w | Base Steps | SR Steps | CLIP Score | CLIP R-Precision | Sampling Time |
| constant=6 | 256 | 128 | 25.19±.03 | 92.12±.53 | 618 sec |
| oscillate(15,1) | 256 | 128 | 25.02±.08 | 89.91±.96 | 618 sec |
| constant=6 | 256 | 8 | 25.29±.05 | 90.88±.50 | 135 sec |
| oscillate(15,1) | 256 | 8 | 25.15±.09 | 88.78±.69 | 135 sec |
| constant=6 | 8 | 8 | 25.03±.05 | 89.68±.38 | 35 sec |
| oscillate(15,1) | 8 | 8 | 25.12±.07 | 90.97±.46 | 35 sec |
| ground truth | | | 24.27 | 86.18 | |
",
"type": "table",
"image_path": "3e4a030f9c6d910b6ae7437a8f7f1c0f96a01de739df18d614ace8792d8ed6c2.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
80,
500,
116.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
116.66666666666666,
500,
153.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
153.33333333333331,
500,
189.99999999999997
],
"spans": [],
"index": 2
}
]
}
],
"index": 1
},
{
"type": "text",
"bbox": [
70,
203,
540,
382
],
"lines": [
{
"bbox": [
69,
202,
541,
216
],
"spans": [
{
"bbox": [
69,
202,
541,
216
],
"score": 1.0,
"content": "Table 1: CLIP scores and CLIP R-Precision (Park et al., 2021) values for generated samples and ground",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
215,
542,
230
],
"spans": [
{
"bbox": [
70,
215,
542,
230
],
"score": 1.0,
"content": "truth videos on prompts from our test set. Cells highlighted in green represent distilled models. We",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
227,
541,
240
],
"spans": [
{
"bbox": [
69,
227,
541,
240
],
"score": 1.0,
"content": "compare three different combinations: original pipeline, distilled SR models on top of original base model,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
239,
542,
253
],
"spans": [
{
"bbox": [
69,
239,
542,
253
],
"score": 1.0,
"content": "and fully distilled pipeline. The original base models use 256 sampling steps, and original SR models use",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
251,
541,
264
],
"spans": [
{
"bbox": [
70,
251,
541,
264
],
"score": 1.0,
"content": "128 steps. All distilled models use 8 sampling steps per stage. Sampling from the original pipeline takes 618",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
263,
541,
276
],
"spans": [
{
"bbox": [
69,
263,
541,
276
],
"score": 1.0,
"content": "seconds for one batch of samples, while sampling from the distilled pipeline takes 35 seconds, making the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
275,
541,
288
],
"spans": [
{
"bbox": [
69,
275,
186,
288
],
"score": 1.0,
"content": "distilled pipeline about 18",
"type": "text"
},
{
"bbox": [
187,
279,
194,
286
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
195,
275,
541,
288
],
"score": 1.0,
"content": "faster. We also explored two different classifier-free guidance settings for the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
288,
541,
299
],
"spans": [
{
"bbox": [
69,
288,
236,
299
],
"score": 1.0,
"content": "base models: constant guidance with",
"type": "text"
},
{
"bbox": [
236,
290,
262,
297
],
"score": 0.91,
"content": "w = 6",
"type": "inline_equation"
},
{
"bbox": [
263,
288,
488,
299
],
"score": 1.0,
"content": "and oscillating guidance which alternates between",
"type": "text"
},
{
"bbox": [
488,
290,
520,
297
],
"score": 0.91,
"content": "w = 1 5",
"type": "inline_equation"
},
{
"bbox": [
520,
288,
541,
299
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
71,
299,
540,
312
],
"spans": [
{
"bbox": [
71,
302,
97,
309
],
"score": 0.9,
"content": "w = 1",
"type": "inline_equation"
},
{
"bbox": [
97,
299,
540,
312
],
"score": 1.0,
"content": ", following Saharia et al. (2022b). When using oscillating guidance, the fully distilled pipeline performs",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
70,
312,
540,
323
],
"spans": [
{
"bbox": [
70,
312,
540,
323
],
"score": 1.0,
"content": "the same as the original model, or even slightly better. When using fixed guidance, our fully distilled pipeline",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
68,
323,
541,
335
],
"spans": [
{
"bbox": [
68,
323,
541,
335
],
"score": 1.0,
"content": "scores slightly lower than the original model, though the difference is minor. Combining the original base",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
335,
541,
348
],
"spans": [
{
"bbox": [
69,
335,
541,
348
],
"score": 1.0,
"content": "model with fixed guidance and distilled super-resolution models produced the highest CLIP score. For all",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
347,
541,
360
],
"spans": [
{
"bbox": [
69,
347,
541,
360
],
"score": 1.0,
"content": "models, generated samples obtain better perceptual quality metrics than the original ground truth data:",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
358,
541,
373
],
"spans": [
{
"bbox": [
69,
358,
541,
373
],
"score": 1.0,
"content": "By using classifier-free guidance our models sample from a distribution tilted towards these quality metrics,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
68,
369,
411,
385
],
"spans": [
{
"bbox": [
68,
369,
411,
385
],
"score": 1.0,
"content": "rather than from an accurate approximation of the original data distribution.",
"type": "text"
}
],
"index": 17
}
],
"index": 10
},
{
"type": "image",
"bbox": [
78,
396,
534,
533
],
"blocks": [
{
"type": "image_body",
"bbox": [
78,
396,
534,
533
],
"group_id": 0,
"lines": [
{
"bbox": [
78,
396,
534,
533
],
"spans": [
{
"bbox": [
78,
396,
534,
533
],
"score": 0.96,
"type": "image",
"image_path": "b497801badbb304195e8c26f2c8d67c7fa5c01f65e1dfcc29bfb88e1439d2f27.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
78,
396,
534,
441.6666666666667
],
"spans": [],
"index": 18
},
{
"bbox": [
78,
441.6666666666667,
534,
487.33333333333337
],
"spans": [],
"index": 19
},
{
"bbox": [
78,
487.33333333333337,
534,
533.0
],
"spans": [],
"index": 20
}
]
},
{
"type": "image_caption",
"bbox": [
70,
545,
540,
594
],
"group_id": 0,
"lines": [
{
"bbox": [
68,
543,
543,
561
],
"spans": [
{
"bbox": [
68,
543,
444,
561
],
"score": 1.0,
"content": "Figure 14: Frames from videos generated by Imagen Video for the text prompt “",
"type": "text"
},
{
"bbox": [
444,
546,
452,
556
],
"score": 0.3,
"content": "A",
"type": "inline_equation"
},
{
"bbox": [
453,
543,
543,
561
],
"score": 1.0,
"content": "teddy bear wearing",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
70,
558,
541,
571
],
"spans": [
{
"bbox": [
70,
558,
541,
571
],
"score": 1.0,
"content": "sunglasses playing guitar next to a cactus.” The samples on the left are produced by our original model",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
68,
568,
541,
584
],
"spans": [
{
"bbox": [
68,
568,
541,
584
],
"score": 1.0,
"content": "cascade, while the samples on the right are from our distilled cascade with 8 sampling steps per stage. Both",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
579,
313,
596
],
"spans": [
{
"bbox": [
69,
579,
198,
596
],
"score": 1.0,
"content": "used constant guidance with",
"type": "text"
},
{
"bbox": [
198,
585,
224,
591
],
"score": 0.89,
"content": "w = 6",
"type": "inline_equation"
},
{
"bbox": [
225,
579,
313,
596
],
"score": 1.0,
"content": "and static clipping.",
"type": "text"
}
],
"index": 24
}
],
"index": 22.5
}
],
"index": 20.75
},
{
"type": "text",
"bbox": [
69,
616,
538,
641
],
"lines": [
{
"bbox": [
69,
616,
541,
630
],
"spans": [
{
"bbox": [
69,
616,
541,
630
],
"score": 1.0,
"content": "to detect and filter. We have decided not to release the Imagen Video model or its source code until these",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
630,
174,
641
],
"spans": [
{
"bbox": [
69,
630,
174,
641
],
"score": 1.0,
"content": "concerns are mitigated.",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5
},
{
"type": "title",
"bbox": [
71,
657,
150,
671
],
"lines": [
{
"bbox": [
68,
654,
152,
674
],
"spans": [
{
"bbox": [
68,
654,
152,
674
],
"score": 1.0,
"content": "5 Conclusion",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
71,
684,
540,
732
],
"lines": [
{
"bbox": [
70,
684,
541,
697
],
"spans": [
{
"bbox": [
70,
684,
541,
697
],
"score": 1.0,
"content": "We presented Imagen Video: a text-conditional video generation system based on a cascade of video diffusion",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
695,
542,
710
],
"spans": [
{
"bbox": [
69,
695,
542,
710
],
"score": 1.0,
"content": "models. By extending the text-to-image diffusion models of Imagen (Saharia et al., 2022b) to the time",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
707,
541,
722
],
"spans": [
{
"bbox": [
69,
707,
541,
722
],
"score": 1.0,
"content": "domain, and training jointly on video and images, we obtained a model capable of generating high fidelity",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
719,
542,
734
],
"spans": [
{
"bbox": [
69,
719,
542,
734
],
"score": 1.0,
"content": "videos with good temporal consistency while maintaining the strong features of the original image system,",
"type": "text"
}
],
"index": 31
}
],
"index": 29.5
}
],
"page_idx": 14,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
26,
239,
38
],
"spans": [
{
"bbox": [
71,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "15",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
110,
80,
500,
190
],
"blocks": [
{
"type": "table_body",
"bbox": [
110,
80,
500,
190
],
"group_id": 0,
"lines": [
{
"bbox": [
110,
80,
500,
190
],
"spans": [
{
"bbox": [
110,
80,
500,
190
],
"score": 0.981,
"html": "| Guidance w | Base Steps | SR Steps | CLIP Score | CLIP R-Precision | Sampling Time |
| constant=6 | 256 | 128 | 25.19±.03 | 92.12±.53 | 618 sec |
| oscillate(15,1) | 256 | 128 | 25.02±.08 | 89.91±.96 | 618 sec |
| constant=6 | 256 | 8 | 25.29±.05 | 90.88±.50 | 135 sec |
| oscillate(15,1) | 256 | 8 | 25.15±.09 | 88.78±.69 | 135 sec |
| constant=6 | 8 | 8 | 25.03±.05 | 89.68±.38 | 35 sec |
| oscillate(15,1) | 8 | 8 | 25.12±.07 | 90.97±.46 | 35 sec |
| ground truth | | | 24.27 | 86.18 | |
",
"type": "table",
"image_path": "3e4a030f9c6d910b6ae7437a8f7f1c0f96a01de739df18d614ace8792d8ed6c2.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
80,
500,
116.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
116.66666666666666,
500,
153.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
153.33333333333331,
500,
189.99999999999997
],
"spans": [],
"index": 2
}
]
}
],
"index": 1
},
{
"type": "text",
"bbox": [
70,
203,
540,
382
],
"lines": [
{
"bbox": [
69,
202,
541,
216
],
"spans": [
{
"bbox": [
69,
202,
541,
216
],
"score": 1.0,
"content": "Table 1: CLIP scores and CLIP R-Precision (Park et al., 2021) values for generated samples and ground",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
70,
215,
542,
230
],
"spans": [
{
"bbox": [
70,
215,
542,
230
],
"score": 1.0,
"content": "truth videos on prompts from our test set. Cells highlighted in green represent distilled models. We",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
227,
541,
240
],
"spans": [
{
"bbox": [
69,
227,
541,
240
],
"score": 1.0,
"content": "compare three different combinations: original pipeline, distilled SR models on top of original base model,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
239,
542,
253
],
"spans": [
{
"bbox": [
69,
239,
542,
253
],
"score": 1.0,
"content": "and fully distilled pipeline. The original base models use 256 sampling steps, and original SR models use",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
70,
251,
541,
264
],
"spans": [
{
"bbox": [
70,
251,
541,
264
],
"score": 1.0,
"content": "128 steps. All distilled models use 8 sampling steps per stage. Sampling from the original pipeline takes 618",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
263,
541,
276
],
"spans": [
{
"bbox": [
69,
263,
541,
276
],
"score": 1.0,
"content": "seconds for one batch of samples, while sampling from the distilled pipeline takes 35 seconds, making the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
69,
275,
541,
288
],
"spans": [
{
"bbox": [
69,
275,
186,
288
],
"score": 1.0,
"content": "distilled pipeline about 18",
"type": "text"
},
{
"bbox": [
187,
279,
194,
286
],
"score": 0.74,
"content": "\\times",
"type": "inline_equation"
},
{
"bbox": [
195,
275,
541,
288
],
"score": 1.0,
"content": "faster. We also explored two different classifier-free guidance settings for the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
288,
541,
299
],
"spans": [
{
"bbox": [
69,
288,
236,
299
],
"score": 1.0,
"content": "base models: constant guidance with",
"type": "text"
},
{
"bbox": [
236,
290,
262,
297
],
"score": 0.91,
"content": "w = 6",
"type": "inline_equation"
},
{
"bbox": [
263,
288,
488,
299
],
"score": 1.0,
"content": "and oscillating guidance which alternates between",
"type": "text"
},
{
"bbox": [
488,
290,
520,
297
],
"score": 0.91,
"content": "w = 1 5",
"type": "inline_equation"
},
{
"bbox": [
520,
288,
541,
299
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
71,
299,
540,
312
],
"spans": [
{
"bbox": [
71,
302,
97,
309
],
"score": 0.9,
"content": "w = 1",
"type": "inline_equation"
},
{
"bbox": [
97,
299,
540,
312
],
"score": 1.0,
"content": ", following Saharia et al. (2022b). When using oscillating guidance, the fully distilled pipeline performs",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
70,
312,
540,
323
],
"spans": [
{
"bbox": [
70,
312,
540,
323
],
"score": 1.0,
"content": "the same as the original model, or even slightly better. When using fixed guidance, our fully distilled pipeline",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
68,
323,
541,
335
],
"spans": [
{
"bbox": [
68,
323,
541,
335
],
"score": 1.0,
"content": "scores slightly lower than the original model, though the difference is minor. Combining the original base",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
335,
541,
348
],
"spans": [
{
"bbox": [
69,
335,
541,
348
],
"score": 1.0,
"content": "model with fixed guidance and distilled super-resolution models produced the highest CLIP score. For all",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
347,
541,
360
],
"spans": [
{
"bbox": [
69,
347,
541,
360
],
"score": 1.0,
"content": "models, generated samples obtain better perceptual quality metrics than the original ground truth data:",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
69,
358,
541,
373
],
"spans": [
{
"bbox": [
69,
358,
541,
373
],
"score": 1.0,
"content": "By using classifier-free guidance our models sample from a distribution tilted towards these quality metrics,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
68,
369,
411,
385
],
"spans": [
{
"bbox": [
68,
369,
411,
385
],
"score": 1.0,
"content": "rather than from an accurate approximation of the original data distribution.",
"type": "text"
}
],
"index": 17
}
],
"index": 10,
"bbox_fs": [
68,
202,
542,
385
]
},
{
"type": "image",
"bbox": [
78,
396,
534,
533
],
"blocks": [
{
"type": "image_body",
"bbox": [
78,
396,
534,
533
],
"group_id": 0,
"lines": [
{
"bbox": [
78,
396,
534,
533
],
"spans": [
{
"bbox": [
78,
396,
534,
533
],
"score": 0.96,
"type": "image",
"image_path": "b497801badbb304195e8c26f2c8d67c7fa5c01f65e1dfcc29bfb88e1439d2f27.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
78,
396,
534,
441.6666666666667
],
"spans": [],
"index": 18
},
{
"bbox": [
78,
441.6666666666667,
534,
487.33333333333337
],
"spans": [],
"index": 19
},
{
"bbox": [
78,
487.33333333333337,
534,
533.0
],
"spans": [],
"index": 20
}
]
},
{
"type": "image_caption",
"bbox": [
70,
545,
540,
594
],
"group_id": 0,
"lines": [
{
"bbox": [
68,
543,
543,
561
],
"spans": [
{
"bbox": [
68,
543,
444,
561
],
"score": 1.0,
"content": "Figure 14: Frames from videos generated by Imagen Video for the text prompt “",
"type": "text"
},
{
"bbox": [
444,
546,
452,
556
],
"score": 0.3,
"content": "A",
"type": "inline_equation"
},
{
"bbox": [
453,
543,
543,
561
],
"score": 1.0,
"content": "teddy bear wearing",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
70,
558,
541,
571
],
"spans": [
{
"bbox": [
70,
558,
541,
571
],
"score": 1.0,
"content": "sunglasses playing guitar next to a cactus.” The samples on the left are produced by our original model",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
68,
568,
541,
584
],
"spans": [
{
"bbox": [
68,
568,
541,
584
],
"score": 1.0,
"content": "cascade, while the samples on the right are from our distilled cascade with 8 sampling steps per stage. Both",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
579,
313,
596
],
"spans": [
{
"bbox": [
69,
579,
198,
596
],
"score": 1.0,
"content": "used constant guidance with",
"type": "text"
},
{
"bbox": [
198,
585,
224,
591
],
"score": 0.89,
"content": "w = 6",
"type": "inline_equation"
},
{
"bbox": [
225,
579,
313,
596
],
"score": 1.0,
"content": "and static clipping.",
"type": "text"
}
],
"index": 24
}
],
"index": 22.5
}
],
"index": 20.75
},
{
"type": "text",
"bbox": [
69,
616,
538,
641
],
"lines": [
{
"bbox": [
69,
616,
541,
630
],
"spans": [
{
"bbox": [
69,
616,
541,
630
],
"score": 1.0,
"content": "to detect and filter. We have decided not to release the Imagen Video model or its source code until these",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
630,
174,
641
],
"spans": [
{
"bbox": [
69,
630,
174,
641
],
"score": 1.0,
"content": "concerns are mitigated.",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5,
"bbox_fs": [
69,
616,
541,
641
]
},
{
"type": "title",
"bbox": [
71,
657,
150,
671
],
"lines": [
{
"bbox": [
68,
654,
152,
674
],
"spans": [
{
"bbox": [
68,
654,
152,
674
],
"score": 1.0,
"content": "5 Conclusion",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
71,
684,
540,
732
],
"lines": [
{
"bbox": [
70,
684,
541,
697
],
"spans": [
{
"bbox": [
70,
684,
541,
697
],
"score": 1.0,
"content": "We presented Imagen Video: a text-conditional video generation system based on a cascade of video diffusion",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
695,
542,
710
],
"spans": [
{
"bbox": [
69,
695,
542,
710
],
"score": 1.0,
"content": "models. By extending the text-to-image diffusion models of Imagen (Saharia et al., 2022b) to the time",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
69,
707,
541,
722
],
"spans": [
{
"bbox": [
69,
707,
541,
722
],
"score": 1.0,
"content": "domain, and training jointly on video and images, we obtained a model capable of generating high fidelity",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
69,
719,
542,
734
],
"spans": [
{
"bbox": [
69,
719,
542,
734
],
"score": 1.0,
"content": "videos with good temporal consistency while maintaining the strong features of the original image system,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
69,
82,
542,
96
],
"spans": [
{
"bbox": [
69,
82,
542,
96
],
"score": 1.0,
"content": "such as the ability to accurately spell text. We transferred multiple methods from the image domain to",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
70,
94,
542,
108
],
"spans": [
{
"bbox": [
70,
94,
133,
108
],
"score": 1.0,
"content": "video, such as",
"type": "text",
"cross_page": true
},
{
"bbox": [
134,
99,
140,
104
],
"score": 0.26,
"content": "\\mathbf { v }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
140,
94,
542,
108
],
"score": 1.0,
"content": "-parameterization (Salimans & Ho, 2022), conditioning augmentation (Ho et al., 2022a), and",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
69,
104,
541,
120
],
"spans": [
{
"bbox": [
69,
104,
541,
120
],
"score": 1.0,
"content": "classifier-free guidance (Ho & Salimans, 2021), and found that these are also useful in the video setting.",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
70,
118,
541,
132
],
"spans": [
{
"bbox": [
70,
118,
541,
132
],
"score": 1.0,
"content": "Video modeling is computationally demanding, and we found that progressive distillation (Salimans & Ho,",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
69,
129,
540,
144
],
"spans": [
{
"bbox": [
69,
129,
540,
144
],
"score": 1.0,
"content": "2022; Meng et al., 2022) is a valuable technique for speeding up video diffusion models at sampling time.",
"type": "text",
"cross_page": true
}
],
"index": 4
},
{
"bbox": [
69,
141,
541,
156
],
"spans": [
{
"bbox": [
69,
141,
541,
156
],
"score": 1.0,
"content": "Given the tremendous recent progress in generative modeling, we believe there is ample scope for further",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
69,
154,
340,
167
],
"spans": [
{
"bbox": [
69,
154,
340,
167
],
"score": 1.0,
"content": "improvements in video generation capabilities in future work.",
"type": "text",
"cross_page": true
}
],
"index": 6
}
],
"index": 29.5,
"bbox_fs": [
69,
684,
542,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
165
],
"lines": [
{
"bbox": [
69,
82,
542,
96
],
"spans": [
{
"bbox": [
69,
82,
542,
96
],
"score": 1.0,
"content": "such as the ability to accurately spell text. We transferred multiple methods from the image domain to",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
70,
94,
542,
108
],
"spans": [
{
"bbox": [
70,
94,
133,
108
],
"score": 1.0,
"content": "video, such as",
"type": "text"
},
{
"bbox": [
134,
99,
140,
104
],
"score": 0.26,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
140,
94,
542,
108
],
"score": 1.0,
"content": "-parameterization (Salimans & Ho, 2022), conditioning augmentation (Ho et al., 2022a), and",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
69,
104,
541,
120
],
"spans": [
{
"bbox": [
69,
104,
541,
120
],
"score": 1.0,
"content": "classifier-free guidance (Ho & Salimans, 2021), and found that these are also useful in the video setting.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
70,
118,
541,
132
],
"spans": [
{
"bbox": [
70,
118,
541,
132
],
"score": 1.0,
"content": "Video modeling is computationally demanding, and we found that progressive distillation (Salimans & Ho,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
69,
129,
540,
144
],
"spans": [
{
"bbox": [
69,
129,
540,
144
],
"score": 1.0,
"content": "2022; Meng et al., 2022) is a valuable technique for speeding up video diffusion models at sampling time.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
69,
141,
541,
156
],
"spans": [
{
"bbox": [
69,
141,
541,
156
],
"score": 1.0,
"content": "Given the tremendous recent progress in generative modeling, we believe there is ample scope for further",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
154,
340,
167
],
"spans": [
{
"bbox": [
69,
154,
340,
167
],
"score": 1.0,
"content": "improvements in video generation capabilities in future work.",
"type": "text"
}
],
"index": 6
}
],
"index": 3
},
{
"type": "title",
"bbox": [
71,
181,
131,
195
],
"lines": [
{
"bbox": [
69,
181,
132,
197
],
"spans": [
{
"bbox": [
69,
181,
132,
197
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
67,
201,
543,
729
],
"lines": [
{
"bbox": [
69,
200,
541,
217
],
"spans": [
{
"bbox": [
69,
200,
541,
217
],
"score": 1.0,
"content": "Mohammad Babaeizadeh, Chelsea Finn, D. Erhan, Roy H. Campbell, and Sergey Levine. Stochastic varia-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
79,
214,
317,
228
],
"spans": [
{
"bbox": [
79,
214,
317,
228
],
"score": 1.0,
"content": "tional video prediction. ArXiv, abs/1710.11252, 2018.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
69,
235,
541,
249
],
"spans": [
{
"bbox": [
69,
235,
541,
249
],
"score": 1.0,
"content": "Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
79,
246,
540,
261
],
"spans": [
{
"bbox": [
79,
246,
540,
261
],
"score": 1.0,
"content": "Erhan. Fitvid: Overfitting in pixel-level video prediction, 2021. URL https://arxiv.org/abs/2106.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
80,
258,
113,
273
],
"spans": [
{
"bbox": [
80,
258,
113,
273
],
"score": 1.0,
"content": "13195.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
69,
279,
542,
295
],
"spans": [
{
"bbox": [
69,
279,
542,
295
],
"score": 1.0,
"content": "Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
79,
292,
487,
307
],
"spans": [
{
"bbox": [
79,
292,
487,
307
],
"score": 1.0,
"content": "stochastic parrots: Can language models be too big? . In Proceedings of FAccT 2021, 2021.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
68,
310,
541,
329
],
"spans": [
{
"bbox": [
68,
310,
541,
329
],
"score": 1.0,
"content": "Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornog-",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
78,
325,
354,
339
],
"spans": [
{
"bbox": [
78,
325,
354,
339
],
"score": 1.0,
"content": "raphy, and malignant stereotypes. In arXiv:2110.01963, 2021.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
345,
541,
361
],
"spans": [
{
"bbox": [
69,
345,
541,
361
],
"score": 1.0,
"content": "Shikha Bordia and Samuel R. Bowman. Identifying and Reducing Gender Bias in Word-Level Language",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
79,
358,
199,
372
],
"spans": [
{
"bbox": [
79,
358,
199,
372
],
"score": 1.0,
"content": "Models. In NAACL, 2017.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
378,
541,
393
],
"spans": [
{
"bbox": [
69,
378,
541,
393
],
"score": 1.0,
"content": "Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, and William Chan. WaveGrad:",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
78,
390,
370,
406
],
"spans": [
{
"bbox": [
78,
390,
370,
406
],
"score": 1.0,
"content": "Estimating Gradients for Waveform Generation. In ICLR, 2021a.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
68,
410,
541,
428
],
"spans": [
{
"bbox": [
68,
410,
541,
428
],
"score": 1.0,
"content": "Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, Najim Dehak, and William Chan.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
79,
424,
492,
438
],
"spans": [
{
"bbox": [
79,
424,
492,
438
],
"score": 1.0,
"content": "WaveGrad 2: Iterative Refinement for Text-to-Speech Synthesis . In INTERSPEECH, 2021b.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
69,
443,
520,
459
],
"spans": [
{
"bbox": [
69,
443,
520,
459
],
"score": 1.0,
"content": "Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2022.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
465,
542,
481
],
"spans": [
{
"bbox": [
69,
465,
542,
481
],
"score": 1.0,
"content": "Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
79,
478,
541,
493
],
"spans": [
{
"bbox": [
79,
478,
541,
493
],
"score": 1.0,
"content": "Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers, 2021.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
79,
489,
279,
505
],
"spans": [
{
"bbox": [
79,
489,
279,
505
],
"score": 1.0,
"content": "URL https://arxiv.org/abs/2105.13290.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
509,
542,
526
],
"spans": [
{
"bbox": [
68,
509,
542,
526
],
"score": 1.0,
"content": "Chelsea Finn, Ian J. Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
79,
523,
289,
537
],
"spans": [
{
"bbox": [
79,
523,
289,
537
],
"score": 1.0,
"content": "video prediction. ArXiv, abs/1605.07157, 2016.",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
70,
545,
540,
558
],
"spans": [
{
"bbox": [
70,
545,
540,
558
],
"score": 1.0,
"content": "Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Mart’in-Mart’in, and Li Fei-Fei. Maskvit:",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
78,
555,
423,
570
],
"spans": [
{
"bbox": [
78,
555,
423,
570
],
"score": 1.0,
"content": "Masked visual pre-training for video prediction. ArXiv, abs/2206.11894, 2022.",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
68,
575,
542,
592
],
"spans": [
{
"bbox": [
68,
575,
542,
592
],
"score": 1.0,
"content": "William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
78,
588,
322,
604
],
"spans": [
{
"bbox": [
78,
588,
322,
604
],
"score": 1.0,
"content": "modeling of long videos. ArXiv, abs/2205.11495, 2022.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
68,
609,
541,
624
],
"spans": [
{
"bbox": [
68,
609,
541,
624
],
"score": 1.0,
"content": "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
79,
622,
430,
636
],
"spans": [
{
"bbox": [
79,
622,
430,
636
],
"score": 1.0,
"content": "evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
641,
541,
657
],
"spans": [
{
"bbox": [
69,
641,
541,
657
],
"score": 1.0,
"content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
78,
654,
542,
670
],
"spans": [
{
"bbox": [
78,
654,
542,
670
],
"score": 1.0,
"content": "trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
79,
667,
189,
679
],
"spans": [
{
"bbox": [
79,
667,
189,
679
],
"score": 1.0,
"content": "arXiv:1706.08500, 2017.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
68,
686,
542,
703
],
"spans": [
{
"bbox": [
68,
686,
542,
703
],
"score": 1.0,
"content": "Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
80,
699,
325,
714
],
"spans": [
{
"bbox": [
80,
699,
325,
714
],
"score": 1.0,
"content": "Generative Models and Downstream Applications, 2021.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
69,
720,
522,
735
],
"spans": [
{
"bbox": [
69,
720,
522,
735
],
"score": 1.0,
"content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.",
"type": "text"
}
],
"index": 41
}
],
"index": 24
},
{
"type": "text",
"bbox": [
68,
720,
522,
732
],
"lines": [
{
"bbox": [
68,
720,
522,
732
],
"spans": [],
"index": 40
}
],
"index": 40
}
],
"page_idx": 15,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
70,
26,
239,
38
],
"spans": [
{
"bbox": [
70,
26,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "16",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
71,
82,
540,
165
],
"lines": [],
"index": 3,
"bbox_fs": [
69,
82,
542,
167
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
71,
181,
131,
195
],
"lines": [
{
"bbox": [
69,
181,
132,
197
],
"spans": [
{
"bbox": [
69,
181,
132,
197
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "list",
"bbox": [
67,
201,
543,
729
],
"lines": [
{
"bbox": [
69,
200,
541,
217
],
"spans": [
{
"bbox": [
69,
200,
541,
217
],
"score": 1.0,
"content": "Mohammad Babaeizadeh, Chelsea Finn, D. Erhan, Roy H. Campbell, and Sergey Levine. Stochastic varia-",
"type": "text"
}
],
"index": 8,
"is_list_start_line": true
},
{
"bbox": [
79,
214,
317,
228
],
"spans": [
{
"bbox": [
79,
214,
317,
228
],
"score": 1.0,
"content": "tional video prediction. ArXiv, abs/1710.11252, 2018.",
"type": "text"
}
],
"index": 9,
"is_list_end_line": true
},
{
"bbox": [
69,
235,
541,
249
],
"spans": [
{
"bbox": [
69,
235,
541,
249
],
"score": 1.0,
"content": "Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru",
"type": "text"
}
],
"index": 10,
"is_list_start_line": true
},
{
"bbox": [
79,
246,
540,
261
],
"spans": [
{
"bbox": [
79,
246,
540,
261
],
"score": 1.0,
"content": "Erhan. Fitvid: Overfitting in pixel-level video prediction, 2021. URL https://arxiv.org/abs/2106.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
80,
258,
113,
273
],
"spans": [
{
"bbox": [
80,
258,
113,
273
],
"score": 1.0,
"content": "13195.",
"type": "text"
}
],
"index": 12,
"is_list_end_line": true
},
{
"bbox": [
69,
279,
542,
295
],
"spans": [
{
"bbox": [
69,
279,
542,
295
],
"score": 1.0,
"content": "Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of",
"type": "text"
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
79,
292,
487,
307
],
"spans": [
{
"bbox": [
79,
292,
487,
307
],
"score": 1.0,
"content": "stochastic parrots: Can language models be too big? . In Proceedings of FAccT 2021, 2021.",
"type": "text"
}
],
"index": 14,
"is_list_end_line": true
},
{
"bbox": [
68,
310,
541,
329
],
"spans": [
{
"bbox": [
68,
310,
541,
329
],
"score": 1.0,
"content": "Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornog-",
"type": "text"
}
],
"index": 15,
"is_list_start_line": true
},
{
"bbox": [
78,
325,
354,
339
],
"spans": [
{
"bbox": [
78,
325,
354,
339
],
"score": 1.0,
"content": "raphy, and malignant stereotypes. In arXiv:2110.01963, 2021.",
"type": "text"
}
],
"index": 16,
"is_list_end_line": true
},
{
"bbox": [
69,
345,
541,
361
],
"spans": [
{
"bbox": [
69,
345,
541,
361
],
"score": 1.0,
"content": "Shikha Bordia and Samuel R. Bowman. Identifying and Reducing Gender Bias in Word-Level Language",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
79,
358,
199,
372
],
"spans": [
{
"bbox": [
79,
358,
199,
372
],
"score": 1.0,
"content": "Models. In NAACL, 2017.",
"type": "text"
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
69,
378,
541,
393
],
"spans": [
{
"bbox": [
69,
378,
541,
393
],
"score": 1.0,
"content": "Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, and William Chan. WaveGrad:",
"type": "text"
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
78,
390,
370,
406
],
"spans": [
{
"bbox": [
78,
390,
370,
406
],
"score": 1.0,
"content": "Estimating Gradients for Waveform Generation. In ICLR, 2021a.",
"type": "text"
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
68,
410,
541,
428
],
"spans": [
{
"bbox": [
68,
410,
541,
428
],
"score": 1.0,
"content": "Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, Najim Dehak, and William Chan.",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true
},
{
"bbox": [
79,
424,
492,
438
],
"spans": [
{
"bbox": [
79,
424,
492,
438
],
"score": 1.0,
"content": "WaveGrad 2: Iterative Refinement for Text-to-Speech Synthesis . In INTERSPEECH, 2021b.",
"type": "text"
}
],
"index": 22,
"is_list_end_line": true
},
{
"bbox": [
69,
443,
520,
459
],
"spans": [
{
"bbox": [
69,
443,
520,
459
],
"score": 1.0,
"content": "Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2022.",
"type": "text"
}
],
"index": 23,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
69,
465,
542,
481
],
"spans": [
{
"bbox": [
69,
465,
542,
481
],
"score": 1.0,
"content": "Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true
},
{
"bbox": [
79,
478,
541,
493
],
"spans": [
{
"bbox": [
79,
478,
541,
493
],
"score": 1.0,
"content": "Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers, 2021.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
79,
489,
279,
505
],
"spans": [
{
"bbox": [
79,
489,
279,
505
],
"score": 1.0,
"content": "URL https://arxiv.org/abs/2105.13290.",
"type": "text"
}
],
"index": 26,
"is_list_end_line": true
},
{
"bbox": [
68,
509,
542,
526
],
"spans": [
{
"bbox": [
68,
509,
542,
526
],
"score": 1.0,
"content": "Chelsea Finn, Ian J. Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through",
"type": "text"
}
],
"index": 27,
"is_list_start_line": true
},
{
"bbox": [
79,
523,
289,
537
],
"spans": [
{
"bbox": [
79,
523,
289,
537
],
"score": 1.0,
"content": "video prediction. ArXiv, abs/1605.07157, 2016.",
"type": "text"
}
],
"index": 28,
"is_list_end_line": true
},
{
"bbox": [
70,
545,
540,
558
],
"spans": [
{
"bbox": [
70,
545,
540,
558
],
"score": 1.0,
"content": "Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Mart’in-Mart’in, and Li Fei-Fei. Maskvit:",
"type": "text"
}
],
"index": 29,
"is_list_start_line": true
},
{
"bbox": [
78,
555,
423,
570
],
"spans": [
{
"bbox": [
78,
555,
423,
570
],
"score": 1.0,
"content": "Masked visual pre-training for video prediction. ArXiv, abs/2206.11894, 2022.",
"type": "text"
}
],
"index": 30,
"is_list_end_line": true
},
{
"bbox": [
68,
575,
542,
592
],
"spans": [
{
"bbox": [
68,
575,
542,
592
],
"score": 1.0,
"content": "William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion",
"type": "text"
}
],
"index": 31,
"is_list_start_line": true
},
{
"bbox": [
78,
588,
322,
604
],
"spans": [
{
"bbox": [
78,
588,
322,
604
],
"score": 1.0,
"content": "modeling of long videos. ArXiv, abs/2205.11495, 2022.",
"type": "text"
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
68,
609,
541,
624
],
"spans": [
{
"bbox": [
68,
609,
541,
624
],
"score": 1.0,
"content": "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
79,
622,
430,
636
],
"spans": [
{
"bbox": [
79,
622,
430,
636
],
"score": 1.0,
"content": "evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.",
"type": "text"
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
69,
641,
541,
657
],
"spans": [
{
"bbox": [
69,
641,
541,
657
],
"score": 1.0,
"content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans",
"type": "text"
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
78,
654,
542,
670
],
"spans": [
{
"bbox": [
78,
654,
542,
670
],
"score": 1.0,
"content": "trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
79,
667,
189,
679
],
"spans": [
{
"bbox": [
79,
667,
189,
679
],
"score": 1.0,
"content": "arXiv:1706.08500, 2017.",
"type": "text"
}
],
"index": 37,
"is_list_end_line": true
},
{
"bbox": [
68,
686,
542,
703
],
"spans": [
{
"bbox": [
68,
686,
542,
703
],
"score": 1.0,
"content": "Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep",
"type": "text"
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
80,
699,
325,
714
],
"spans": [
{
"bbox": [
80,
699,
325,
714
],
"score": 1.0,
"content": "Generative Models and Downstream Applications, 2021.",
"type": "text"
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
69,
720,
522,
735
],
"spans": [
{
"bbox": [
69,
720,
522,
735
],
"score": 1.0,
"content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.",
"type": "text"
}
],
"index": 41,
"is_list_start_line": true,
"is_list_end_line": true
}
],
"index": 24,
"bbox_fs": [
68,
200,
542,
735
]
},
{
"type": "text",
"bbox": [
68,
720,
522,
732
],
"lines": [
{
"bbox": [
68,
720,
522,
732
],
"spans": [],
"index": 40
}
],
"index": 40,
"bbox_fs": [
68,
720,
522,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
67,
73,
542,
735
],
"lines": [
{
"bbox": [
69,
82,
541,
96
],
"spans": [
{
"bbox": [
69,
82,
541,
96
],
"score": 1.0,
"content": "Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
79,
95,
417,
109
],
"spans": [
{
"bbox": [
79,
95,
417,
109
],
"score": 1.0,
"content": "Cascaded diffusion models for high fidelity image generation. JMLR, 2022a.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
68,
113,
542,
129
],
"spans": [
{
"bbox": [
68,
113,
542,
129
],
"score": 1.0,
"content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
81,
128,
315,
140
],
"spans": [
{
"bbox": [
81,
128,
315,
140
],
"score": 1.0,
"content": "Video Diffusion Models. In arXiv:2204.03458, 2022b.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
68,
146,
542,
162
],
"spans": [
{
"bbox": [
68,
146,
542,
162
],
"score": 1.0,
"content": "Nal Kalchbrenner, Aäron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
79,
160,
403,
174
],
"spans": [
{
"bbox": [
79,
160,
403,
174
],
"score": 1.0,
"content": "Koray Kavukcuoglu. Video pixel networks. ArXiv, abs/1610.00527, 2017.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
69,
178,
542,
195
],
"spans": [
{
"bbox": [
69,
178,
542,
195
],
"score": 1.0,
"content": "Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
78,
192,
339,
207
],
"spans": [
{
"bbox": [
78,
192,
339,
207
],
"score": 1.0,
"content": "generative models. arXiv preprint arXiv:2206.00364, 2022.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
69,
213,
542,
226
],
"spans": [
{
"bbox": [
69,
213,
542,
226
],
"score": 1.0,
"content": "Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
78,
225,
227,
239
],
"spans": [
{
"bbox": [
78,
225,
227,
239
],
"score": 1.0,
"content": "preprint arXiv:2107.00630, 2021.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
70,
245,
542,
259
],
"spans": [
{
"bbox": [
70,
245,
542,
259
],
"score": 1.0,
"content": "Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
80,
258,
272,
270
],
"spans": [
{
"bbox": [
80,
258,
272,
270
],
"score": 1.0,
"content": "Model for Audio Synthesis. In ICLR, 2021.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
69,
278,
541,
292
],
"spans": [
{
"bbox": [
69,
278,
541,
292
],
"score": 1.0,
"content": "Manoj Kumar, Mohammad Babaeizadeh, D. Erhan, Chelsea Finn, Sergey Levine, Laurent Dinh, and Durk",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
79,
289,
543,
305
],
"spans": [
{
"bbox": [
79,
289,
543,
305
],
"score": 1.0,
"content": "Kingma. Videoflow: A conditional flow-based model for stochastic video generation. arXiv: Computer",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
81,
302,
249,
316
],
"spans": [
{
"bbox": [
81,
302,
249,
316
],
"score": 1.0,
"content": "Vision and Pattern Recognition, 2020.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
69,
321,
542,
338
],
"spans": [
{
"bbox": [
69,
321,
542,
338
],
"score": 1.0,
"content": "Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
78,
334,
538,
350
],
"spans": [
{
"bbox": [
78,
334,
538,
350
],
"score": 1.0,
"content": "captions. In Proceedings of the IEEE international conference on computer vision, pp. 1426–1434, 2017.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
354,
542,
370
],
"spans": [
{
"bbox": [
69,
354,
542,
370
],
"score": 1.0,
"content": "Michaël Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
79,
367,
241,
380
],
"spans": [
{
"bbox": [
79,
367,
241,
380
],
"score": 1.0,
"content": "error. CoRR, abs/1511.05440, 2016.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
387,
542,
403
],
"spans": [
{
"bbox": [
69,
387,
542,
403
],
"score": 1.0,
"content": "Chenlin Meng, Ruiqi Gao, Diederik P Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
79,
400,
425,
414
],
"spans": [
{
"bbox": [
79,
400,
425,
414
],
"score": 1.0,
"content": "distillation of guided diffusion models. arXiv preprint arXiv:2210.03142, 2022.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
70,
420,
541,
434
],
"spans": [
{
"bbox": [
70,
420,
541,
434
],
"score": 1.0,
"content": "Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
80,
433,
543,
447
],
"spans": [
{
"bbox": [
80,
433,
543,
447
],
"score": 1.0,
"content": "using deep recurrent attentive architectures. In Proceedings of the 25th ACM international conference on",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
79,
444,
229,
459
],
"spans": [
{
"bbox": [
79,
444,
229,
459
],
"score": 1.0,
"content": "Multimedia, pp. 1096–1104, 2017.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
69,
463,
543,
479
],
"spans": [
{
"bbox": [
69,
463,
543,
479
],
"score": 1.0,
"content": "Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
79,
476,
362,
492
],
"spans": [
{
"bbox": [
79,
476,
362,
492
],
"score": 1.0,
"content": "Conference on Machine Learning, pp. 8162–8171. PMLR, 2021.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
69,
496,
541,
511
],
"spans": [
{
"bbox": [
69,
496,
541,
511
],
"score": 1.0,
"content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Bob McGrew Pamela Mishkin, Ilya",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
79,
509,
541,
523
],
"spans": [
{
"bbox": [
79,
509,
541,
523
],
"score": 1.0,
"content": "Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
79,
522,
316,
536
],
"spans": [
{
"bbox": [
79,
522,
316,
536
],
"score": 1.0,
"content": "Guided Diffusion Models. In arXiv:2112.10741, 2021.",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
69,
542,
541,
556
],
"spans": [
{
"bbox": [
69,
542,
541,
556
],
"score": 1.0,
"content": "Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for com-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
79,
554,
541,
568
],
"spans": [
{
"bbox": [
79,
554,
541,
568
],
"score": 1.0,
"content": "positional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Sys-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
79,
565,
542,
581
],
"spans": [
{
"bbox": [
79,
565,
542,
581
],
"score": 1.0,
"content": "tems Datasets and Benchmarks Track (Round 1), 2021. URL https://openreview.net/forum?id=",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
80,
579,
143,
591
],
"spans": [
{
"bbox": [
80,
579,
143,
591
],
"score": 1.0,
"content": "bKBhQhPeKaF.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
69,
596,
542,
613
],
"spans": [
{
"bbox": [
69,
596,
542,
613
],
"score": 1.0,
"content": "Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
79,
610,
182,
624
],
"spans": [
{
"bbox": [
79,
610,
182,
624
],
"score": 1.0,
"content": "Diffusion. arXiv, 2022.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
69,
629,
542,
647
],
"spans": [
{
"bbox": [
69,
629,
542,
647
],
"score": 1.0,
"content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
80,
643,
540,
657
],
"spans": [
{
"bbox": [
80,
643,
540,
657
],
"score": 1.0,
"content": "Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Trans-",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
79,
655,
214,
669
],
"spans": [
{
"bbox": [
79,
655,
214,
669
],
"score": 1.0,
"content": "former. JMLR, 21(140), 2020.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
70,
676,
541,
690
],
"spans": [
{
"bbox": [
70,
676,
541,
690
],
"score": 1.0,
"content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
79,
688,
319,
702
],
"spans": [
{
"bbox": [
79,
688,
319,
702
],
"score": 1.0,
"content": "Image Generation with CLIP Latents. In arXiv, 2022.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
69,
707,
540,
723
],
"spans": [
{
"bbox": [
69,
707,
540,
723
],
"score": 1.0,
"content": "Marc’Aurelio Ranzato, Arthur D. Szlam, Joan Bruna, Michaël Mathieu, Ronan Collobert, and Sumit Chopra.",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
79,
720,
542,
736
],
"spans": [
{
"bbox": [
79,
720,
542,
736
],
"score": 1.0,
"content": "Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014.",
"type": "text"
}
],
"index": 41
}
],
"index": 20.5
}
],
"page_idx": 16,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "list",
"bbox": [
67,
73,
542,
735
],
"lines": [
{
"bbox": [
69,
82,
541,
96
],
"spans": [
{
"bbox": [
69,
82,
541,
96
],
"score": 1.0,
"content": "Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans.",
"type": "text"
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
79,
95,
417,
109
],
"spans": [
{
"bbox": [
79,
95,
417,
109
],
"score": 1.0,
"content": "Cascaded diffusion models for high fidelity image generation. JMLR, 2022a.",
"type": "text"
}
],
"index": 1,
"is_list_end_line": true
},
{
"bbox": [
68,
113,
542,
129
],
"spans": [
{
"bbox": [
68,
113,
542,
129
],
"score": 1.0,
"content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet.",
"type": "text"
}
],
"index": 2,
"is_list_start_line": true
},
{
"bbox": [
81,
128,
315,
140
],
"spans": [
{
"bbox": [
81,
128,
315,
140
],
"score": 1.0,
"content": "Video Diffusion Models. In arXiv:2204.03458, 2022b.",
"type": "text"
}
],
"index": 3,
"is_list_end_line": true
},
{
"bbox": [
68,
146,
542,
162
],
"spans": [
{
"bbox": [
68,
146,
542,
162
],
"score": 1.0,
"content": "Nal Kalchbrenner, Aäron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and",
"type": "text"
}
],
"index": 4,
"is_list_start_line": true
},
{
"bbox": [
79,
160,
403,
174
],
"spans": [
{
"bbox": [
79,
160,
403,
174
],
"score": 1.0,
"content": "Koray Kavukcuoglu. Video pixel networks. ArXiv, abs/1610.00527, 2017.",
"type": "text"
}
],
"index": 5,
"is_list_end_line": true
},
{
"bbox": [
69,
178,
542,
195
],
"spans": [
{
"bbox": [
69,
178,
542,
195
],
"score": 1.0,
"content": "Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based",
"type": "text"
}
],
"index": 6,
"is_list_start_line": true
},
{
"bbox": [
78,
192,
339,
207
],
"spans": [
{
"bbox": [
78,
192,
339,
207
],
"score": 1.0,
"content": "generative models. arXiv preprint arXiv:2206.00364, 2022.",
"type": "text"
}
],
"index": 7,
"is_list_end_line": true
},
{
"bbox": [
69,
213,
542,
226
],
"spans": [
{
"bbox": [
69,
213,
542,
226
],
"score": 1.0,
"content": "Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv",
"type": "text"
}
],
"index": 8,
"is_list_start_line": true
},
{
"bbox": [
78,
225,
227,
239
],
"spans": [
{
"bbox": [
78,
225,
227,
239
],
"score": 1.0,
"content": "preprint arXiv:2107.00630, 2021.",
"type": "text"
}
],
"index": 9,
"is_list_end_line": true
},
{
"bbox": [
70,
245,
542,
259
],
"spans": [
{
"bbox": [
70,
245,
542,
259
],
"score": 1.0,
"content": "Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion",
"type": "text"
}
],
"index": 10,
"is_list_start_line": true
},
{
"bbox": [
80,
258,
272,
270
],
"spans": [
{
"bbox": [
80,
258,
272,
270
],
"score": 1.0,
"content": "Model for Audio Synthesis. In ICLR, 2021.",
"type": "text"
}
],
"index": 11,
"is_list_end_line": true
},
{
"bbox": [
69,
278,
541,
292
],
"spans": [
{
"bbox": [
69,
278,
541,
292
],
"score": 1.0,
"content": "Manoj Kumar, Mohammad Babaeizadeh, D. Erhan, Chelsea Finn, Sergey Levine, Laurent Dinh, and Durk",
"type": "text"
}
],
"index": 12,
"is_list_start_line": true
},
{
"bbox": [
79,
289,
543,
305
],
"spans": [
{
"bbox": [
79,
289,
543,
305
],
"score": 1.0,
"content": "Kingma. Videoflow: A conditional flow-based model for stochastic video generation. arXiv: Computer",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
81,
302,
249,
316
],
"spans": [
{
"bbox": [
81,
302,
249,
316
],
"score": 1.0,
"content": "Vision and Pattern Recognition, 2020.",
"type": "text"
}
],
"index": 14,
"is_list_end_line": true
},
{
"bbox": [
69,
321,
542,
338
],
"spans": [
{
"bbox": [
69,
321,
542,
338
],
"score": 1.0,
"content": "Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using",
"type": "text"
}
],
"index": 15,
"is_list_start_line": true
},
{
"bbox": [
78,
334,
538,
350
],
"spans": [
{
"bbox": [
78,
334,
538,
350
],
"score": 1.0,
"content": "captions. In Proceedings of the IEEE international conference on computer vision, pp. 1426–1434, 2017.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
69,
354,
542,
370
],
"spans": [
{
"bbox": [
69,
354,
542,
370
],
"score": 1.0,
"content": "Michaël Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
79,
367,
241,
380
],
"spans": [
{
"bbox": [
79,
367,
241,
380
],
"score": 1.0,
"content": "error. CoRR, abs/1511.05440, 2016.",
"type": "text"
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
69,
387,
542,
403
],
"spans": [
{
"bbox": [
69,
387,
542,
403
],
"score": 1.0,
"content": "Chenlin Meng, Ruiqi Gao, Diederik P Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On",
"type": "text"
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
79,
400,
425,
414
],
"spans": [
{
"bbox": [
79,
400,
425,
414
],
"score": 1.0,
"content": "distillation of guided diffusion models. arXiv preprint arXiv:2210.03142, 2022.",
"type": "text"
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
70,
420,
541,
434
],
"spans": [
{
"bbox": [
70,
420,
541,
434
],
"score": 1.0,
"content": "Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true
},
{
"bbox": [
80,
433,
543,
447
],
"spans": [
{
"bbox": [
80,
433,
543,
447
],
"score": 1.0,
"content": "using deep recurrent attentive architectures. In Proceedings of the 25th ACM international conference on",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
79,
444,
229,
459
],
"spans": [
{
"bbox": [
79,
444,
229,
459
],
"score": 1.0,
"content": "Multimedia, pp. 1096–1104, 2017.",
"type": "text"
}
],
"index": 23,
"is_list_end_line": true
},
{
"bbox": [
69,
463,
543,
479
],
"spans": [
{
"bbox": [
69,
463,
543,
479
],
"score": 1.0,
"content": "Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true
},
{
"bbox": [
79,
476,
362,
492
],
"spans": [
{
"bbox": [
79,
476,
362,
492
],
"score": 1.0,
"content": "Conference on Machine Learning, pp. 8162–8171. PMLR, 2021.",
"type": "text"
}
],
"index": 25,
"is_list_end_line": true
},
{
"bbox": [
69,
496,
541,
511
],
"spans": [
{
"bbox": [
69,
496,
541,
511
],
"score": 1.0,
"content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Bob McGrew Pamela Mishkin, Ilya",
"type": "text"
}
],
"index": 26,
"is_list_start_line": true
},
{
"bbox": [
79,
509,
541,
523
],
"spans": [
{
"bbox": [
79,
509,
541,
523
],
"score": 1.0,
"content": "Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
79,
522,
316,
536
],
"spans": [
{
"bbox": [
79,
522,
316,
536
],
"score": 1.0,
"content": "Guided Diffusion Models. In arXiv:2112.10741, 2021.",
"type": "text"
}
],
"index": 28,
"is_list_end_line": true
},
{
"bbox": [
69,
542,
541,
556
],
"spans": [
{
"bbox": [
69,
542,
541,
556
],
"score": 1.0,
"content": "Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for com-",
"type": "text"
}
],
"index": 29,
"is_list_start_line": true
},
{
"bbox": [
79,
554,
541,
568
],
"spans": [
{
"bbox": [
79,
554,
541,
568
],
"score": 1.0,
"content": "positional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Sys-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
79,
565,
542,
581
],
"spans": [
{
"bbox": [
79,
565,
542,
581
],
"score": 1.0,
"content": "tems Datasets and Benchmarks Track (Round 1), 2021. URL https://openreview.net/forum?id=",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
80,
579,
143,
591
],
"spans": [
{
"bbox": [
80,
579,
143,
591
],
"score": 1.0,
"content": "bKBhQhPeKaF.",
"type": "text"
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
69,
596,
542,
613
],
"spans": [
{
"bbox": [
69,
596,
542,
613
],
"score": 1.0,
"content": "Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
79,
610,
182,
624
],
"spans": [
{
"bbox": [
79,
610,
182,
624
],
"score": 1.0,
"content": "Diffusion. arXiv, 2022.",
"type": "text"
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
69,
629,
542,
647
],
"spans": [
{
"bbox": [
69,
629,
542,
647
],
"score": 1.0,
"content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,",
"type": "text"
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
80,
643,
540,
657
],
"spans": [
{
"bbox": [
80,
643,
540,
657
],
"score": 1.0,
"content": "Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Trans-",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
79,
655,
214,
669
],
"spans": [
{
"bbox": [
79,
655,
214,
669
],
"score": 1.0,
"content": "former. JMLR, 21(140), 2020.",
"type": "text"
}
],
"index": 37,
"is_list_end_line": true
},
{
"bbox": [
70,
676,
541,
690
],
"spans": [
{
"bbox": [
70,
676,
541,
690
],
"score": 1.0,
"content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional",
"type": "text"
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
79,
688,
319,
702
],
"spans": [
{
"bbox": [
79,
688,
319,
702
],
"score": 1.0,
"content": "Image Generation with CLIP Latents. In arXiv, 2022.",
"type": "text"
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
69,
707,
540,
723
],
"spans": [
{
"bbox": [
69,
707,
540,
723
],
"score": 1.0,
"content": "Marc’Aurelio Ranzato, Arthur D. Szlam, Joan Bruna, Michaël Mathieu, Ronan Collobert, and Sumit Chopra.",
"type": "text"
}
],
"index": 40,
"is_list_start_line": true
},
{
"bbox": [
79,
720,
542,
736
],
"spans": [
{
"bbox": [
79,
720,
542,
736
],
"score": 1.0,
"content": "Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
69,
81,
541,
96
],
"spans": [
{
"bbox": [
69,
81,
541,
96
],
"score": 1.0,
"content": "Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
79,
94,
361,
108
],
"spans": [
{
"bbox": [
79,
94,
361,
108
],
"score": 1.0,
"content": "Image Synthesis with Latent Diffusion Models. In CVPR, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 1,
"is_list_end_line": true
},
{
"bbox": [
68,
111,
543,
131
],
"spans": [
{
"bbox": [
68,
111,
543,
131
],
"score": 1.0,
"content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_start_line": true
},
{
"bbox": [
79,
127,
541,
141
],
"spans": [
{
"bbox": [
79,
127,
541,
141
],
"score": 1.0,
"content": "segmentation. In International Conference on Medical image computing and computer-assisted interven-",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
79,
139,
230,
153
],
"spans": [
{
"bbox": [
79,
139,
230,
153
],
"score": 1.0,
"content": "tion, pp. 234–241. Springer, 2015.",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
68,
156,
541,
174
],
"spans": [
{
"bbox": [
68,
156,
541,
174
],
"score": 1.0,
"content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet,",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
79,
171,
489,
185
],
"spans": [
{
"bbox": [
79,
171,
489,
185
],
"score": 1.0,
"content": "and Mohammad Norouzi. Palette: Image-to-Image Diffusion Models. In SIGGRAPH, 2022a.",
"type": "text",
"cross_page": true
}
],
"index": 6,
"is_list_end_line": true
},
{
"bbox": [
69,
191,
541,
205
],
"spans": [
{
"bbox": [
69,
191,
541,
205
],
"score": 1.0,
"content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed",
"type": "text",
"cross_page": true
}
],
"index": 7,
"is_list_start_line": true
},
{
"bbox": [
79,
204,
541,
218
],
"spans": [
{
"bbox": [
79,
204,
541,
218
],
"score": 1.0,
"content": "Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
77,
213,
543,
232
],
"spans": [
{
"bbox": [
77,
213,
543,
232
],
"score": 1.0,
"content": "Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
79,
228,
285,
242
],
"spans": [
{
"bbox": [
79,
228,
285,
242
],
"score": 1.0,
"content": "Language Understanding. In NeurIPS, 2022b.",
"type": "text",
"cross_page": true
}
],
"index": 10,
"is_list_end_line": true
},
{
"bbox": [
68,
246,
541,
263
],
"spans": [
{
"bbox": [
68,
246,
541,
263
],
"score": 1.0,
"content": "Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi.",
"type": "text",
"cross_page": true
}
],
"index": 11,
"is_list_start_line": true
},
{
"bbox": [
79,
260,
542,
273
],
"spans": [
{
"bbox": [
79,
260,
542,
273
],
"score": 1.0,
"content": "Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine",
"type": "text",
"cross_page": true
}
],
"index": 12
},
{
"bbox": [
80,
273,
166,
285
],
"spans": [
{
"bbox": [
80,
273,
166,
285
],
"score": 1.0,
"content": "Intelligence, 2022c.",
"type": "text",
"cross_page": true
}
],
"index": 13,
"is_list_end_line": true
},
{
"bbox": [
69,
291,
541,
306
],
"spans": [
{
"bbox": [
69,
291,
541,
306
],
"score": 1.0,
"content": "Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In ICLR,",
"type": "text",
"cross_page": true
}
],
"index": 14,
"is_list_start_line": true
},
{
"bbox": [
78,
303,
107,
318
],
"spans": [
{
"bbox": [
78,
303,
107,
318
],
"score": 1.0,
"content": "2022.",
"type": "text",
"cross_page": true
}
],
"index": 15,
"is_list_end_line": true
},
{
"bbox": [
70,
325,
541,
338
],
"spans": [
{
"bbox": [
70,
325,
423,
338
],
"score": 1.0,
"content": "Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. PixelCNN",
"type": "text",
"cross_page": true
},
{
"bbox": [
424,
328,
439,
335
],
"score": 0.42,
"content": "^ { + + }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
439,
325,
541,
338
],
"score": 1.0,
"content": ": Improving the Pixel-",
"type": "text",
"cross_page": true
}
],
"index": 16,
"is_list_start_line": true
},
{
"bbox": [
79,
337,
542,
351
],
"spans": [
{
"bbox": [
79,
337,
542,
351
],
"score": 1.0,
"content": "CNN with discretized logistic mixture likelihood and other modifications. In International Conference on",
"type": "text",
"cross_page": true
}
],
"index": 17
},
{
"bbox": [
80,
348,
222,
362
],
"spans": [
{
"bbox": [
80,
348,
222,
362
],
"score": 1.0,
"content": "Learning Representations, 2017.",
"type": "text",
"cross_page": true
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
69,
369,
541,
383
],
"spans": [
{
"bbox": [
69,
369,
541,
383
],
"score": 1.0,
"content": "Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush",
"type": "text",
"cross_page": true
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
80,
382,
541,
393
],
"spans": [
{
"bbox": [
80,
382,
541,
393
],
"score": 1.0,
"content": "Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered",
"type": "text",
"cross_page": true
}
],
"index": 20
},
{
"bbox": [
79,
393,
384,
407
],
"spans": [
{
"bbox": [
79,
393,
384,
407
],
"score": 1.0,
"content": "400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 21,
"is_list_end_line": true
},
{
"bbox": [
69,
413,
540,
427
],
"spans": [
{
"bbox": [
69,
413,
540,
427
],
"score": 1.0,
"content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolu-",
"type": "text",
"cross_page": true
}
],
"index": 22,
"is_list_start_line": true
},
{
"bbox": [
79,
425,
500,
439
],
"spans": [
{
"bbox": [
79,
425,
500,
439
],
"score": 1.0,
"content": "tional lstm network: A machine learning approach for precipitation nowcasting. In NIPS, 2015.",
"type": "text",
"cross_page": true
}
],
"index": 23,
"is_list_end_line": true
},
{
"bbox": [
68,
443,
542,
461
],
"spans": [
{
"bbox": [
68,
443,
542,
461
],
"score": 1.0,
"content": "Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang,",
"type": "text",
"cross_page": true
}
],
"index": 24,
"is_list_start_line": true
},
{
"bbox": [
79,
457,
541,
471
],
"spans": [
{
"bbox": [
79,
457,
541,
471
],
"score": 1.0,
"content": "Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video",
"type": "text",
"cross_page": true
}
],
"index": 25
},
{
"bbox": [
79,
470,
369,
484
],
"spans": [
{
"bbox": [
79,
470,
369,
484
],
"score": 1.0,
"content": "Generation without Text-Video Data. In arXiv:2209.14792, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 26,
"is_list_end_line": true
},
{
"bbox": [
68,
488,
542,
505
],
"spans": [
{
"bbox": [
68,
488,
542,
505
],
"score": 1.0,
"content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning",
"type": "text",
"cross_page": true
}
],
"index": 27,
"is_list_start_line": true
},
{
"bbox": [
79,
502,
540,
515
],
"spans": [
{
"bbox": [
79,
502,
540,
515
],
"score": 1.0,
"content": "using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp. 2256–2265.",
"type": "text",
"cross_page": true
}
],
"index": 28
},
{
"bbox": [
79,
514,
141,
528
],
"spans": [
{
"bbox": [
79,
514,
141,
528
],
"score": 1.0,
"content": "PMLR, 2015.",
"type": "text",
"cross_page": true
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
68,
533,
543,
550
],
"spans": [
{
"bbox": [
68,
533,
543,
550
],
"score": 1.0,
"content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint",
"type": "text",
"cross_page": true
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
80,
547,
189,
559
],
"spans": [
{
"bbox": [
80,
547,
189,
559
],
"score": 1.0,
"content": "arXiv:2010.02502, 2020.",
"type": "text",
"cross_page": true
}
],
"index": 31,
"is_list_end_line": true
},
{
"bbox": [
68,
565,
541,
582
],
"spans": [
{
"bbox": [
68,
565,
541,
582
],
"score": 1.0,
"content": "Yang Song and Stefano Ermon. Generative Modeling by Estimating Gradients of the Data Distribution.",
"type": "text",
"cross_page": true
}
],
"index": 32,
"is_list_start_line": true
},
{
"bbox": [
79,
579,
150,
592
],
"spans": [
{
"bbox": [
79,
579,
150,
592
],
"score": 1.0,
"content": "NeurIPS, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 33,
"is_list_end_line": true
},
{
"bbox": [
69,
598,
540,
613
],
"spans": [
{
"bbox": [
69,
598,
540,
613
],
"score": 1.0,
"content": "Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole.",
"type": "text",
"cross_page": true
}
],
"index": 34,
"is_list_start_line": true
},
{
"bbox": [
79,
612,
477,
625
],
"spans": [
{
"bbox": [
79,
612,
477,
625
],
"score": 1.0,
"content": "Score-based generative modeling through stochastic differential equations. In ICLR, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 35,
"is_list_end_line": true
},
{
"bbox": [
69,
632,
541,
645
],
"spans": [
{
"bbox": [
69,
632,
541,
645
],
"score": 1.0,
"content": "Belinda Tzen and Maxim Raginsky. Neural Stochastic Differential Equations: Deep Latent Gaussian Models",
"type": "text",
"cross_page": true
}
],
"index": 36,
"is_list_start_line": true
},
{
"bbox": [
79,
644,
304,
658
],
"spans": [
{
"bbox": [
79,
644,
304,
658
],
"score": 1.0,
"content": "in the Diffusion Limit. In arXiv:1905.09883, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 37,
"is_list_end_line": true
},
{
"bbox": [
70,
664,
541,
677
],
"spans": [
{
"bbox": [
70,
664,
541,
677
],
"score": 1.0,
"content": "Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain",
"type": "text",
"cross_page": true
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
79,
676,
542,
690
],
"spans": [
{
"bbox": [
79,
676,
542,
690
],
"score": 1.0,
"content": "Gelly. FVD: A new Metric for Video Generation. In ICLR 2022 Workshop: Deep Generative Models for",
"type": "text",
"cross_page": true
}
],
"index": 39
},
{
"bbox": [
80,
687,
211,
702
],
"spans": [
{
"bbox": [
80,
687,
211,
702
],
"score": 1.0,
"content": "Highly Structured Data, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 40,
"is_list_end_line": true
},
{
"bbox": [
68,
707,
542,
723
],
"spans": [
{
"bbox": [
68,
707,
542,
723
],
"score": 1.0,
"content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. ArXiv,",
"type": "text",
"cross_page": true
}
],
"index": 41,
"is_list_start_line": true
},
{
"bbox": [
79,
720,
180,
734
],
"spans": [
{
"bbox": [
79,
720,
180,
734
],
"score": 1.0,
"content": "abs/1609.02612, 2016.",
"type": "text",
"cross_page": true
}
],
"index": 42,
"is_list_end_line": true
},
{
"bbox": [
70,
82,
541,
95
],
"spans": [
{
"bbox": [
70,
82,
541,
95
],
"score": 1.0,
"content": "Daniel Watson, Ricardo Chan, William Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
80,
95,
376,
106
],
"spans": [
{
"bbox": [
80,
95,
376,
106
],
"score": 1.0,
"content": "Norouzi. Novel View Synthesis with Diffusion Models. arXiv, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 1,
"is_list_end_line": true
},
{
"bbox": [
70,
114,
541,
127
],
"spans": [
{
"bbox": [
70,
114,
541,
127
],
"score": 1.0,
"content": "Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G. Dimakis, and Peyman",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_start_line": true
},
{
"bbox": [
79,
126,
367,
139
],
"spans": [
{
"bbox": [
79,
126,
367,
139
],
"score": 1.0,
"content": "Milanfar. Deblurring via Stochastic Refinement. In CVPR, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 3,
"is_list_end_line": true
},
{
"bbox": [
68,
144,
540,
160
],
"spans": [
{
"bbox": [
68,
144,
540,
160
],
"score": 1.0,
"content": "Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Genera-",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_start_line": true
},
{
"bbox": [
79,
158,
224,
170
],
"spans": [
{
"bbox": [
79,
158,
224,
170
],
"score": 1.0,
"content": "tion. In arXiv:2203.09481, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_end_line": true
},
{
"bbox": [
68,
177,
540,
192
],
"spans": [
{
"bbox": [
68,
177,
540,
192
],
"score": 1.0,
"content": "Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexan-",
"type": "text",
"cross_page": true
}
],
"index": 6,
"is_list_start_line": true
},
{
"bbox": [
79,
188,
541,
205
],
"spans": [
{
"bbox": [
79,
188,
541,
205
],
"score": 1.0,
"content": "der Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang,",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
78,
201,
541,
216
],
"spans": [
{
"bbox": [
78,
201,
541,
216
],
"score": 1.0,
"content": "and Yonghui Wu Jason Baldridge. Scaling Autoregressive Models for Content-Rich Text-to-Image Gener-",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
79,
214,
230,
227
],
"spans": [
{
"bbox": [
79,
214,
230,
227
],
"score": 1.0,
"content": "ation. In arXiv:2206.10789, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 9,
"is_list_end_line": true
}
],
"index": 20.5,
"bbox_fs": [
68,
82,
543,
736
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
63,
71,
542,
732
],
"lines": [
{
"bbox": [
69,
81,
541,
96
],
"spans": [
{
"bbox": [
69,
81,
541,
96
],
"score": 1.0,
"content": "Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
79,
94,
361,
108
],
"spans": [
{
"bbox": [
79,
94,
361,
108
],
"score": 1.0,
"content": "Image Synthesis with Latent Diffusion Models. In CVPR, 2022.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
68,
111,
543,
131
],
"spans": [
{
"bbox": [
68,
111,
543,
131
],
"score": 1.0,
"content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
79,
127,
541,
141
],
"spans": [
{
"bbox": [
79,
127,
541,
141
],
"score": 1.0,
"content": "segmentation. In International Conference on Medical image computing and computer-assisted interven-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
79,
139,
230,
153
],
"spans": [
{
"bbox": [
79,
139,
230,
153
],
"score": 1.0,
"content": "tion, pp. 234–241. Springer, 2015.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
68,
156,
541,
174
],
"spans": [
{
"bbox": [
68,
156,
541,
174
],
"score": 1.0,
"content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
79,
171,
489,
185
],
"spans": [
{
"bbox": [
79,
171,
489,
185
],
"score": 1.0,
"content": "and Mohammad Norouzi. Palette: Image-to-Image Diffusion Models. In SIGGRAPH, 2022a.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
69,
191,
541,
205
],
"spans": [
{
"bbox": [
69,
191,
541,
205
],
"score": 1.0,
"content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
79,
204,
541,
218
],
"spans": [
{
"bbox": [
79,
204,
541,
218
],
"score": 1.0,
"content": "Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
77,
213,
543,
232
],
"spans": [
{
"bbox": [
77,
213,
543,
232
],
"score": 1.0,
"content": "Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
79,
228,
285,
242
],
"spans": [
{
"bbox": [
79,
228,
285,
242
],
"score": 1.0,
"content": "Language Understanding. In NeurIPS, 2022b.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
68,
246,
541,
263
],
"spans": [
{
"bbox": [
68,
246,
541,
263
],
"score": 1.0,
"content": "Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
79,
260,
542,
273
],
"spans": [
{
"bbox": [
79,
260,
542,
273
],
"score": 1.0,
"content": "Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
80,
273,
166,
285
],
"spans": [
{
"bbox": [
80,
273,
166,
285
],
"score": 1.0,
"content": "Intelligence, 2022c.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
69,
291,
541,
306
],
"spans": [
{
"bbox": [
69,
291,
541,
306
],
"score": 1.0,
"content": "Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In ICLR,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
78,
303,
107,
318
],
"spans": [
{
"bbox": [
78,
303,
107,
318
],
"score": 1.0,
"content": "2022.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
70,
325,
541,
338
],
"spans": [
{
"bbox": [
70,
325,
423,
338
],
"score": 1.0,
"content": "Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. PixelCNN",
"type": "text"
},
{
"bbox": [
424,
328,
439,
335
],
"score": 0.42,
"content": "^ { + + }",
"type": "inline_equation"
},
{
"bbox": [
439,
325,
541,
338
],
"score": 1.0,
"content": ": Improving the Pixel-",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
79,
337,
542,
351
],
"spans": [
{
"bbox": [
79,
337,
542,
351
],
"score": 1.0,
"content": "CNN with discretized logistic mixture likelihood and other modifications. In International Conference on",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
80,
348,
222,
362
],
"spans": [
{
"bbox": [
80,
348,
222,
362
],
"score": 1.0,
"content": "Learning Representations, 2017.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
69,
369,
541,
383
],
"spans": [
{
"bbox": [
69,
369,
541,
383
],
"score": 1.0,
"content": "Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
80,
382,
541,
393
],
"spans": [
{
"bbox": [
80,
382,
541,
393
],
"score": 1.0,
"content": "Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
79,
393,
384,
407
],
"spans": [
{
"bbox": [
79,
393,
384,
407
],
"score": 1.0,
"content": "400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
69,
413,
540,
427
],
"spans": [
{
"bbox": [
69,
413,
540,
427
],
"score": 1.0,
"content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolu-",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
79,
425,
500,
439
],
"spans": [
{
"bbox": [
79,
425,
500,
439
],
"score": 1.0,
"content": "tional lstm network: A machine learning approach for precipitation nowcasting. In NIPS, 2015.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
68,
443,
542,
461
],
"spans": [
{
"bbox": [
68,
443,
542,
461
],
"score": 1.0,
"content": "Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
79,
457,
541,
471
],
"spans": [
{
"bbox": [
79,
457,
541,
471
],
"score": 1.0,
"content": "Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
79,
470,
369,
484
],
"spans": [
{
"bbox": [
79,
470,
369,
484
],
"score": 1.0,
"content": "Generation without Text-Video Data. In arXiv:2209.14792, 2022.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
68,
488,
542,
505
],
"spans": [
{
"bbox": [
68,
488,
542,
505
],
"score": 1.0,
"content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
79,
502,
540,
515
],
"spans": [
{
"bbox": [
79,
502,
540,
515
],
"score": 1.0,
"content": "using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp. 2256–2265.",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
79,
514,
141,
528
],
"spans": [
{
"bbox": [
79,
514,
141,
528
],
"score": 1.0,
"content": "PMLR, 2015.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
68,
533,
543,
550
],
"spans": [
{
"bbox": [
68,
533,
543,
550
],
"score": 1.0,
"content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
80,
547,
189,
559
],
"spans": [
{
"bbox": [
80,
547,
189,
559
],
"score": 1.0,
"content": "arXiv:2010.02502, 2020.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
68,
565,
541,
582
],
"spans": [
{
"bbox": [
68,
565,
541,
582
],
"score": 1.0,
"content": "Yang Song and Stefano Ermon. Generative Modeling by Estimating Gradients of the Data Distribution.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
79,
579,
150,
592
],
"spans": [
{
"bbox": [
79,
579,
150,
592
],
"score": 1.0,
"content": "NeurIPS, 2019.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
69,
598,
540,
613
],
"spans": [
{
"bbox": [
69,
598,
540,
613
],
"score": 1.0,
"content": "Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
79,
612,
477,
625
],
"spans": [
{
"bbox": [
79,
612,
477,
625
],
"score": 1.0,
"content": "Score-based generative modeling through stochastic differential equations. In ICLR, 2021.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
69,
632,
541,
645
],
"spans": [
{
"bbox": [
69,
632,
541,
645
],
"score": 1.0,
"content": "Belinda Tzen and Maxim Raginsky. Neural Stochastic Differential Equations: Deep Latent Gaussian Models",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
79,
644,
304,
658
],
"spans": [
{
"bbox": [
79,
644,
304,
658
],
"score": 1.0,
"content": "in the Diffusion Limit. In arXiv:1905.09883, 2019.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
70,
664,
541,
677
],
"spans": [
{
"bbox": [
70,
664,
541,
677
],
"score": 1.0,
"content": "Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
79,
676,
542,
690
],
"spans": [
{
"bbox": [
79,
676,
542,
690
],
"score": 1.0,
"content": "Gelly. FVD: A new Metric for Video Generation. In ICLR 2022 Workshop: Deep Generative Models for",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
80,
687,
211,
702
],
"spans": [
{
"bbox": [
80,
687,
211,
702
],
"score": 1.0,
"content": "Highly Structured Data, 2019.",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
68,
707,
542,
723
],
"spans": [
{
"bbox": [
68,
707,
542,
723
],
"score": 1.0,
"content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. ArXiv,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
79,
720,
180,
734
],
"spans": [
{
"bbox": [
79,
720,
180,
734
],
"score": 1.0,
"content": "abs/1609.02612, 2016.",
"type": "text"
}
],
"index": 42
}
],
"index": 21
}
],
"page_idx": 17,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
71,
25,
239,
38
],
"spans": [
{
"bbox": [
71,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "list",
"bbox": [
63,
71,
542,
732
],
"lines": [],
"index": 21,
"bbox_fs": [
68,
81,
543,
734
],
"lines_deleted": true
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
68,
79,
542,
225
],
"lines": [
{
"bbox": [
70,
82,
541,
95
],
"spans": [
{
"bbox": [
70,
82,
541,
95
],
"score": 1.0,
"content": "Daniel Watson, Ricardo Chan, William Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
80,
95,
376,
106
],
"spans": [
{
"bbox": [
80,
95,
376,
106
],
"score": 1.0,
"content": "Norouzi. Novel View Synthesis with Diffusion Models. arXiv, 2022.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
70,
114,
541,
127
],
"spans": [
{
"bbox": [
70,
114,
541,
127
],
"score": 1.0,
"content": "Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G. Dimakis, and Peyman",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
79,
126,
367,
139
],
"spans": [
{
"bbox": [
79,
126,
367,
139
],
"score": 1.0,
"content": "Milanfar. Deblurring via Stochastic Refinement. In CVPR, 2022.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
68,
144,
540,
160
],
"spans": [
{
"bbox": [
68,
144,
540,
160
],
"score": 1.0,
"content": "Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Genera-",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
79,
158,
224,
170
],
"spans": [
{
"bbox": [
79,
158,
224,
170
],
"score": 1.0,
"content": "tion. In arXiv:2203.09481, 2022.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
68,
177,
540,
192
],
"spans": [
{
"bbox": [
68,
177,
540,
192
],
"score": 1.0,
"content": "Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexan-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
79,
188,
541,
205
],
"spans": [
{
"bbox": [
79,
188,
541,
205
],
"score": 1.0,
"content": "der Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
78,
201,
541,
216
],
"spans": [
{
"bbox": [
78,
201,
541,
216
],
"score": 1.0,
"content": "and Yonghui Wu Jason Baldridge. Scaling Autoregressive Models for Content-Rich Text-to-Image Gener-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
79,
214,
230,
227
],
"spans": [
{
"bbox": [
79,
214,
230,
227
],
"score": 1.0,
"content": "ation. In arXiv:2206.10789, 2022.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5
}
],
"page_idx": 18,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
72,
27,
237,
37
],
"lines": [
{
"bbox": [
70,
25,
239,
38
],
"spans": [
{
"bbox": [
70,
25,
239,
38
],
"score": 1.0,
"content": "Under review as submission to TMLR",
"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": "19",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
68,
79,
542,
225
],
"lines": [],
"index": 4.5,
"bbox_fs": [
68,
82,
541,
227
],
"lines_deleted": true
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.1.11"
}