{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
122,
97,
489,
137
],
"lines": [
{
"bbox": [
133,
97,
478,
116
],
"spans": [
{
"bbox": [
133,
97,
478,
116
],
"score": 1.0,
"content": "DPM-Solver: A Fast ODE Solver for Diffusion",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
121,
117,
489,
138
],
"spans": [
{
"bbox": [
121,
117,
489,
138
],
"score": 1.0,
"content": "Probabilistic Model Sampling in Around 10 Steps",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
127,
178,
484,
260
],
"lines": [
{
"bbox": [
136,
178,
479,
191
],
"spans": [
{
"bbox": [
136,
178,
166,
191
],
"score": 1.0,
"content": "Cheng",
"type": "text"
},
{
"bbox": [
167,
178,
184,
190
],
"score": 0.47,
"content": "\\mathbf { L } \\mathbf { u } ^ { \\dagger }",
"type": "inline_equation"
},
{
"bbox": [
185,
178,
326,
191
],
"score": 1.0,
"content": ", Yuhao Zhou†, Fan Bao†, Jianfei",
"type": "text"
},
{
"bbox": [
326,
178,
357,
190
],
"score": 0.36,
"content": "\\mathbf { C h e n } ^ { \\dagger * }",
"type": "inline_equation"
},
{
"bbox": [
358,
178,
412,
191
],
"score": 1.0,
"content": ", Chongxuan",
"type": "text"
},
{
"bbox": [
413,
178,
428,
190
],
"score": 0.56,
"content": "\\mathbf { L i } ^ { \\dagger }",
"type": "inline_equation"
},
{
"bbox": [
428,
178,
479,
191
],
"score": 1.0,
"content": ", Jun Zhu†∗",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
156,
189,
455,
205
],
"spans": [
{
"bbox": [
156,
189,
455,
205
],
"score": 1.0,
"content": "†Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, THBI Lab",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
145,
200,
466,
217
],
"spans": [
{
"bbox": [
145,
200,
466,
217
],
"score": 1.0,
"content": "†Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084 China",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
160,
212,
450,
229
],
"spans": [
{
"bbox": [
160,
212,
450,
229
],
"score": 1.0,
"content": "‡Gaoling School of Artificial Intelligence, Renmin University of China,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
124,
225,
486,
240
],
"spans": [
{
"bbox": [
124,
225,
486,
240
],
"score": 1.0,
"content": "‡Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
138,
236,
472,
251
],
"spans": [
{
"bbox": [
138,
236,
472,
251
],
"score": 1.0,
"content": "{lucheng.lc15, yuhaoz.cs}@gmail.com; bf19@mails.tsinghua.edu.cn",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
154,
248,
457,
262
],
"spans": [
{
"bbox": [
154,
248,
457,
262
],
"score": 1.0,
"content": "chongxuanli@ruc.edu.cn; {jianfeic, dcszj}@tsinghua.edu.cn",
"type": "text"
}
],
"index": 8
}
],
"index": 5
},
{
"type": "title",
"bbox": [
283,
288,
328,
301
],
"lines": [
{
"bbox": [
281,
287,
330,
302
],
"spans": [
{
"bbox": [
281,
287,
330,
302
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
143,
313,
469,
510
],
"lines": [
{
"bbox": [
141,
314,
471,
327
],
"spans": [
{
"bbox": [
141,
314,
471,
327
],
"score": 1.0,
"content": "Diffusion probabilistic models (DPMs) are emerging powerful generative models.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
325,
470,
338
],
"spans": [
{
"bbox": [
141,
325,
470,
338
],
"score": 1.0,
"content": "Despite their high-quality generation performance, DPMs still suffer from their",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
141,
336,
469,
349
],
"spans": [
{
"bbox": [
141,
336,
469,
349
],
"score": 1.0,
"content": "slow sampling as they generally need hundreds or thousands of sequential function",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
347,
469,
360
],
"spans": [
{
"bbox": [
141,
347,
469,
360
],
"score": 1.0,
"content": "evaluations (steps) of large neural networks to draw a sample. Sampling from",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
357,
469,
370
],
"spans": [
{
"bbox": [
141,
357,
469,
370
],
"score": 1.0,
"content": "DPMs can be viewed alternatively as solving the corresponding diffusion ordinary",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
369,
470,
381
],
"spans": [
{
"bbox": [
141,
369,
470,
381
],
"score": 1.0,
"content": "differential equations (ODEs). In this work, we propose an exact formulation of",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
380,
469,
392
],
"spans": [
{
"bbox": [
141,
380,
469,
392
],
"score": 1.0,
"content": "the solution of diffusion ODEs. The formulation analytically computes the linear",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
390,
469,
403
],
"spans": [
{
"bbox": [
141,
390,
469,
403
],
"score": 1.0,
"content": "part of the solution, rather than leaving all terms to black-box ODE solvers as",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
401,
470,
414
],
"spans": [
{
"bbox": [
141,
401,
470,
414
],
"score": 1.0,
"content": "adopted in previous works. By applying change-of-variable, the solution can be",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
412,
470,
424
],
"spans": [
{
"bbox": [
141,
412,
470,
424
],
"score": 1.0,
"content": "equivalently simplified to an exponentially weighted integral of the neural network.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
423,
470,
436
],
"spans": [
{
"bbox": [
141,
423,
470,
436
],
"score": 1.0,
"content": "Based on our formulation, we propose DPM-Solver, a fast dedicated high-order",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
141,
433,
470,
447
],
"spans": [
{
"bbox": [
141,
433,
470,
447
],
"score": 1.0,
"content": "solver for diffusion ODEs with the convergence order guarantee. DPM-Solver",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
445,
470,
457
],
"spans": [
{
"bbox": [
141,
445,
470,
457
],
"score": 1.0,
"content": "is suitable for both discrete-time and continuous-time DPMs without any further",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
456,
469,
469
],
"spans": [
{
"bbox": [
141,
456,
469,
469
],
"score": 1.0,
"content": "training. Experimental results show that DPM-Solver can generate high-quality",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
141,
467,
469,
479
],
"spans": [
{
"bbox": [
141,
467,
469,
479
],
"score": 1.0,
"content": "samples in only 10 to 20 function evaluations on various datasets. We achieve",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
142,
478,
469,
489
],
"spans": [
{
"bbox": [
142,
478,
179,
488
],
"score": 0.27,
"content": "4 . 7 0 \\ : \\mathrm { F I D }",
"type": "inline_equation"
},
{
"bbox": [
179,
478,
469,
489
],
"score": 1.0,
"content": "in 10 function evaluations and 2.87 FID in 20 function evaluations on the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
141,
488,
470,
501
],
"spans": [
{
"bbox": [
141,
488,
238,
501
],
"score": 1.0,
"content": "CIFAR10 dataset, and a",
"type": "text"
},
{
"bbox": [
238,
489,
276,
499
],
"score": 0.91,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
276,
488,
470,
501
],
"score": 1.0,
"content": "speedup compared with previous state-of-the-art",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
141,
500,
317,
511
],
"spans": [
{
"bbox": [
141,
500,
317,
511
],
"score": 1.0,
"content": "training-free samplers on various datasets.2",
"type": "text"
}
],
"index": 27
}
],
"index": 18.5
},
{
"type": "title",
"bbox": [
107,
533,
190,
547
],
"lines": [
{
"bbox": [
105,
532,
192,
549
],
"spans": [
{
"bbox": [
105,
532,
192,
549
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
559,
506,
680
],
"lines": [
{
"bbox": [
105,
558,
507,
573
],
"spans": [
{
"bbox": [
105,
558,
507,
573
],
"score": 1.0,
"content": "Diffusion probabilistic models (DPMs) [1–3] are emerging powerful generative models with promis-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
570,
506,
585
],
"spans": [
{
"bbox": [
105,
570,
506,
585
],
"score": 1.0,
"content": "ing performance on many tasks, such as image generation [4, 5], video generation [6], text-to-image",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
581,
507,
595
],
"spans": [
{
"bbox": [
105,
581,
507,
595
],
"score": 1.0,
"content": "generation [7], speech synthesis [8, 9] and lossless compression [10]. DPMs are defined by discrete-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
592,
505,
605
],
"spans": [
{
"bbox": [
106,
592,
505,
605
],
"score": 1.0,
"content": "time random processes [1, 2] or continuous-time stochastic differential equations (SDEs) [3], which",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
602,
506,
617
],
"spans": [
{
"bbox": [
105,
602,
506,
617
],
"score": 1.0,
"content": "learn to gradually remove the noise added to the data points. Compared with the widely-used genera-",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
614,
505,
627
],
"spans": [
{
"bbox": [
105,
614,
505,
627
],
"score": 1.0,
"content": "tive adversarial networks (GANs) [11] and variational auto-encoders (VAEs) [12], DPMs can not only",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
625,
506,
638
],
"spans": [
{
"bbox": [
106,
625,
506,
638
],
"score": 1.0,
"content": "compute exact likelihood [3], but also achieve even better sample quality for image generation [4].",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
636,
505,
648
],
"spans": [
{
"bbox": [
106,
636,
505,
648
],
"score": 1.0,
"content": "However, to obtain high-quality samples, DPMs usually need hundreds or thousands of sequential",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
648,
506,
659
],
"spans": [
{
"bbox": [
105,
648,
506,
659
],
"score": 1.0,
"content": "steps of large neural network evaluations, thereby resulting in a much slower sampling speed than the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
658,
506,
671
],
"spans": [
{
"bbox": [
106,
658,
506,
671
],
"score": 1.0,
"content": "single-step GANs or VAEs. Such inefficiency is becoming a critical bottleneck for the adoption of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
667,
471,
682
],
"spans": [
{
"bbox": [
105,
667,
471,
682
],
"score": 1.0,
"content": "DPMs in downstream tasks, leading to an urgent request to design fast samplers for DPMs.",
"type": "text"
}
],
"index": 39
}
],
"index": 34
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
118,
691,
385,
712
],
"lines": [
{
"bbox": [
117,
688,
208,
703
],
"spans": [
{
"bbox": [
117,
688,
208,
703
],
"score": 1.0,
"content": "∗Corresponding Author.",
"type": "text"
}
]
},
{
"bbox": [
118,
699,
386,
714
],
"spans": [
{
"bbox": [
118,
699,
386,
714
],
"score": 1.0,
"content": "2Code is available at https://github.com/LuChengTHU/dpm-solver",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
106,
731,
385,
742
],
"lines": [
{
"bbox": [
106,
730,
385,
743
],
"spans": [
{
"bbox": [
106,
730,
385,
743
],
"score": 1.0,
"content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
121,
77,
490,
83
],
"lines": []
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
122,
97,
489,
137
],
"lines": [
{
"bbox": [
133,
97,
478,
116
],
"spans": [
{
"bbox": [
133,
97,
478,
116
],
"score": 1.0,
"content": "DPM-Solver: A Fast ODE Solver for Diffusion",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
121,
117,
489,
138
],
"spans": [
{
"bbox": [
121,
117,
489,
138
],
"score": 1.0,
"content": "Probabilistic Model Sampling in Around 10 Steps",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
127,
178,
484,
260
],
"lines": [
{
"bbox": [
136,
178,
479,
191
],
"spans": [
{
"bbox": [
136,
178,
166,
191
],
"score": 1.0,
"content": "Cheng",
"type": "text"
},
{
"bbox": [
167,
178,
184,
190
],
"score": 0.47,
"content": "\\mathbf { L } \\mathbf { u } ^ { \\dagger }",
"type": "inline_equation"
},
{
"bbox": [
185,
178,
326,
191
],
"score": 1.0,
"content": ", Yuhao Zhou†, Fan Bao†, Jianfei",
"type": "text"
},
{
"bbox": [
326,
178,
357,
190
],
"score": 0.36,
"content": "\\mathbf { C h e n } ^ { \\dagger * }",
"type": "inline_equation"
},
{
"bbox": [
358,
178,
412,
191
],
"score": 1.0,
"content": ", Chongxuan",
"type": "text"
},
{
"bbox": [
413,
178,
428,
190
],
"score": 0.56,
"content": "\\mathbf { L i } ^ { \\dagger }",
"type": "inline_equation"
},
{
"bbox": [
428,
178,
479,
191
],
"score": 1.0,
"content": ", Jun Zhu†∗",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
156,
189,
455,
205
],
"spans": [
{
"bbox": [
156,
189,
455,
205
],
"score": 1.0,
"content": "†Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, THBI Lab",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
145,
200,
466,
217
],
"spans": [
{
"bbox": [
145,
200,
466,
217
],
"score": 1.0,
"content": "†Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084 China",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
160,
212,
450,
229
],
"spans": [
{
"bbox": [
160,
212,
450,
229
],
"score": 1.0,
"content": "‡Gaoling School of Artificial Intelligence, Renmin University of China,",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
124,
225,
486,
240
],
"spans": [
{
"bbox": [
124,
225,
486,
240
],
"score": 1.0,
"content": "‡Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
138,
236,
472,
251
],
"spans": [
{
"bbox": [
138,
236,
472,
251
],
"score": 1.0,
"content": "{lucheng.lc15, yuhaoz.cs}@gmail.com; bf19@mails.tsinghua.edu.cn",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
154,
248,
457,
262
],
"spans": [
{
"bbox": [
154,
248,
457,
262
],
"score": 1.0,
"content": "chongxuanli@ruc.edu.cn; {jianfeic, dcszj}@tsinghua.edu.cn",
"type": "text"
}
],
"index": 8
}
],
"index": 5,
"bbox_fs": [
124,
178,
486,
262
]
},
{
"type": "title",
"bbox": [
283,
288,
328,
301
],
"lines": [
{
"bbox": [
281,
287,
330,
302
],
"spans": [
{
"bbox": [
281,
287,
330,
302
],
"score": 1.0,
"content": "Abstract",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
143,
313,
469,
510
],
"lines": [
{
"bbox": [
141,
314,
471,
327
],
"spans": [
{
"bbox": [
141,
314,
471,
327
],
"score": 1.0,
"content": "Diffusion probabilistic models (DPMs) are emerging powerful generative models.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
325,
470,
338
],
"spans": [
{
"bbox": [
141,
325,
470,
338
],
"score": 1.0,
"content": "Despite their high-quality generation performance, DPMs still suffer from their",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
141,
336,
469,
349
],
"spans": [
{
"bbox": [
141,
336,
469,
349
],
"score": 1.0,
"content": "slow sampling as they generally need hundreds or thousands of sequential function",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
347,
469,
360
],
"spans": [
{
"bbox": [
141,
347,
469,
360
],
"score": 1.0,
"content": "evaluations (steps) of large neural networks to draw a sample. Sampling from",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
357,
469,
370
],
"spans": [
{
"bbox": [
141,
357,
469,
370
],
"score": 1.0,
"content": "DPMs can be viewed alternatively as solving the corresponding diffusion ordinary",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
369,
470,
381
],
"spans": [
{
"bbox": [
141,
369,
470,
381
],
"score": 1.0,
"content": "differential equations (ODEs). In this work, we propose an exact formulation of",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
380,
469,
392
],
"spans": [
{
"bbox": [
141,
380,
469,
392
],
"score": 1.0,
"content": "the solution of diffusion ODEs. The formulation analytically computes the linear",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
390,
469,
403
],
"spans": [
{
"bbox": [
141,
390,
469,
403
],
"score": 1.0,
"content": "part of the solution, rather than leaving all terms to black-box ODE solvers as",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
401,
470,
414
],
"spans": [
{
"bbox": [
141,
401,
470,
414
],
"score": 1.0,
"content": "adopted in previous works. By applying change-of-variable, the solution can be",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
412,
470,
424
],
"spans": [
{
"bbox": [
141,
412,
470,
424
],
"score": 1.0,
"content": "equivalently simplified to an exponentially weighted integral of the neural network.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
423,
470,
436
],
"spans": [
{
"bbox": [
141,
423,
470,
436
],
"score": 1.0,
"content": "Based on our formulation, we propose DPM-Solver, a fast dedicated high-order",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
141,
433,
470,
447
],
"spans": [
{
"bbox": [
141,
433,
470,
447
],
"score": 1.0,
"content": "solver for diffusion ODEs with the convergence order guarantee. DPM-Solver",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
445,
470,
457
],
"spans": [
{
"bbox": [
141,
445,
470,
457
],
"score": 1.0,
"content": "is suitable for both discrete-time and continuous-time DPMs without any further",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
456,
469,
469
],
"spans": [
{
"bbox": [
141,
456,
469,
469
],
"score": 1.0,
"content": "training. Experimental results show that DPM-Solver can generate high-quality",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
141,
467,
469,
479
],
"spans": [
{
"bbox": [
141,
467,
469,
479
],
"score": 1.0,
"content": "samples in only 10 to 20 function evaluations on various datasets. We achieve",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
142,
478,
469,
489
],
"spans": [
{
"bbox": [
142,
478,
179,
488
],
"score": 0.27,
"content": "4 . 7 0 \\ : \\mathrm { F I D }",
"type": "inline_equation"
},
{
"bbox": [
179,
478,
469,
489
],
"score": 1.0,
"content": "in 10 function evaluations and 2.87 FID in 20 function evaluations on the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
141,
488,
470,
501
],
"spans": [
{
"bbox": [
141,
488,
238,
501
],
"score": 1.0,
"content": "CIFAR10 dataset, and a",
"type": "text"
},
{
"bbox": [
238,
489,
276,
499
],
"score": 0.91,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
276,
488,
470,
501
],
"score": 1.0,
"content": "speedup compared with previous state-of-the-art",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
141,
500,
317,
511
],
"spans": [
{
"bbox": [
141,
500,
317,
511
],
"score": 1.0,
"content": "training-free samplers on various datasets.2",
"type": "text"
}
],
"index": 27
}
],
"index": 18.5,
"bbox_fs": [
141,
314,
471,
511
]
},
{
"type": "title",
"bbox": [
107,
533,
190,
547
],
"lines": [
{
"bbox": [
105,
532,
192,
549
],
"spans": [
{
"bbox": [
105,
532,
192,
549
],
"score": 1.0,
"content": "1 Introduction",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
559,
506,
680
],
"lines": [
{
"bbox": [
105,
558,
507,
573
],
"spans": [
{
"bbox": [
105,
558,
507,
573
],
"score": 1.0,
"content": "Diffusion probabilistic models (DPMs) [1–3] are emerging powerful generative models with promis-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
570,
506,
585
],
"spans": [
{
"bbox": [
105,
570,
506,
585
],
"score": 1.0,
"content": "ing performance on many tasks, such as image generation [4, 5], video generation [6], text-to-image",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
581,
507,
595
],
"spans": [
{
"bbox": [
105,
581,
507,
595
],
"score": 1.0,
"content": "generation [7], speech synthesis [8, 9] and lossless compression [10]. DPMs are defined by discrete-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
592,
505,
605
],
"spans": [
{
"bbox": [
106,
592,
505,
605
],
"score": 1.0,
"content": "time random processes [1, 2] or continuous-time stochastic differential equations (SDEs) [3], which",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
602,
506,
617
],
"spans": [
{
"bbox": [
105,
602,
506,
617
],
"score": 1.0,
"content": "learn to gradually remove the noise added to the data points. Compared with the widely-used genera-",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
614,
505,
627
],
"spans": [
{
"bbox": [
105,
614,
505,
627
],
"score": 1.0,
"content": "tive adversarial networks (GANs) [11] and variational auto-encoders (VAEs) [12], DPMs can not only",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
625,
506,
638
],
"spans": [
{
"bbox": [
106,
625,
506,
638
],
"score": 1.0,
"content": "compute exact likelihood [3], but also achieve even better sample quality for image generation [4].",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
636,
505,
648
],
"spans": [
{
"bbox": [
106,
636,
505,
648
],
"score": 1.0,
"content": "However, to obtain high-quality samples, DPMs usually need hundreds or thousands of sequential",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
648,
506,
659
],
"spans": [
{
"bbox": [
105,
648,
506,
659
],
"score": 1.0,
"content": "steps of large neural network evaluations, thereby resulting in a much slower sampling speed than the",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
658,
506,
671
],
"spans": [
{
"bbox": [
106,
658,
506,
671
],
"score": 1.0,
"content": "single-step GANs or VAEs. Such inefficiency is becoming a critical bottleneck for the adoption of",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
667,
471,
682
],
"spans": [
{
"bbox": [
105,
667,
471,
682
],
"score": 1.0,
"content": "DPMs in downstream tasks, leading to an urgent request to design fast samplers for DPMs.",
"type": "text"
}
],
"index": 39
}
],
"index": 34,
"bbox_fs": [
105,
558,
507,
682
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
107,
70,
491,
243
],
"blocks": [
{
"type": "image_body",
"bbox": [
107,
70,
491,
243
],
"group_id": 0,
"lines": [
{
"bbox": [
107,
70,
491,
243
],
"spans": [
{
"bbox": [
107,
70,
491,
243
],
"score": 0.973,
"type": "image",
"image_path": "4d6dd94407a2a26c38622fe661183fe6ba98b37b4ec1a90af34b958f1acc3ac9.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
107,
70,
491,
127.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
107,
127.66666666666666,
491,
185.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
107,
185.33333333333331,
491,
242.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
105,
249,
505,
270
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
248,
506,
261
],
"spans": [
{
"bbox": [
105,
248,
506,
261
],
"score": 1.0,
"content": "Figure 1: Samples by DDIM [19] with 10, 15, 20, 100 number of function evaluations (NFE), and DPM-Solver",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
258,
488,
271
],
"spans": [
{
"bbox": [
105,
258,
349,
271
],
"score": 1.0,
"content": "(ours) with only 10 NFE, using the pre-trained DPMs on ImageNet",
"type": "text"
},
{
"bbox": [
349,
259,
385,
269
],
"score": 0.86,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
385,
258,
488,
271
],
"score": 1.0,
"content": "with classifier guidance [4].",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
107,
281,
505,
401
],
"lines": [
{
"bbox": [
105,
280,
505,
293
],
"spans": [
{
"bbox": [
105,
280,
505,
293
],
"score": 1.0,
"content": "Existing fast samplers for DPMs can be divided into two categories. The first category includes",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
291,
505,
304
],
"spans": [
{
"bbox": [
106,
291,
505,
304
],
"score": 1.0,
"content": "knowledge distillation [13, 14] and noise level or sample trajectory learning [15–18]. Such methods",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
302,
506,
316
],
"spans": [
{
"bbox": [
105,
302,
506,
316
],
"score": 1.0,
"content": "require a possibly expensive training stage before they can be used for efficient sampling. Furthermore,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
314,
505,
326
],
"spans": [
{
"bbox": [
106,
314,
505,
326
],
"score": 1.0,
"content": "their applicability and flexibility might be limited. It might require nontrivial effort to adapt the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
324,
506,
338
],
"spans": [
{
"bbox": [
105,
324,
506,
338
],
"score": 1.0,
"content": "method to different models, datasets, and number of sampling steps. The second category consists",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
335,
506,
348
],
"spans": [
{
"bbox": [
105,
335,
506,
348
],
"score": 1.0,
"content": "of training-free [19–21] samplers, which are suitable for all pre-trained DPMs in a simple plug-and-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
346,
506,
359
],
"spans": [
{
"bbox": [
106,
346,
506,
359
],
"score": 1.0,
"content": "play manner. Training-free samplers include adopting implicit [19] or analytical [21] generation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
357,
506,
370
],
"spans": [
{
"bbox": [
106,
357,
506,
370
],
"score": 1.0,
"content": "process, advanced differential equation (DE) solvers [3, 20, 22–24] and dynamic programming [18].",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
367,
506,
381
],
"spans": [
{
"bbox": [
105,
367,
251,
381
],
"score": 1.0,
"content": "However, these methods still require",
"type": "text"
},
{
"bbox": [
252,
369,
273,
378
],
"score": 0.85,
"content": "\\sim 5 0",
"type": "inline_equation"
},
{
"bbox": [
274,
367,
506,
381
],
"score": 1.0,
"content": "function evaluations [21] to generate high-quality samples",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
379,
505,
392
],
"spans": [
{
"bbox": [
106,
379,
505,
392
],
"score": 1.0,
"content": "(comparable to those generated by plain samplers in about 1000 function evaluations), thereby are",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
389,
194,
403
],
"spans": [
{
"bbox": [
105,
389,
194,
403
],
"score": 1.0,
"content": "still time-consuming.",
"type": "text"
}
],
"index": 15
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
406,
505,
537
],
"lines": [
{
"bbox": [
105,
406,
505,
419
],
"spans": [
{
"bbox": [
105,
406,
505,
419
],
"score": 1.0,
"content": "In this work, we bring the efficiency of training-free samplers to a new level to produce high-quality",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
417,
505,
430
],
"spans": [
{
"bbox": [
105,
417,
505,
430
],
"score": 1.0,
"content": "samples in the “few-step sampling” regime, where the sampling can be done within around 10 steps",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
428,
505,
441
],
"spans": [
{
"bbox": [
106,
428,
505,
441
],
"score": 1.0,
"content": "of sequential function evaluations. We tackle the alternative problem of sampling from DPMs as",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
438,
505,
452
],
"spans": [
{
"bbox": [
105,
438,
505,
452
],
"score": 1.0,
"content": "solving the corresponding diffusion ordinary differential equations (ODEs) of DPMs, and carefully",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
450,
506,
462
],
"spans": [
{
"bbox": [
105,
450,
506,
462
],
"score": 1.0,
"content": "examine the structure of diffusion ODEs. Diffusion ODEs have a semi-linear structure — they consist",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
461,
506,
473
],
"spans": [
{
"bbox": [
105,
461,
506,
473
],
"score": 1.0,
"content": "of a linear function of the data variable and a nonlinear function parameterized by neural networks.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
472,
505,
484
],
"spans": [
{
"bbox": [
105,
472,
505,
484
],
"score": 1.0,
"content": "Such structure is omitted in previous training-free samplers [3, 20], which directly use black-box",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
505,
495
],
"score": 1.0,
"content": "DE solvers. To utilize the semi-linear structure, we derive an exact formulation of the solutions of",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
492,
505,
507
],
"spans": [
{
"bbox": [
105,
492,
505,
507
],
"score": 1.0,
"content": "diffusion ODEs by analytically computing the linear part of the solutions, avoiding the corresponding",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
504,
505,
518
],
"spans": [
{
"bbox": [
105,
504,
505,
518
],
"score": 1.0,
"content": "discretization error. Furthermore, by applying change-of-variable, the solutions can be equivalently",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
515,
506,
528
],
"spans": [
{
"bbox": [
105,
515,
506,
528
],
"score": 1.0,
"content": "simplified to an exponentially weighted integral of the neural network. Such integral is very special",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
526,
487,
539
],
"spans": [
{
"bbox": [
105,
526,
487,
539
],
"score": 1.0,
"content": "and can be efficiently approximated by the numerical methods for exponential integrators [25].",
"type": "text"
}
],
"index": 27
}
],
"index": 21.5
},
{
"type": "text",
"bbox": [
107,
542,
505,
663
],
"lines": [
{
"bbox": [
106,
543,
505,
555
],
"spans": [
{
"bbox": [
106,
543,
505,
555
],
"score": 1.0,
"content": "Based on our formulation of solutions, we propose DPM-Solver, a fast dedicated solver for diffusion",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
553,
506,
567
],
"spans": [
{
"bbox": [
105,
553,
506,
567
],
"score": 1.0,
"content": "ODEs by approximating the above integral. Specifically, we propose first-order, second-order and",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
506,
577
],
"score": 1.0,
"content": "third-order versions of DPM-Solver with convergence order guarantees. We further propose an",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
576,
506,
588
],
"spans": [
{
"bbox": [
106,
576,
506,
588
],
"score": 1.0,
"content": "adaptive step size schedule for DPM-Solver. In general, DPM-Solver is applicable to both continuous-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
585,
507,
600
],
"spans": [
{
"bbox": [
105,
585,
507,
600
],
"score": 1.0,
"content": "time and discrete-time DPMs, and also conditional sampling with classifier guidance [4]. Fig. 1",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
597,
506,
611
],
"spans": [
{
"bbox": [
105,
597,
506,
611
],
"score": 1.0,
"content": "demonstrates the speedup performance of a Denoising Diffusion Implicit Models (DDIM) [19]",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
607,
505,
621
],
"spans": [
{
"bbox": [
105,
607,
505,
621
],
"score": 1.0,
"content": "baseline and DPM-Solver, which shows that DPM-Solver can generate high-quality samples with as",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
619,
506,
631
],
"spans": [
{
"bbox": [
106,
619,
506,
631
],
"score": 1.0,
"content": "few as 10 function evaluations and is much faster than DDIM on the ImageNet 256x256 dataset [26].",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
629,
506,
642
],
"spans": [
{
"bbox": [
105,
629,
506,
642
],
"score": 1.0,
"content": "Our additional experimental results show that DPM-Solver can greatly improve the sampling speed of",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
641,
505,
653
],
"spans": [
{
"bbox": [
105,
641,
505,
653
],
"score": 1.0,
"content": "both discrete-time and continuous-time DPMs, and it can achieve excellent sample quality in around",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
651,
489,
664
],
"spans": [
{
"bbox": [
105,
651,
489,
664
],
"score": 1.0,
"content": "10 function evaluations, which is much faster than all previous training-free samplers of DPMs.",
"type": "text"
}
],
"index": 38
}
],
"index": 33
},
{
"type": "title",
"bbox": [
108,
682,
281,
696
],
"lines": [
{
"bbox": [
104,
682,
282,
698
],
"spans": [
{
"bbox": [
104,
682,
282,
698
],
"score": 1.0,
"content": "2 Diffusion Probabilistic Models",
"type": "text"
}
],
"index": 39
}
],
"index": 39
},
{
"type": "text",
"bbox": [
105,
711,
500,
722
],
"lines": [
{
"bbox": [
106,
710,
501,
724
],
"spans": [
{
"bbox": [
106,
710,
501,
724
],
"score": 1.0,
"content": "We review diffusion probabilistic models and their associated differential equations in this section.",
"type": "text"
}
],
"index": 40
}
],
"index": 40
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
742,
309,
750
],
"lines": [
{
"bbox": [
301,
741,
310,
753
],
"spans": [
{
"bbox": [
301,
741,
310,
753
],
"score": 1.0,
"content": "2",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
107,
70,
491,
243
],
"blocks": [
{
"type": "image_body",
"bbox": [
107,
70,
491,
243
],
"group_id": 0,
"lines": [
{
"bbox": [
107,
70,
491,
243
],
"spans": [
{
"bbox": [
107,
70,
491,
243
],
"score": 0.973,
"type": "image",
"image_path": "4d6dd94407a2a26c38622fe661183fe6ba98b37b4ec1a90af34b958f1acc3ac9.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
107,
70,
491,
127.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
107,
127.66666666666666,
491,
185.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
107,
185.33333333333331,
491,
242.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
105,
249,
505,
270
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
248,
506,
261
],
"spans": [
{
"bbox": [
105,
248,
506,
261
],
"score": 1.0,
"content": "Figure 1: Samples by DDIM [19] with 10, 15, 20, 100 number of function evaluations (NFE), and DPM-Solver",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
258,
488,
271
],
"spans": [
{
"bbox": [
105,
258,
349,
271
],
"score": 1.0,
"content": "(ours) with only 10 NFE, using the pre-trained DPMs on ImageNet",
"type": "text"
},
{
"bbox": [
349,
259,
385,
269
],
"score": 0.86,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
385,
258,
488,
271
],
"score": 1.0,
"content": "with classifier guidance [4].",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
107,
281,
505,
401
],
"lines": [
{
"bbox": [
105,
280,
505,
293
],
"spans": [
{
"bbox": [
105,
280,
505,
293
],
"score": 1.0,
"content": "Existing fast samplers for DPMs can be divided into two categories. The first category includes",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
291,
505,
304
],
"spans": [
{
"bbox": [
106,
291,
505,
304
],
"score": 1.0,
"content": "knowledge distillation [13, 14] and noise level or sample trajectory learning [15–18]. Such methods",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
302,
506,
316
],
"spans": [
{
"bbox": [
105,
302,
506,
316
],
"score": 1.0,
"content": "require a possibly expensive training stage before they can be used for efficient sampling. Furthermore,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
314,
505,
326
],
"spans": [
{
"bbox": [
106,
314,
505,
326
],
"score": 1.0,
"content": "their applicability and flexibility might be limited. It might require nontrivial effort to adapt the",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
324,
506,
338
],
"spans": [
{
"bbox": [
105,
324,
506,
338
],
"score": 1.0,
"content": "method to different models, datasets, and number of sampling steps. The second category consists",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
335,
506,
348
],
"spans": [
{
"bbox": [
105,
335,
506,
348
],
"score": 1.0,
"content": "of training-free [19–21] samplers, which are suitable for all pre-trained DPMs in a simple plug-and-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
346,
506,
359
],
"spans": [
{
"bbox": [
106,
346,
506,
359
],
"score": 1.0,
"content": "play manner. Training-free samplers include adopting implicit [19] or analytical [21] generation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
357,
506,
370
],
"spans": [
{
"bbox": [
106,
357,
506,
370
],
"score": 1.0,
"content": "process, advanced differential equation (DE) solvers [3, 20, 22–24] and dynamic programming [18].",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
367,
506,
381
],
"spans": [
{
"bbox": [
105,
367,
251,
381
],
"score": 1.0,
"content": "However, these methods still require",
"type": "text"
},
{
"bbox": [
252,
369,
273,
378
],
"score": 0.85,
"content": "\\sim 5 0",
"type": "inline_equation"
},
{
"bbox": [
274,
367,
506,
381
],
"score": 1.0,
"content": "function evaluations [21] to generate high-quality samples",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
379,
505,
392
],
"spans": [
{
"bbox": [
106,
379,
505,
392
],
"score": 1.0,
"content": "(comparable to those generated by plain samplers in about 1000 function evaluations), thereby are",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
389,
194,
403
],
"spans": [
{
"bbox": [
105,
389,
194,
403
],
"score": 1.0,
"content": "still time-consuming.",
"type": "text"
}
],
"index": 15
}
],
"index": 10,
"bbox_fs": [
105,
280,
506,
403
]
},
{
"type": "text",
"bbox": [
107,
406,
505,
537
],
"lines": [
{
"bbox": [
105,
406,
505,
419
],
"spans": [
{
"bbox": [
105,
406,
505,
419
],
"score": 1.0,
"content": "In this work, we bring the efficiency of training-free samplers to a new level to produce high-quality",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
417,
505,
430
],
"spans": [
{
"bbox": [
105,
417,
505,
430
],
"score": 1.0,
"content": "samples in the “few-step sampling” regime, where the sampling can be done within around 10 steps",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
428,
505,
441
],
"spans": [
{
"bbox": [
106,
428,
505,
441
],
"score": 1.0,
"content": "of sequential function evaluations. We tackle the alternative problem of sampling from DPMs as",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
438,
505,
452
],
"spans": [
{
"bbox": [
105,
438,
505,
452
],
"score": 1.0,
"content": "solving the corresponding diffusion ordinary differential equations (ODEs) of DPMs, and carefully",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
450,
506,
462
],
"spans": [
{
"bbox": [
105,
450,
506,
462
],
"score": 1.0,
"content": "examine the structure of diffusion ODEs. Diffusion ODEs have a semi-linear structure — they consist",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
461,
506,
473
],
"spans": [
{
"bbox": [
105,
461,
506,
473
],
"score": 1.0,
"content": "of a linear function of the data variable and a nonlinear function parameterized by neural networks.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
472,
505,
484
],
"spans": [
{
"bbox": [
105,
472,
505,
484
],
"score": 1.0,
"content": "Such structure is omitted in previous training-free samplers [3, 20], which directly use black-box",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
505,
495
],
"score": 1.0,
"content": "DE solvers. To utilize the semi-linear structure, we derive an exact formulation of the solutions of",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
492,
505,
507
],
"spans": [
{
"bbox": [
105,
492,
505,
507
],
"score": 1.0,
"content": "diffusion ODEs by analytically computing the linear part of the solutions, avoiding the corresponding",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
504,
505,
518
],
"spans": [
{
"bbox": [
105,
504,
505,
518
],
"score": 1.0,
"content": "discretization error. Furthermore, by applying change-of-variable, the solutions can be equivalently",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
515,
506,
528
],
"spans": [
{
"bbox": [
105,
515,
506,
528
],
"score": 1.0,
"content": "simplified to an exponentially weighted integral of the neural network. Such integral is very special",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
526,
487,
539
],
"spans": [
{
"bbox": [
105,
526,
487,
539
],
"score": 1.0,
"content": "and can be efficiently approximated by the numerical methods for exponential integrators [25].",
"type": "text"
}
],
"index": 27
}
],
"index": 21.5,
"bbox_fs": [
105,
406,
506,
539
]
},
{
"type": "text",
"bbox": [
107,
542,
505,
663
],
"lines": [
{
"bbox": [
106,
543,
505,
555
],
"spans": [
{
"bbox": [
106,
543,
505,
555
],
"score": 1.0,
"content": "Based on our formulation of solutions, we propose DPM-Solver, a fast dedicated solver for diffusion",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
553,
506,
567
],
"spans": [
{
"bbox": [
105,
553,
506,
567
],
"score": 1.0,
"content": "ODEs by approximating the above integral. Specifically, we propose first-order, second-order and",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
506,
577
],
"score": 1.0,
"content": "third-order versions of DPM-Solver with convergence order guarantees. We further propose an",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
576,
506,
588
],
"spans": [
{
"bbox": [
106,
576,
506,
588
],
"score": 1.0,
"content": "adaptive step size schedule for DPM-Solver. In general, DPM-Solver is applicable to both continuous-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
585,
507,
600
],
"spans": [
{
"bbox": [
105,
585,
507,
600
],
"score": 1.0,
"content": "time and discrete-time DPMs, and also conditional sampling with classifier guidance [4]. Fig. 1",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
597,
506,
611
],
"spans": [
{
"bbox": [
105,
597,
506,
611
],
"score": 1.0,
"content": "demonstrates the speedup performance of a Denoising Diffusion Implicit Models (DDIM) [19]",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
607,
505,
621
],
"spans": [
{
"bbox": [
105,
607,
505,
621
],
"score": 1.0,
"content": "baseline and DPM-Solver, which shows that DPM-Solver can generate high-quality samples with as",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
619,
506,
631
],
"spans": [
{
"bbox": [
106,
619,
506,
631
],
"score": 1.0,
"content": "few as 10 function evaluations and is much faster than DDIM on the ImageNet 256x256 dataset [26].",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
629,
506,
642
],
"spans": [
{
"bbox": [
105,
629,
506,
642
],
"score": 1.0,
"content": "Our additional experimental results show that DPM-Solver can greatly improve the sampling speed of",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
641,
505,
653
],
"spans": [
{
"bbox": [
105,
641,
505,
653
],
"score": 1.0,
"content": "both discrete-time and continuous-time DPMs, and it can achieve excellent sample quality in around",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
651,
489,
664
],
"spans": [
{
"bbox": [
105,
651,
489,
664
],
"score": 1.0,
"content": "10 function evaluations, which is much faster than all previous training-free samplers of DPMs.",
"type": "text"
}
],
"index": 38
}
],
"index": 33,
"bbox_fs": [
105,
543,
507,
664
]
},
{
"type": "title",
"bbox": [
108,
682,
281,
696
],
"lines": [
{
"bbox": [
104,
682,
282,
698
],
"spans": [
{
"bbox": [
104,
682,
282,
698
],
"score": 1.0,
"content": "2 Diffusion Probabilistic Models",
"type": "text"
}
],
"index": 39
}
],
"index": 39
},
{
"type": "text",
"bbox": [
105,
711,
500,
722
],
"lines": [
{
"bbox": [
106,
710,
501,
724
],
"spans": [
{
"bbox": [
106,
710,
501,
724
],
"score": 1.0,
"content": "We review diffusion probabilistic models and their associated differential equations in this section.",
"type": "text"
}
],
"index": 40
}
],
"index": 40,
"bbox_fs": [
106,
710,
501,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
72,
289,
84
],
"lines": [
{
"bbox": [
106,
72,
289,
86
],
"spans": [
{
"bbox": [
106,
72,
289,
86
],
"score": 1.0,
"content": "2.1 Forward Process and Diffusion SDEs",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
92,
505,
128
],
"lines": [
{
"bbox": [
105,
92,
505,
106
],
"spans": [
{
"bbox": [
105,
92,
204,
106
],
"score": 1.0,
"content": "Assume that we have a",
"type": "text"
},
{
"bbox": [
205,
94,
214,
103
],
"score": 0.82,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
214,
92,
339,
106
],
"score": 1.0,
"content": "-dimensional random variable",
"type": "text"
},
{
"bbox": [
339,
92,
379,
104
],
"score": 0.92,
"content": "\\pmb { x } _ { 0 } \\in \\mathbb { R } ^ { D }",
"type": "inline_equation"
},
{
"bbox": [
380,
92,
505,
106
],
"score": 1.0,
"content": "with an unknown distribution",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
107,
103,
505,
118
],
"spans": [
{
"bbox": [
107,
104,
135,
116
],
"score": 0.9,
"content": "q _ { 0 } ( { \\pmb x } _ { 0 } )",
"type": "inline_equation"
},
{
"bbox": [
135,
103,
437,
118
],
"score": 1.0,
"content": ". Diffusion Probabilistic Models (DPMs) [1–3, 10] define a forward process",
"type": "text"
},
{
"bbox": [
438,
104,
483,
117
],
"score": 0.93,
"content": "\\{ \\pmb { x } _ { t } \\} _ { t \\in [ 0 , T ] }",
"type": "inline_equation"
},
{
"bbox": [
484,
103,
505,
118
],
"score": 1.0,
"content": "with",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
107,
114,
506,
130
],
"spans": [
{
"bbox": [
107,
117,
133,
127
],
"score": 0.88,
"content": "T > 0",
"type": "inline_equation"
},
{
"bbox": [
134,
114,
186,
130
],
"score": 1.0,
"content": "starting with",
"type": "text"
},
{
"bbox": [
186,
118,
198,
127
],
"score": 0.85,
"content": "\\scriptstyle { \\mathbf { { \\mathit { x } } } } _ { 0 }",
"type": "inline_equation"
},
{
"bbox": [
198,
114,
270,
130
],
"score": 1.0,
"content": ", such that for any",
"type": "text"
},
{
"bbox": [
270,
116,
308,
128
],
"score": 0.93,
"content": "t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
309,
114,
385,
130
],
"score": 1.0,
"content": ", the distribution of",
"type": "text"
},
{
"bbox": [
385,
117,
396,
127
],
"score": 0.86,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
396,
114,
458,
130
],
"score": 1.0,
"content": "conditioned on",
"type": "text"
},
{
"bbox": [
458,
118,
470,
127
],
"score": 0.86,
"content": "\\scriptstyle { \\mathbf { { \\mathit { x } } } } _ { 0 }",
"type": "inline_equation"
},
{
"bbox": [
470,
114,
506,
130
],
"score": 1.0,
"content": "satisfies",
"type": "text"
}
],
"index": 3
}
],
"index": 2
},
{
"type": "interline_equation",
"bbox": [
228,
133,
381,
149
],
"lines": [
{
"bbox": [
228,
133,
381,
149
],
"spans": [
{
"bbox": [
228,
133,
381,
149
],
"score": 0.92,
"content": "\\begin{array} { r } { q _ { 0 t } ( \\pmb { x } _ { t } | \\pmb { x } _ { 0 } ) = \\mathcal { N } ( \\pmb { x } _ { t } | \\alpha ( t ) \\pmb { x } _ { 0 } , \\sigma ^ { 2 } ( t ) \\pmb { I } ) , } \\end{array}",
"type": "interline_equation",
"image_path": "b98f04c1bbffa9531e257210ae6112c502489822e557b5a3afdb274392b83d54.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
228,
133,
381,
149
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
153,
505,
221
],
"lines": [
{
"bbox": [
106,
153,
505,
166
],
"spans": [
{
"bbox": [
106,
153,
134,
166
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
154,
202,
166
],
"score": 0.92,
"content": "\\alpha ( t ) , \\sigma ( t ) \\in \\mathbb { R } ^ { + }",
"type": "inline_equation"
},
{
"bbox": [
202,
153,
328,
166
],
"score": 1.0,
"content": "are differentiable functions of",
"type": "text"
},
{
"bbox": [
329,
155,
334,
164
],
"score": 0.81,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
334,
153,
505,
166
],
"score": 1.0,
"content": "with bounded derivatives, and we denote",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
165,
506,
178
],
"spans": [
{
"bbox": [
105,
165,
141,
178
],
"score": 1.0,
"content": "them as",
"type": "text"
},
{
"bbox": [
142,
168,
166,
177
],
"score": 0.88,
"content": "\\alpha _ { t } , \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
167,
165,
295,
178
],
"score": 1.0,
"content": "for simplicity. The choice for",
"type": "text"
},
{
"bbox": [
295,
167,
307,
176
],
"score": 0.86,
"content": "\\alpha _ { t }",
"type": "inline_equation"
},
{
"bbox": [
307,
165,
326,
178
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
326,
167,
337,
176
],
"score": 0.86,
"content": "\\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
337,
165,
506,
178
],
"score": 1.0,
"content": "is referred to as the noise schedule of a",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
175,
505,
188
],
"spans": [
{
"bbox": [
105,
175,
150,
188
],
"score": 1.0,
"content": "DPM. Let",
"type": "text"
},
{
"bbox": [
151,
177,
177,
187
],
"score": 0.9,
"content": "q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
177,
175,
323,
188
],
"score": 1.0,
"content": "denote the marginal distribution of",
"type": "text"
},
{
"bbox": [
324,
178,
334,
187
],
"score": 0.87,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
335,
175,
505,
188
],
"score": 1.0,
"content": ", DPMs choose noise schedules to ensure",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
186,
506,
200
],
"spans": [
{
"bbox": [
105,
186,
124,
200
],
"score": 1.0,
"content": "that",
"type": "text"
},
{
"bbox": [
124,
187,
229,
199
],
"score": 0.89,
"content": "q _ { T } ( \\pmb { x } _ { T } ) \\overset { \\cdot } { \\approx } \\dot { \\mathcal { N } } ( \\pmb { x } _ { T } | \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\pmb { I } )",
"type": "inline_equation"
},
{
"bbox": [
229,
186,
267,
200
],
"score": 1.0,
"content": "for some",
"type": "text"
},
{
"bbox": [
267,
187,
293,
198
],
"score": 0.9,
"content": "\\tilde { \\sigma } > 0",
"type": "inline_equation"
},
{
"bbox": [
293,
186,
438,
200
],
"score": 1.0,
"content": ", and the signal-to-noise-ratio (SNR)",
"type": "text"
},
{
"bbox": [
438,
186,
465,
199
],
"score": 0.93,
"content": "\\alpha _ { t } ^ { 2 } / \\sigma _ { t } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
466,
186,
506,
200
],
"score": 1.0,
"content": "is strictly",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
198,
506,
211
],
"spans": [
{
"bbox": [
106,
198,
173,
211
],
"score": 1.0,
"content": "decreasing w.r.t.",
"type": "text"
},
{
"bbox": [
174,
200,
179,
208
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
179,
198,
506,
211
],
"score": 1.0,
"content": "[10]. Moreover, Kingma et al. [10] prove that the following stochastic differential",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
209,
492,
221
],
"spans": [
{
"bbox": [
105,
209,
313,
221
],
"score": 1.0,
"content": "equation (SDE) has the same transition distribution",
"type": "text"
},
{
"bbox": [
313,
209,
357,
221
],
"score": 0.93,
"content": "q _ { 0 t } ( \\pmb { x } _ { t } | \\pmb { x } _ { 0 } )",
"type": "inline_equation"
},
{
"bbox": [
358,
209,
448,
221
],
"score": 1.0,
"content": "as in Eq. (2.1) for any",
"type": "text"
},
{
"bbox": [
449,
209,
487,
221
],
"score": 0.93,
"content": "t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
488,
209,
492,
221
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 10
}
],
"index": 7.5
},
{
"type": "interline_equation",
"bbox": [
213,
226,
397,
240
],
"lines": [
{
"bbox": [
213,
226,
397,
240
],
"spans": [
{
"bbox": [
213,
226,
397,
240
],
"score": 0.91,
"content": "\\mathrm { d } \\pmb { x } _ { t } = f ( t ) \\pmb { x } _ { t } \\mathrm { d } t + g ( t ) \\mathrm { d } \\pmb { w } _ { t } , \\quad \\pmb { x } _ { 0 } \\sim q _ { 0 } ( \\pmb { x } _ { 0 } ) ,",
"type": "interline_equation",
"image_path": "8433318473b1628737ca5404d0a929d6a4570d9919cf8824f87c274bc4a0bbfd.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
213,
226,
397,
240
],
"spans": [],
"index": 11
}
]
},
{
"type": "text",
"bbox": [
106,
245,
317,
257
],
"lines": [
{
"bbox": [
105,
243,
318,
260
],
"spans": [
{
"bbox": [
105,
243,
133,
260
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
245,
172,
257
],
"score": 0.94,
"content": "{ \\pmb w } _ { t } \\in \\mathbb { R } ^ { D }",
"type": "inline_equation"
},
{
"bbox": [
172,
243,
318,
260
],
"score": 1.0,
"content": "is the standard Wiener process, and",
"type": "text"
}
],
"index": 12
}
],
"index": 12
},
{
"type": "interline_equation",
"bbox": [
206,
263,
405,
289
],
"lines": [
{
"bbox": [
206,
263,
405,
289
],
"spans": [
{
"bbox": [
206,
263,
405,
289
],
"score": 0.93,
"content": "f ( t ) = \\frac { \\mathrm { d } \\log \\alpha _ { t } } { \\mathrm { d } t } , \\quad g ^ { 2 } ( t ) = \\frac { \\mathrm { d } \\sigma _ { t } ^ { 2 } } { \\mathrm { d } t } - 2 \\frac { \\mathrm { d } \\log \\alpha _ { t } } { \\mathrm { d } t } \\sigma _ { t } ^ { 2 } .",
"type": "interline_equation",
"image_path": "e5d12cd513ddf8ae3fa219a1feddb4594e4cf22e33260b27389e15d81ab87dc7.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
206,
263,
405,
289
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
107,
294,
506,
317
],
"lines": [
{
"bbox": [
106,
294,
505,
307
],
"spans": [
{
"bbox": [
106,
294,
505,
307
],
"score": 1.0,
"content": "Under some regularity conditions, Song et al. [3] show that the forward process in Eq. (2.2) has an",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
304,
476,
318
],
"spans": [
{
"bbox": [
105,
304,
255,
318
],
"score": 1.0,
"content": "equivalent reverse process from time",
"type": "text"
},
{
"bbox": [
256,
306,
264,
315
],
"score": 0.84,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
264,
304,
276,
318
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
276,
306,
282,
315
],
"score": 0.29,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
282,
304,
439,
318
],
"score": 1.0,
"content": ", starting with the marginal distribution",
"type": "text"
},
{
"bbox": [
439,
306,
471,
317
],
"score": 0.93,
"content": "q _ { T } ( { \\pmb x } _ { T } )",
"type": "inline_equation"
},
{
"bbox": [
471,
304,
476,
318
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5
},
{
"type": "interline_equation",
"bbox": [
163,
322,
446,
338
],
"lines": [
{
"bbox": [
163,
322,
446,
338
],
"spans": [
{
"bbox": [
163,
322,
446,
338
],
"score": 0.88,
"content": "\\mathrm { d } \\pmb { x } _ { t } = [ f ( t ) \\pmb { x } _ { t } - g ^ { 2 } ( t ) \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } ) ] \\mathrm { d } t + g ( t ) \\mathrm { d } \\bar { \\pmb { w } } _ { t } , \\quad \\pmb { x } _ { T } \\sim q _ { T } ( \\pmb { x } _ { T } ) ,",
"type": "interline_equation",
"image_path": "098b1c09edd91e1e8b539459bb92a3c5f191abb42aa1157755522d3dce62bcae.jpg"
}
]
}
],
"index": 16,
"virtual_lines": [
{
"bbox": [
163,
322,
446,
338
],
"spans": [],
"index": 16
}
]
},
{
"type": "text",
"bbox": [
108,
343,
505,
388
],
"lines": [
{
"bbox": [
106,
343,
505,
355
],
"spans": [
{
"bbox": [
106,
343,
133,
355
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
344,
146,
354
],
"score": 0.88,
"content": "\\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { w } } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
147,
343,
505,
355
],
"score": 1.0,
"content": "is a standard Wiener process in the reverse time. The only unknown term in Eq. (2.4) is",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
354,
505,
367
],
"spans": [
{
"bbox": [
105,
354,
182,
367
],
"score": 1.0,
"content": "the score function",
"type": "text"
},
{
"bbox": [
183,
354,
239,
366
],
"score": 0.92,
"content": "\\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
239,
354,
293,
367
],
"score": 1.0,
"content": "at each time",
"type": "text"
},
{
"bbox": [
293,
356,
298,
364
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
299,
354,
469,
367
],
"score": 1.0,
"content": ". In practice, DPMs use a neural network",
"type": "text"
},
{
"bbox": [
469,
355,
505,
366
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
}
],
"index": 18
},
{
"bbox": [
105,
365,
506,
377
],
"spans": [
{
"bbox": [
105,
365,
180,
377
],
"score": 1.0,
"content": "parameterized by",
"type": "text"
},
{
"bbox": [
181,
366,
186,
375
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
187,
365,
345,
377
],
"score": 1.0,
"content": "to estimate the scaled score function:",
"type": "text"
},
{
"bbox": [
345,
365,
419,
377
],
"score": 0.92,
"content": "- \\sigma _ { t } \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
420,
365,
488,
377
],
"score": 1.0,
"content": ". The parameter",
"type": "text"
},
{
"bbox": [
488,
366,
494,
375
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
495,
365,
506,
377
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
376,
331,
389
],
"spans": [
{
"bbox": [
105,
376,
331,
389
],
"score": 1.0,
"content": "optimized by minimizing the following objective [2, 3]:",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5
},
{
"type": "interline_equation",
"bbox": [
165,
393,
445,
453
],
"lines": [
{
"bbox": [
165,
393,
445,
453
],
"spans": [
{
"bbox": [
165,
393,
445,
453
],
"score": 0.94,
"content": "\\begin{array} { l } { \\displaystyle \\mathcal { L } ( \\theta ; \\omega ( t ) ) : = \\frac { 1 } { 2 } \\int _ { 0 } ^ { T } \\omega ( t ) \\mathbb { E } _ { q _ { t } ( \\mathbf { \\Delta x } _ { t } ) } \\Big [ \\| \\epsilon _ { \\theta } ( \\mathbf { x } _ { t } , t ) + \\sigma _ { t } \\nabla _ { \\mathbf { x } } \\log q _ { t } ( \\mathbf { \\Delta x } _ { t } ) \\| _ { 2 } ^ { 2 } \\Big ] \\mathrm { d } t } \\\\ { \\displaystyle \\qquad = \\frac { 1 } { 2 } \\int _ { 0 } ^ { T } \\omega ( t ) \\mathbb { E } _ { q _ { 0 } ( \\mathbf { x } _ { 0 } ) } \\mathbb { E } _ { q ( \\epsilon ) } \\Big [ \\| \\epsilon _ { \\theta } ( \\mathbf { x } _ { t } , t ) - \\epsilon \\| _ { 2 } ^ { 2 } \\Big ] \\mathrm { d } t + C , } \\end{array}",
"type": "interline_equation",
"image_path": "18da74f3ada9dcef2cf0de53303bacedefeb327430ef720bfe85381fb0de2f45.jpg"
}
]
}
],
"index": 22,
"virtual_lines": [
{
"bbox": [
165,
393,
445,
413.0
],
"spans": [],
"index": 21
},
{
"bbox": [
165,
413.0,
445,
433.0
],
"spans": [],
"index": 22
},
{
"bbox": [
165,
433.0,
445,
453.0
],
"spans": [],
"index": 23
}
]
},
{
"type": "text",
"bbox": [
106,
456,
506,
513
],
"lines": [
{
"bbox": [
106,
456,
506,
470
],
"spans": [
{
"bbox": [
106,
456,
133,
470
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
457,
153,
469
],
"score": 0.92,
"content": "\\omega ( t )",
"type": "inline_equation"
},
{
"bbox": [
153,
456,
253,
470
],
"score": 1.0,
"content": "is a weighting function,",
"type": "text"
},
{
"bbox": [
253,
456,
345,
469
],
"score": 0.88,
"content": "\\epsilon \\sim q ( \\epsilon ) = \\mathcal { N } ( \\epsilon | \\mathbf { 0 } , I )",
"type": "inline_equation"
},
{
"bbox": [
345,
456,
348,
470
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
349,
458,
421,
468
],
"score": 0.88,
"content": "{ \\pmb x } _ { t } = \\alpha _ { t } { \\pmb x } _ { 0 } + \\sigma _ { t } { \\pmb \\epsilon }",
"type": "inline_equation"
},
{
"bbox": [
421,
456,
442,
470
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
442,
457,
451,
467
],
"score": 0.86,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
451,
456,
506,
470
],
"score": 1.0,
"content": "is a constant",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
468,
506,
480
],
"spans": [
{
"bbox": [
106,
468,
169,
480
],
"score": 1.0,
"content": "independent of",
"type": "text"
},
{
"bbox": [
169,
469,
174,
478
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
175,
468,
192,
480
],
"score": 1.0,
"content": ". As",
"type": "text"
},
{
"bbox": [
193,
468,
229,
480
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
229,
468,
482,
480
],
"score": 1.0,
"content": "can also be regarded as predicting the Gaussian noise added to",
"type": "text"
},
{
"bbox": [
482,
469,
493,
479
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
493,
468,
506,
480
],
"score": 1.0,
"content": ", it",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
477,
507,
493
],
"spans": [
{
"bbox": [
104,
477,
381,
493
],
"score": 1.0,
"content": "is usually called the noise prediction model. Since the ground truth of",
"type": "text"
},
{
"bbox": [
382,
479,
418,
491
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
418,
477,
429,
493
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
429,
479,
503,
491
],
"score": 0.89,
"content": "- \\sigma _ { t } \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
503,
477,
507,
493
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
488,
506,
503
],
"spans": [
{
"bbox": [
105,
488,
303,
503
],
"score": 1.0,
"content": "DPMs replace the score function in Eq. (2.4) by",
"type": "text"
},
{
"bbox": [
303,
489,
361,
501
],
"score": 0.92,
"content": "- \\mathbf { \\epsilon } \\mathbf { \\epsilon } \\bar { \\mathbf { \\alpha } } ( \\mathbf { x } _ { t } , t ) / \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
361,
488,
506,
503
],
"score": 1.0,
"content": "and define a parameterized reverse",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
500,
403,
513
],
"spans": [
{
"bbox": [
105,
500,
179,
513
],
"score": 1.0,
"content": "process (diffusion",
"type": "text"
},
{
"bbox": [
180,
501,
199,
511
],
"score": 0.45,
"content": "S D E",
"type": "inline_equation"
},
{
"bbox": [
200,
500,
246,
513
],
"score": 1.0,
"content": ") from time",
"type": "text"
},
{
"bbox": [
246,
502,
254,
510
],
"score": 0.85,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
254,
500,
266,
513
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
266,
502,
272,
510
],
"score": 0.38,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
272,
500,
328,
513
],
"score": 1.0,
"content": ", starting with",
"type": "text"
},
{
"bbox": [
329,
500,
399,
513
],
"score": 0.92,
"content": "\\pmb { x } _ { T } \\overset { \\cdot } { \\sim } \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\pmb { I } )",
"type": "inline_equation"
},
{
"bbox": [
399,
500,
403,
513
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 28
}
],
"index": 26
},
{
"type": "interline_equation",
"bbox": [
163,
518,
448,
546
],
"lines": [
{
"bbox": [
163,
518,
448,
546
],
"spans": [
{
"bbox": [
163,
518,
448,
546
],
"score": 0.94,
"content": "\\mathrm { d } x _ { t } = \\left[ f ( t ) x _ { t } + \\frac { g ^ { 2 } ( t ) } { \\sigma _ { t } } \\epsilon _ { \\theta } ( x _ { t } , t ) \\right] \\mathrm { d } t + g ( t ) \\mathrm { d } \\bar { w } _ { t } , \\quad x _ { T } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } I ) .",
"type": "interline_equation",
"image_path": "65fdde0d1d949853db59a598710255c20a35d6c68ca6401f8b46cd9f514525f2.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
163,
518,
448,
546
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
107,
551,
506,
607
],
"lines": [
{
"bbox": [
106,
551,
505,
564
],
"spans": [
{
"bbox": [
106,
551,
505,
564
],
"score": 1.0,
"content": "Samples can be generated from DPMs by solving the diffusion SDE in Eq. (2.5) with numerical",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
563,
505,
574
],
"spans": [
{
"bbox": [
106,
563,
268,
574
],
"score": 1.0,
"content": "solvers, which discretize the SDE from",
"type": "text"
},
{
"bbox": [
268,
563,
276,
572
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
277,
563,
505,
574
],
"score": 1.0,
"content": "to 0. Song et al. [3] proved that the traditional ancestral",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
573,
506,
586
],
"spans": [
{
"bbox": [
105,
573,
506,
586
],
"score": 1.0,
"content": "sampling method for DPMs [2] can be viewed as a first-order SDE solver for Eq. (2.5). However, these",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
583,
506,
597
],
"spans": [
{
"bbox": [
105,
583,
506,
597
],
"score": 1.0,
"content": "first-order methods usually need hundreds of or thousands of function evaluations to converge [3],",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
595,
279,
608
],
"spans": [
{
"bbox": [
106,
595,
279,
608
],
"score": 1.0,
"content": "leading to extremely slow sampling speed.",
"type": "text"
}
],
"index": 34
}
],
"index": 32
},
{
"type": "title",
"bbox": [
108,
619,
278,
632
],
"lines": [
{
"bbox": [
105,
618,
280,
635
],
"spans": [
{
"bbox": [
105,
618,
280,
635
],
"score": 1.0,
"content": "2.2 Diffusion (Probability Flow) ODEs",
"type": "text"
}
],
"index": 35
}
],
"index": 35
},
{
"type": "text",
"bbox": [
106,
640,
506,
696
],
"lines": [
{
"bbox": [
105,
640,
506,
654
],
"spans": [
{
"bbox": [
105,
640,
506,
654
],
"score": 1.0,
"content": "When discretizing SDEs, the step size is limited by the randomness of the Wiener process [27, Chap.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
651,
505,
665
],
"spans": [
{
"bbox": [
105,
651,
505,
665
],
"score": 1.0,
"content": "11]. A large step size (small number of steps) often causes non-convergence, especially in high",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
660,
507,
676
],
"spans": [
{
"bbox": [
105,
660,
507,
676
],
"score": 1.0,
"content": "dimensional spaces. For faster sampling, one can consider the associated probability flow ODE [3],",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
672,
507,
687
],
"spans": [
{
"bbox": [
105,
672,
325,
687
],
"score": 1.0,
"content": "which has the same marginal distribution at each time",
"type": "text"
},
{
"bbox": [
325,
675,
330,
683
],
"score": 0.8,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
330,
672,
507,
687
],
"score": 1.0,
"content": "as that of the SDE. Specifically, for DPMs,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
684,
375,
698
],
"spans": [
{
"bbox": [
106,
684,
375,
698
],
"score": 1.0,
"content": "Song et al. [3] proved that the probability flow ODE of Eq. (2.4) is",
"type": "text"
}
],
"index": 40
}
],
"index": 38
},
{
"type": "interline_equation",
"bbox": [
189,
702,
421,
726
],
"lines": [
{
"bbox": [
189,
702,
421,
726
],
"spans": [
{
"bbox": [
189,
702,
421,
726
],
"score": 0.92,
"content": "\\frac { \\mathrm { d } \\pmb { x } _ { t } } { \\mathrm { d } t } = f ( t ) \\pmb { x } _ { t } - \\frac { 1 } { 2 } g ^ { 2 } ( t ) \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } ) , \\quad \\pmb { x } _ { T } \\sim q _ { T } ( \\pmb { x } _ { T } ) ,",
"type": "interline_equation",
"image_path": "db0329942038b0f1cd9de0b5b9716a202705a11feeab57d38ae6c43085d15745.jpg"
}
]
}
],
"index": 41,
"virtual_lines": [
{
"bbox": [
189,
702,
421,
726
],
"spans": [],
"index": 41
}
]
}
],
"page_idx": 2,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
301,
740,
310,
752
],
"spans": [
{
"bbox": [
301,
740,
310,
752
],
"score": 1.0,
"content": "3",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
72,
289,
84
],
"lines": [
{
"bbox": [
106,
72,
289,
86
],
"spans": [
{
"bbox": [
106,
72,
289,
86
],
"score": 1.0,
"content": "2.1 Forward Process and Diffusion SDEs",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
92,
505,
128
],
"lines": [
{
"bbox": [
105,
92,
505,
106
],
"spans": [
{
"bbox": [
105,
92,
204,
106
],
"score": 1.0,
"content": "Assume that we have a",
"type": "text"
},
{
"bbox": [
205,
94,
214,
103
],
"score": 0.82,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
214,
92,
339,
106
],
"score": 1.0,
"content": "-dimensional random variable",
"type": "text"
},
{
"bbox": [
339,
92,
379,
104
],
"score": 0.92,
"content": "\\pmb { x } _ { 0 } \\in \\mathbb { R } ^ { D }",
"type": "inline_equation"
},
{
"bbox": [
380,
92,
505,
106
],
"score": 1.0,
"content": "with an unknown distribution",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
107,
103,
505,
118
],
"spans": [
{
"bbox": [
107,
104,
135,
116
],
"score": 0.9,
"content": "q _ { 0 } ( { \\pmb x } _ { 0 } )",
"type": "inline_equation"
},
{
"bbox": [
135,
103,
437,
118
],
"score": 1.0,
"content": ". Diffusion Probabilistic Models (DPMs) [1–3, 10] define a forward process",
"type": "text"
},
{
"bbox": [
438,
104,
483,
117
],
"score": 0.93,
"content": "\\{ \\pmb { x } _ { t } \\} _ { t \\in [ 0 , T ] }",
"type": "inline_equation"
},
{
"bbox": [
484,
103,
505,
118
],
"score": 1.0,
"content": "with",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
107,
114,
506,
130
],
"spans": [
{
"bbox": [
107,
117,
133,
127
],
"score": 0.88,
"content": "T > 0",
"type": "inline_equation"
},
{
"bbox": [
134,
114,
186,
130
],
"score": 1.0,
"content": "starting with",
"type": "text"
},
{
"bbox": [
186,
118,
198,
127
],
"score": 0.85,
"content": "\\scriptstyle { \\mathbf { { \\mathit { x } } } } _ { 0 }",
"type": "inline_equation"
},
{
"bbox": [
198,
114,
270,
130
],
"score": 1.0,
"content": ", such that for any",
"type": "text"
},
{
"bbox": [
270,
116,
308,
128
],
"score": 0.93,
"content": "t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
309,
114,
385,
130
],
"score": 1.0,
"content": ", the distribution of",
"type": "text"
},
{
"bbox": [
385,
117,
396,
127
],
"score": 0.86,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
396,
114,
458,
130
],
"score": 1.0,
"content": "conditioned on",
"type": "text"
},
{
"bbox": [
458,
118,
470,
127
],
"score": 0.86,
"content": "\\scriptstyle { \\mathbf { { \\mathit { x } } } } _ { 0 }",
"type": "inline_equation"
},
{
"bbox": [
470,
114,
506,
130
],
"score": 1.0,
"content": "satisfies",
"type": "text"
}
],
"index": 3
}
],
"index": 2,
"bbox_fs": [
105,
92,
506,
130
]
},
{
"type": "interline_equation",
"bbox": [
228,
133,
381,
149
],
"lines": [
{
"bbox": [
228,
133,
381,
149
],
"spans": [
{
"bbox": [
228,
133,
381,
149
],
"score": 0.92,
"content": "\\begin{array} { r } { q _ { 0 t } ( \\pmb { x } _ { t } | \\pmb { x } _ { 0 } ) = \\mathcal { N } ( \\pmb { x } _ { t } | \\alpha ( t ) \\pmb { x } _ { 0 } , \\sigma ^ { 2 } ( t ) \\pmb { I } ) , } \\end{array}",
"type": "interline_equation",
"image_path": "b98f04c1bbffa9531e257210ae6112c502489822e557b5a3afdb274392b83d54.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
228,
133,
381,
149
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
153,
505,
221
],
"lines": [
{
"bbox": [
106,
153,
505,
166
],
"spans": [
{
"bbox": [
106,
153,
134,
166
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
154,
202,
166
],
"score": 0.92,
"content": "\\alpha ( t ) , \\sigma ( t ) \\in \\mathbb { R } ^ { + }",
"type": "inline_equation"
},
{
"bbox": [
202,
153,
328,
166
],
"score": 1.0,
"content": "are differentiable functions of",
"type": "text"
},
{
"bbox": [
329,
155,
334,
164
],
"score": 0.81,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
334,
153,
505,
166
],
"score": 1.0,
"content": "with bounded derivatives, and we denote",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
165,
506,
178
],
"spans": [
{
"bbox": [
105,
165,
141,
178
],
"score": 1.0,
"content": "them as",
"type": "text"
},
{
"bbox": [
142,
168,
166,
177
],
"score": 0.88,
"content": "\\alpha _ { t } , \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
167,
165,
295,
178
],
"score": 1.0,
"content": "for simplicity. The choice for",
"type": "text"
},
{
"bbox": [
295,
167,
307,
176
],
"score": 0.86,
"content": "\\alpha _ { t }",
"type": "inline_equation"
},
{
"bbox": [
307,
165,
326,
178
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
326,
167,
337,
176
],
"score": 0.86,
"content": "\\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
337,
165,
506,
178
],
"score": 1.0,
"content": "is referred to as the noise schedule of a",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
175,
505,
188
],
"spans": [
{
"bbox": [
105,
175,
150,
188
],
"score": 1.0,
"content": "DPM. Let",
"type": "text"
},
{
"bbox": [
151,
177,
177,
187
],
"score": 0.9,
"content": "q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
177,
175,
323,
188
],
"score": 1.0,
"content": "denote the marginal distribution of",
"type": "text"
},
{
"bbox": [
324,
178,
334,
187
],
"score": 0.87,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
335,
175,
505,
188
],
"score": 1.0,
"content": ", DPMs choose noise schedules to ensure",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
186,
506,
200
],
"spans": [
{
"bbox": [
105,
186,
124,
200
],
"score": 1.0,
"content": "that",
"type": "text"
},
{
"bbox": [
124,
187,
229,
199
],
"score": 0.89,
"content": "q _ { T } ( \\pmb { x } _ { T } ) \\overset { \\cdot } { \\approx } \\dot { \\mathcal { N } } ( \\pmb { x } _ { T } | \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\pmb { I } )",
"type": "inline_equation"
},
{
"bbox": [
229,
186,
267,
200
],
"score": 1.0,
"content": "for some",
"type": "text"
},
{
"bbox": [
267,
187,
293,
198
],
"score": 0.9,
"content": "\\tilde { \\sigma } > 0",
"type": "inline_equation"
},
{
"bbox": [
293,
186,
438,
200
],
"score": 1.0,
"content": ", and the signal-to-noise-ratio (SNR)",
"type": "text"
},
{
"bbox": [
438,
186,
465,
199
],
"score": 0.93,
"content": "\\alpha _ { t } ^ { 2 } / \\sigma _ { t } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
466,
186,
506,
200
],
"score": 1.0,
"content": "is strictly",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
198,
506,
211
],
"spans": [
{
"bbox": [
106,
198,
173,
211
],
"score": 1.0,
"content": "decreasing w.r.t.",
"type": "text"
},
{
"bbox": [
174,
200,
179,
208
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
179,
198,
506,
211
],
"score": 1.0,
"content": "[10]. Moreover, Kingma et al. [10] prove that the following stochastic differential",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
209,
492,
221
],
"spans": [
{
"bbox": [
105,
209,
313,
221
],
"score": 1.0,
"content": "equation (SDE) has the same transition distribution",
"type": "text"
},
{
"bbox": [
313,
209,
357,
221
],
"score": 0.93,
"content": "q _ { 0 t } ( \\pmb { x } _ { t } | \\pmb { x } _ { 0 } )",
"type": "inline_equation"
},
{
"bbox": [
358,
209,
448,
221
],
"score": 1.0,
"content": "as in Eq. (2.1) for any",
"type": "text"
},
{
"bbox": [
449,
209,
487,
221
],
"score": 0.93,
"content": "t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
488,
209,
492,
221
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 10
}
],
"index": 7.5,
"bbox_fs": [
105,
153,
506,
221
]
},
{
"type": "interline_equation",
"bbox": [
213,
226,
397,
240
],
"lines": [
{
"bbox": [
213,
226,
397,
240
],
"spans": [
{
"bbox": [
213,
226,
397,
240
],
"score": 0.91,
"content": "\\mathrm { d } \\pmb { x } _ { t } = f ( t ) \\pmb { x } _ { t } \\mathrm { d } t + g ( t ) \\mathrm { d } \\pmb { w } _ { t } , \\quad \\pmb { x } _ { 0 } \\sim q _ { 0 } ( \\pmb { x } _ { 0 } ) ,",
"type": "interline_equation",
"image_path": "8433318473b1628737ca5404d0a929d6a4570d9919cf8824f87c274bc4a0bbfd.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
213,
226,
397,
240
],
"spans": [],
"index": 11
}
]
},
{
"type": "text",
"bbox": [
106,
245,
317,
257
],
"lines": [
{
"bbox": [
105,
243,
318,
260
],
"spans": [
{
"bbox": [
105,
243,
133,
260
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
245,
172,
257
],
"score": 0.94,
"content": "{ \\pmb w } _ { t } \\in \\mathbb { R } ^ { D }",
"type": "inline_equation"
},
{
"bbox": [
172,
243,
318,
260
],
"score": 1.0,
"content": "is the standard Wiener process, and",
"type": "text"
}
],
"index": 12
}
],
"index": 12,
"bbox_fs": [
105,
243,
318,
260
]
},
{
"type": "interline_equation",
"bbox": [
206,
263,
405,
289
],
"lines": [
{
"bbox": [
206,
263,
405,
289
],
"spans": [
{
"bbox": [
206,
263,
405,
289
],
"score": 0.93,
"content": "f ( t ) = \\frac { \\mathrm { d } \\log \\alpha _ { t } } { \\mathrm { d } t } , \\quad g ^ { 2 } ( t ) = \\frac { \\mathrm { d } \\sigma _ { t } ^ { 2 } } { \\mathrm { d } t } - 2 \\frac { \\mathrm { d } \\log \\alpha _ { t } } { \\mathrm { d } t } \\sigma _ { t } ^ { 2 } .",
"type": "interline_equation",
"image_path": "e5d12cd513ddf8ae3fa219a1feddb4594e4cf22e33260b27389e15d81ab87dc7.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
206,
263,
405,
289
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
107,
294,
506,
317
],
"lines": [
{
"bbox": [
106,
294,
505,
307
],
"spans": [
{
"bbox": [
106,
294,
505,
307
],
"score": 1.0,
"content": "Under some regularity conditions, Song et al. [3] show that the forward process in Eq. (2.2) has an",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
304,
476,
318
],
"spans": [
{
"bbox": [
105,
304,
255,
318
],
"score": 1.0,
"content": "equivalent reverse process from time",
"type": "text"
},
{
"bbox": [
256,
306,
264,
315
],
"score": 0.84,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
264,
304,
276,
318
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
276,
306,
282,
315
],
"score": 0.29,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
282,
304,
439,
318
],
"score": 1.0,
"content": ", starting with the marginal distribution",
"type": "text"
},
{
"bbox": [
439,
306,
471,
317
],
"score": 0.93,
"content": "q _ { T } ( { \\pmb x } _ { T } )",
"type": "inline_equation"
},
{
"bbox": [
471,
304,
476,
318
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5,
"bbox_fs": [
105,
294,
505,
318
]
},
{
"type": "interline_equation",
"bbox": [
163,
322,
446,
338
],
"lines": [
{
"bbox": [
163,
322,
446,
338
],
"spans": [
{
"bbox": [
163,
322,
446,
338
],
"score": 0.88,
"content": "\\mathrm { d } \\pmb { x } _ { t } = [ f ( t ) \\pmb { x } _ { t } - g ^ { 2 } ( t ) \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } ) ] \\mathrm { d } t + g ( t ) \\mathrm { d } \\bar { \\pmb { w } } _ { t } , \\quad \\pmb { x } _ { T } \\sim q _ { T } ( \\pmb { x } _ { T } ) ,",
"type": "interline_equation",
"image_path": "098b1c09edd91e1e8b539459bb92a3c5f191abb42aa1157755522d3dce62bcae.jpg"
}
]
}
],
"index": 16,
"virtual_lines": [
{
"bbox": [
163,
322,
446,
338
],
"spans": [],
"index": 16
}
]
},
{
"type": "text",
"bbox": [
108,
343,
505,
388
],
"lines": [
{
"bbox": [
106,
343,
505,
355
],
"spans": [
{
"bbox": [
106,
343,
133,
355
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
344,
146,
354
],
"score": 0.88,
"content": "\\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { \\nabla } } \\bar { \\mathbf { w } } _ { t }",
"type": "inline_equation"
},
{
"bbox": [
147,
343,
505,
355
],
"score": 1.0,
"content": "is a standard Wiener process in the reverse time. The only unknown term in Eq. (2.4) is",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
354,
505,
367
],
"spans": [
{
"bbox": [
105,
354,
182,
367
],
"score": 1.0,
"content": "the score function",
"type": "text"
},
{
"bbox": [
183,
354,
239,
366
],
"score": 0.92,
"content": "\\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
239,
354,
293,
367
],
"score": 1.0,
"content": "at each time",
"type": "text"
},
{
"bbox": [
293,
356,
298,
364
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
299,
354,
469,
367
],
"score": 1.0,
"content": ". In practice, DPMs use a neural network",
"type": "text"
},
{
"bbox": [
469,
355,
505,
366
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
}
],
"index": 18
},
{
"bbox": [
105,
365,
506,
377
],
"spans": [
{
"bbox": [
105,
365,
180,
377
],
"score": 1.0,
"content": "parameterized by",
"type": "text"
},
{
"bbox": [
181,
366,
186,
375
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
187,
365,
345,
377
],
"score": 1.0,
"content": "to estimate the scaled score function:",
"type": "text"
},
{
"bbox": [
345,
365,
419,
377
],
"score": 0.92,
"content": "- \\sigma _ { t } \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
420,
365,
488,
377
],
"score": 1.0,
"content": ". The parameter",
"type": "text"
},
{
"bbox": [
488,
366,
494,
375
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
495,
365,
506,
377
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
376,
331,
389
],
"spans": [
{
"bbox": [
105,
376,
331,
389
],
"score": 1.0,
"content": "optimized by minimizing the following objective [2, 3]:",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5,
"bbox_fs": [
105,
343,
506,
389
]
},
{
"type": "interline_equation",
"bbox": [
165,
393,
445,
453
],
"lines": [
{
"bbox": [
165,
393,
445,
453
],
"spans": [
{
"bbox": [
165,
393,
445,
453
],
"score": 0.94,
"content": "\\begin{array} { l } { \\displaystyle \\mathcal { L } ( \\theta ; \\omega ( t ) ) : = \\frac { 1 } { 2 } \\int _ { 0 } ^ { T } \\omega ( t ) \\mathbb { E } _ { q _ { t } ( \\mathbf { \\Delta x } _ { t } ) } \\Big [ \\| \\epsilon _ { \\theta } ( \\mathbf { x } _ { t } , t ) + \\sigma _ { t } \\nabla _ { \\mathbf { x } } \\log q _ { t } ( \\mathbf { \\Delta x } _ { t } ) \\| _ { 2 } ^ { 2 } \\Big ] \\mathrm { d } t } \\\\ { \\displaystyle \\qquad = \\frac { 1 } { 2 } \\int _ { 0 } ^ { T } \\omega ( t ) \\mathbb { E } _ { q _ { 0 } ( \\mathbf { x } _ { 0 } ) } \\mathbb { E } _ { q ( \\epsilon ) } \\Big [ \\| \\epsilon _ { \\theta } ( \\mathbf { x } _ { t } , t ) - \\epsilon \\| _ { 2 } ^ { 2 } \\Big ] \\mathrm { d } t + C , } \\end{array}",
"type": "interline_equation",
"image_path": "18da74f3ada9dcef2cf0de53303bacedefeb327430ef720bfe85381fb0de2f45.jpg"
}
]
}
],
"index": 22,
"virtual_lines": [
{
"bbox": [
165,
393,
445,
413.0
],
"spans": [],
"index": 21
},
{
"bbox": [
165,
413.0,
445,
433.0
],
"spans": [],
"index": 22
},
{
"bbox": [
165,
433.0,
445,
453.0
],
"spans": [],
"index": 23
}
]
},
{
"type": "text",
"bbox": [
106,
456,
506,
513
],
"lines": [
{
"bbox": [
106,
456,
506,
470
],
"spans": [
{
"bbox": [
106,
456,
133,
470
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
457,
153,
469
],
"score": 0.92,
"content": "\\omega ( t )",
"type": "inline_equation"
},
{
"bbox": [
153,
456,
253,
470
],
"score": 1.0,
"content": "is a weighting function,",
"type": "text"
},
{
"bbox": [
253,
456,
345,
469
],
"score": 0.88,
"content": "\\epsilon \\sim q ( \\epsilon ) = \\mathcal { N } ( \\epsilon | \\mathbf { 0 } , I )",
"type": "inline_equation"
},
{
"bbox": [
345,
456,
348,
470
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
349,
458,
421,
468
],
"score": 0.88,
"content": "{ \\pmb x } _ { t } = \\alpha _ { t } { \\pmb x } _ { 0 } + \\sigma _ { t } { \\pmb \\epsilon }",
"type": "inline_equation"
},
{
"bbox": [
421,
456,
442,
470
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
442,
457,
451,
467
],
"score": 0.86,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
451,
456,
506,
470
],
"score": 1.0,
"content": "is a constant",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
468,
506,
480
],
"spans": [
{
"bbox": [
106,
468,
169,
480
],
"score": 1.0,
"content": "independent of",
"type": "text"
},
{
"bbox": [
169,
469,
174,
478
],
"score": 0.82,
"content": "\\theta",
"type": "inline_equation"
},
{
"bbox": [
175,
468,
192,
480
],
"score": 1.0,
"content": ". As",
"type": "text"
},
{
"bbox": [
193,
468,
229,
480
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
229,
468,
482,
480
],
"score": 1.0,
"content": "can also be regarded as predicting the Gaussian noise added to",
"type": "text"
},
{
"bbox": [
482,
469,
493,
479
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
493,
468,
506,
480
],
"score": 1.0,
"content": ", it",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
477,
507,
493
],
"spans": [
{
"bbox": [
104,
477,
381,
493
],
"score": 1.0,
"content": "is usually called the noise prediction model. Since the ground truth of",
"type": "text"
},
{
"bbox": [
382,
479,
418,
491
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
418,
477,
429,
493
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
429,
479,
503,
491
],
"score": 0.89,
"content": "- \\sigma _ { t } \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
503,
477,
507,
493
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
488,
506,
503
],
"spans": [
{
"bbox": [
105,
488,
303,
503
],
"score": 1.0,
"content": "DPMs replace the score function in Eq. (2.4) by",
"type": "text"
},
{
"bbox": [
303,
489,
361,
501
],
"score": 0.92,
"content": "- \\mathbf { \\epsilon } \\mathbf { \\epsilon } \\bar { \\mathbf { \\alpha } } ( \\mathbf { x } _ { t } , t ) / \\sigma _ { t }",
"type": "inline_equation"
},
{
"bbox": [
361,
488,
506,
503
],
"score": 1.0,
"content": "and define a parameterized reverse",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
500,
403,
513
],
"spans": [
{
"bbox": [
105,
500,
179,
513
],
"score": 1.0,
"content": "process (diffusion",
"type": "text"
},
{
"bbox": [
180,
501,
199,
511
],
"score": 0.45,
"content": "S D E",
"type": "inline_equation"
},
{
"bbox": [
200,
500,
246,
513
],
"score": 1.0,
"content": ") from time",
"type": "text"
},
{
"bbox": [
246,
502,
254,
510
],
"score": 0.85,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
254,
500,
266,
513
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
266,
502,
272,
510
],
"score": 0.38,
"content": "0",
"type": "inline_equation"
},
{
"bbox": [
272,
500,
328,
513
],
"score": 1.0,
"content": ", starting with",
"type": "text"
},
{
"bbox": [
329,
500,
399,
513
],
"score": 0.92,
"content": "\\pmb { x } _ { T } \\overset { \\cdot } { \\sim } \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\pmb { I } )",
"type": "inline_equation"
},
{
"bbox": [
399,
500,
403,
513
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 28
}
],
"index": 26,
"bbox_fs": [
104,
456,
507,
513
]
},
{
"type": "interline_equation",
"bbox": [
163,
518,
448,
546
],
"lines": [
{
"bbox": [
163,
518,
448,
546
],
"spans": [
{
"bbox": [
163,
518,
448,
546
],
"score": 0.94,
"content": "\\mathrm { d } x _ { t } = \\left[ f ( t ) x _ { t } + \\frac { g ^ { 2 } ( t ) } { \\sigma _ { t } } \\epsilon _ { \\theta } ( x _ { t } , t ) \\right] \\mathrm { d } t + g ( t ) \\mathrm { d } \\bar { w } _ { t } , \\quad x _ { T } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } I ) .",
"type": "interline_equation",
"image_path": "65fdde0d1d949853db59a598710255c20a35d6c68ca6401f8b46cd9f514525f2.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
163,
518,
448,
546
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
107,
551,
506,
607
],
"lines": [
{
"bbox": [
106,
551,
505,
564
],
"spans": [
{
"bbox": [
106,
551,
505,
564
],
"score": 1.0,
"content": "Samples can be generated from DPMs by solving the diffusion SDE in Eq. (2.5) with numerical",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
563,
505,
574
],
"spans": [
{
"bbox": [
106,
563,
268,
574
],
"score": 1.0,
"content": "solvers, which discretize the SDE from",
"type": "text"
},
{
"bbox": [
268,
563,
276,
572
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
277,
563,
505,
574
],
"score": 1.0,
"content": "to 0. Song et al. [3] proved that the traditional ancestral",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
573,
506,
586
],
"spans": [
{
"bbox": [
105,
573,
506,
586
],
"score": 1.0,
"content": "sampling method for DPMs [2] can be viewed as a first-order SDE solver for Eq. (2.5). However, these",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
583,
506,
597
],
"spans": [
{
"bbox": [
105,
583,
506,
597
],
"score": 1.0,
"content": "first-order methods usually need hundreds of or thousands of function evaluations to converge [3],",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
595,
279,
608
],
"spans": [
{
"bbox": [
106,
595,
279,
608
],
"score": 1.0,
"content": "leading to extremely slow sampling speed.",
"type": "text"
}
],
"index": 34
}
],
"index": 32,
"bbox_fs": [
105,
551,
506,
608
]
},
{
"type": "title",
"bbox": [
108,
619,
278,
632
],
"lines": [
{
"bbox": [
105,
618,
280,
635
],
"spans": [
{
"bbox": [
105,
618,
280,
635
],
"score": 1.0,
"content": "2.2 Diffusion (Probability Flow) ODEs",
"type": "text"
}
],
"index": 35
}
],
"index": 35
},
{
"type": "text",
"bbox": [
106,
640,
506,
696
],
"lines": [
{
"bbox": [
105,
640,
506,
654
],
"spans": [
{
"bbox": [
105,
640,
506,
654
],
"score": 1.0,
"content": "When discretizing SDEs, the step size is limited by the randomness of the Wiener process [27, Chap.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
651,
505,
665
],
"spans": [
{
"bbox": [
105,
651,
505,
665
],
"score": 1.0,
"content": "11]. A large step size (small number of steps) often causes non-convergence, especially in high",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
660,
507,
676
],
"spans": [
{
"bbox": [
105,
660,
507,
676
],
"score": 1.0,
"content": "dimensional spaces. For faster sampling, one can consider the associated probability flow ODE [3],",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
672,
507,
687
],
"spans": [
{
"bbox": [
105,
672,
325,
687
],
"score": 1.0,
"content": "which has the same marginal distribution at each time",
"type": "text"
},
{
"bbox": [
325,
675,
330,
683
],
"score": 0.8,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
330,
672,
507,
687
],
"score": 1.0,
"content": "as that of the SDE. Specifically, for DPMs,",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
684,
375,
698
],
"spans": [
{
"bbox": [
106,
684,
375,
698
],
"score": 1.0,
"content": "Song et al. [3] proved that the probability flow ODE of Eq. (2.4) is",
"type": "text"
}
],
"index": 40
}
],
"index": 38,
"bbox_fs": [
105,
640,
507,
698
]
},
{
"type": "interline_equation",
"bbox": [
189,
702,
421,
726
],
"lines": [
{
"bbox": [
189,
702,
421,
726
],
"spans": [
{
"bbox": [
189,
702,
421,
726
],
"score": 0.92,
"content": "\\frac { \\mathrm { d } \\pmb { x } _ { t } } { \\mathrm { d } t } = f ( t ) \\pmb { x } _ { t } - \\frac { 1 } { 2 } g ^ { 2 } ( t ) \\nabla _ { \\pmb { x } } \\log q _ { t } ( \\pmb { x } _ { t } ) , \\quad \\pmb { x } _ { T } \\sim q _ { T } ( \\pmb { x } _ { T } ) ,",
"type": "interline_equation",
"image_path": "db0329942038b0f1cd9de0b5b9716a202705a11feeab57d38ae6c43085d15745.jpg"
}
]
}
],
"index": 41,
"virtual_lines": [
{
"bbox": [
189,
702,
421,
726
],
"spans": [],
"index": 41
}
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
95
],
"lines": [
{
"bbox": [
106,
73,
505,
86
],
"spans": [
{
"bbox": [
106,
73,
245,
86
],
"score": 1.0,
"content": "where the marginal distribution of",
"type": "text"
},
{
"bbox": [
246,
74,
257,
84
],
"score": 0.86,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
257,
73,
286,
86
],
"score": 1.0,
"content": "is also",
"type": "text"
},
{
"bbox": [
286,
73,
313,
85
],
"score": 0.92,
"content": "q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
313,
73,
505,
86
],
"score": 1.0,
"content": ". By replacing the score function with the noise",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
478,
97
],
"spans": [
{
"bbox": [
105,
83,
448,
97
],
"score": 1.0,
"content": "prediction model, Song et al. [3] defined the following parameterized ODE (diffusion",
"type": "text"
},
{
"bbox": [
448,
84,
470,
94
],
"score": 0.41,
"content": "O D E",
"type": "inline_equation"
},
{
"bbox": [
470,
83,
478,
97
],
"score": 1.0,
"content": "):",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "interline_equation",
"bbox": [
170,
96,
441,
123
],
"lines": [
{
"bbox": [
170,
96,
441,
123
],
"spans": [
{
"bbox": [
170,
96,
441,
123
],
"score": 0.94,
"content": "\\frac { \\mathrm { d } \\pmb { x } _ { t } } { \\mathrm { d } t } = \\pmb { h } _ { \\theta } ( \\pmb { x } _ { t } , t ) : = f ( t ) \\pmb { x } _ { t } + \\frac { g ^ { 2 } ( t ) } { 2 \\sigma _ { t } } \\epsilon _ { \\theta } ( \\pmb { x } _ { t } , t ) , \\quad \\pmb { x } _ { T } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\mathbf { I } ) .",
"type": "interline_equation",
"image_path": "772c489895a83788baac27c0b2cc0da51d29a5ff2c72c378ea8301efa3cfe52d.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
170,
96,
441,
105.0
],
"spans": [],
"index": 2
},
{
"bbox": [
170,
105.0,
441,
114.0
],
"spans": [],
"index": 3
},
{
"bbox": [
170,
114.0,
441,
123.0
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
123,
506,
212
],
"lines": [
{
"bbox": [
106,
124,
506,
137
],
"spans": [
{
"bbox": [
106,
124,
299,
137
],
"score": 1.0,
"content": "Samples can be drawn by solving the ODE from",
"type": "text"
},
{
"bbox": [
299,
124,
308,
134
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
308,
124,
506,
137
],
"score": 1.0,
"content": "to 0. Comparing with SDEs, ODEs can be solved",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
135,
506,
147
],
"spans": [
{
"bbox": [
105,
135,
506,
147
],
"score": 1.0,
"content": "with larger step sizes as they have no randomness. Furthermore, we can take advantage of efficient",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
146,
505,
158
],
"spans": [
{
"bbox": [
105,
146,
505,
158
],
"score": 1.0,
"content": "numerical ODE solvers to accelerate the sampling. Song et al. [3] used the RK45 ODE solver [28]",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
156,
506,
169
],
"spans": [
{
"bbox": [
105,
156,
315,
169
],
"score": 1.0,
"content": "for the diffusion ODEs, which generates samples in",
"type": "text"
},
{
"bbox": [
315,
157,
337,
167
],
"score": 0.86,
"content": "\\sim 6 0",
"type": "inline_equation"
},
{
"bbox": [
338,
156,
506,
169
],
"score": 1.0,
"content": "function evaluations to reach comparable",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
167,
506,
182
],
"spans": [
{
"bbox": [
105,
167,
506,
182
],
"score": 1.0,
"content": "quality with a 1000-step SDE solver for Eq. (2.5) on the CIFAR-10 dataset [29]. However, existing",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
178,
506,
192
],
"spans": [
{
"bbox": [
105,
178,
457,
192
],
"score": 1.0,
"content": "general-purpose ODE solvers still cannot generate satisfactory samples in the few-step",
"type": "text"
},
{
"bbox": [
457,
179,
479,
189
],
"score": 0.8,
"content": "\\sim 1 0",
"type": "inline_equation"
},
{
"bbox": [
479,
178,
506,
192
],
"score": 1.0,
"content": "steps)",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
190,
506,
201
],
"spans": [
{
"bbox": [
106,
190,
506,
201
],
"score": 1.0,
"content": "sampling regime. To the best of our knowledge, there is still a lack of training-free samplers for",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
200,
491,
213
],
"spans": [
{
"bbox": [
106,
200,
491,
213
],
"score": 1.0,
"content": "DPMs in the few-step sampling regime, and the sampling speed of DPMs is still a critical issue.",
"type": "text"
}
],
"index": 12
}
],
"index": 8.5
},
{
"type": "title",
"bbox": [
106,
226,
353,
240
],
"lines": [
{
"bbox": [
104,
225,
354,
242
],
"spans": [
{
"bbox": [
104,
225,
354,
242
],
"score": 1.0,
"content": "3 Customized Fast Solvers for Diffusion ODEs",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
249,
505,
316
],
"lines": [
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "As highlighted in Sec. 2.2, discretizing SDEs is generally difficult in high dimensions [27, Chap. 11]",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
262,
505,
273
],
"spans": [
{
"bbox": [
106,
262,
505,
273
],
"score": 1.0,
"content": "and it is hard to converge within few steps. In contrast, ODEs are easier to solve, yielding a potential",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
272,
505,
284
],
"spans": [
{
"bbox": [
106,
272,
505,
284
],
"score": 1.0,
"content": "for fast samplers. However, as mentioned in Sec. 2.2, the general black-box ODE solver used in",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
283,
506,
295
],
"spans": [
{
"bbox": [
105,
283,
506,
295
],
"score": 1.0,
"content": "previous work [3] empirically fails to converge in few steps. This motivates us to design a dedicated",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
293,
505,
306
],
"score": 1.0,
"content": "solver for diffusion ODEs to enable fast and high-quality few-step sampling. We start with a detailed",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
305,
334,
316
],
"spans": [
{
"bbox": [
106,
305,
334,
316
],
"score": 1.0,
"content": "investigation of the specific structure of diffusion ODEs.",
"type": "text"
}
],
"index": 19
}
],
"index": 16.5
},
{
"type": "title",
"bbox": [
107,
327,
390,
340
],
"lines": [
{
"bbox": [
104,
327,
392,
342
],
"spans": [
{
"bbox": [
104,
327,
392,
342
],
"score": 1.0,
"content": "3.1 Simplified Formulation of Exact Solutions of Diffusion ODEs",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
108,
348,
504,
382
],
"lines": [
{
"bbox": [
106,
348,
504,
361
],
"spans": [
{
"bbox": [
106,
348,
338,
361
],
"score": 1.0,
"content": "The key insight of this work is that given an initial value",
"type": "text"
},
{
"bbox": [
338,
351,
351,
360
],
"score": 0.86,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
351,
348,
382,
361
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
383,
349,
407,
359
],
"score": 0.89,
"content": "s > 0",
"type": "inline_equation"
},
{
"bbox": [
407,
348,
461,
361
],
"score": 1.0,
"content": ", the solution",
"type": "text"
},
{
"bbox": [
461,
351,
473,
360
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
473,
348,
504,
361
],
"score": 1.0,
"content": "at each",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
359,
505,
371
],
"spans": [
{
"bbox": [
106,
359,
127,
371
],
"score": 1.0,
"content": "time",
"type": "text"
},
{
"bbox": [
127,
360,
151,
370
],
"score": 0.89,
"content": "t < s",
"type": "inline_equation"
},
{
"bbox": [
152,
359,
505,
371
],
"score": 1.0,
"content": "of diffusion ODEs in Eq. (2.7) can be simplified into a very special exact formulation",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
370,
264,
383
],
"spans": [
{
"bbox": [
106,
370,
264,
383
],
"score": 1.0,
"content": "which can be efficiently approximated.",
"type": "text"
}
],
"index": 23
}
],
"index": 22
},
{
"type": "text",
"bbox": [
106,
386,
505,
480
],
"lines": [
{
"bbox": [
105,
386,
505,
399
],
"spans": [
{
"bbox": [
105,
386,
317,
399
],
"score": 1.0,
"content": "Our first key observation is that a part of the solution",
"type": "text"
},
{
"bbox": [
317,
389,
328,
398
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
329,
386,
505,
399
],
"score": 1.0,
"content": "can be exactly computed by considering the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
104,
397,
507,
411
],
"spans": [
{
"bbox": [
104,
397,
507,
411
],
"score": 1.0,
"content": "particular structure of diffusion ODEs. The r.h.s. of diffusion ODEs in Eq. (2.7) consists of two parts:",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
408,
506,
426
],
"spans": [
{
"bbox": [
105,
409,
140,
426
],
"score": 1.0,
"content": "the part",
"type": "text"
},
{
"bbox": [
140,
411,
169,
424
],
"score": 0.93,
"content": "f ( t ) x _ { t }",
"type": "inline_equation"
},
{
"bbox": [
169,
409,
260,
426
],
"score": 1.0,
"content": "is a linear function of",
"type": "text"
},
{
"bbox": [
260,
414,
271,
423
],
"score": 0.85,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
272,
409,
350,
426
],
"score": 1.0,
"content": ", and the other part",
"type": "text"
},
{
"bbox": [
351,
408,
406,
426
],
"score": 0.93,
"content": "\\frac { g ^ { 2 } ( t ) } { 2 \\sigma _ { t } } \\epsilon _ { \\theta } ( \\pmb { x } _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
406,
409,
506,
426
],
"score": 1.0,
"content": "is generally a nonlinear",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
425,
505,
438
],
"spans": [
{
"bbox": [
106,
425,
151,
438
],
"score": 1.0,
"content": "function of",
"type": "text"
},
{
"bbox": [
152,
426,
163,
436
],
"score": 0.87,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
163,
425,
283,
438
],
"score": 1.0,
"content": "because of the neural network",
"type": "text"
},
{
"bbox": [
283,
425,
319,
437
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
320,
425,
505,
438
],
"score": 1.0,
"content": ". This type of ODE is referred to as semi-linear",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
435,
506,
448
],
"spans": [
{
"bbox": [
105,
435,
506,
448
],
"score": 1.0,
"content": "ODE. The black-box ODE solvers adopted by previous work [3] are ignorant of this semi-linear",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
446,
505,
459
],
"spans": [
{
"bbox": [
105,
447,
232,
459
],
"score": 1.0,
"content": "structure as they take the whole",
"type": "text"
},
{
"bbox": [
233,
446,
271,
459
],
"score": 0.93,
"content": "h _ { \\theta } ( x _ { t } , \\bar t ) ",
"type": "inline_equation"
},
{
"bbox": [
271,
447,
505,
459
],
"score": 1.0,
"content": "in Eq. (2.7) as the input, which causes discretization errors",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
458,
505,
469
],
"spans": [
{
"bbox": [
106,
458,
482,
469
],
"score": 1.0,
"content": "of both the linear and nonlinear term. We note that for semi-linear ODEs, the solution at time",
"type": "text"
},
{
"bbox": [
483,
459,
488,
467
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
488,
458,
505,
469
],
"score": 1.0,
"content": "can",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
469,
379,
480
],
"spans": [
{
"bbox": [
106,
469,
379,
480
],
"score": 1.0,
"content": "be exactly formulated by the “variation of constants” formula [30]:",
"type": "text"
}
],
"index": 31
}
],
"index": 27.5
},
{
"type": "interline_equation",
"bbox": [
188,
481,
423,
510
],
"lines": [
{
"bbox": [
188,
481,
423,
510
],
"spans": [
{
"bbox": [
188,
481,
423,
510
],
"score": 0.93,
"content": "\\pmb { x } _ { t } = e ^ { \\int _ { s } ^ { t } f ( \\tau ) \\mathrm { d } \\tau } \\pmb { x } _ { s } + \\int _ { s } ^ { t } \\left( e ^ { \\int _ { \\tau } ^ { t } f ( r ) \\mathrm { d } r } \\frac { g ^ { 2 } \\big ( \\tau \\big ) } { 2 \\sigma _ { \\tau } } \\epsilon _ { \\theta } ( \\pmb { x } _ { \\tau } , \\tau ) \\right) \\mathrm { d } \\tau .",
"type": "interline_equation",
"image_path": "6f1153f8f759866e204632c40d87e1d736b7589259f6710fccd3282a51e399e5.jpg"
}
]
}
],
"index": 32.5,
"virtual_lines": [
{
"bbox": [
188,
481,
423,
495.5
],
"spans": [],
"index": 32
},
{
"bbox": [
188,
495.5,
423,
510.0
],
"spans": [],
"index": 33
}
]
},
{
"type": "text",
"bbox": [
106,
510,
505,
565
],
"lines": [
{
"bbox": [
106,
510,
505,
523
],
"spans": [
{
"bbox": [
106,
510,
505,
523
],
"score": 1.0,
"content": "This formulation decouples the linear part and the nonlinear part. In contrast to black-box ODE",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
522,
504,
532
],
"spans": [
{
"bbox": [
106,
522,
504,
532
],
"score": 1.0,
"content": "solvers, the linear part is now exactly computed, which eliminates the approximation error of the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
532,
504,
544
],
"spans": [
{
"bbox": [
106,
532,
504,
544
],
"score": 1.0,
"content": "linear term. However, the integral of the nonlinear part is still complicated because it couples the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
542,
506,
556
],
"spans": [
{
"bbox": [
105,
542,
273,
556
],
"score": 1.0,
"content": "coefficients about the noise schedule (i.e.,",
"type": "text"
},
{
"bbox": [
273,
543,
334,
555
],
"score": 0.95,
"content": "f ( \\tau ) , g ( \\tau ) \\bar { , } \\sigma _ { \\tau } )",
"type": "inline_equation"
},
{
"bbox": [
335,
542,
464,
556
],
"score": 1.0,
"content": "and the complex neural network",
"type": "text"
},
{
"bbox": [
464,
544,
474,
554
],
"score": 0.85,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
475,
542,
506,
556
],
"score": 1.0,
"content": ", which",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
554,
217,
567
],
"spans": [
{
"bbox": [
105,
554,
217,
567
],
"score": 1.0,
"content": "is still hard to approximate.",
"type": "text"
}
],
"index": 38
}
],
"index": 36
},
{
"type": "text",
"bbox": [
106,
570,
505,
614
],
"lines": [
{
"bbox": [
105,
569,
505,
583
],
"spans": [
{
"bbox": [
105,
569,
505,
583
],
"score": 1.0,
"content": "Our second key observation is that the integral of the nonlinear part can be greatly simplified by",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
581,
505,
594
],
"spans": [
{
"bbox": [
105,
581,
246,
594
],
"score": 1.0,
"content": "introducing a special variable. Let",
"type": "text"
},
{
"bbox": [
247,
581,
317,
593
],
"score": 0.92,
"content": "\\lambda _ { t } : = \\log \\bar { ( \\alpha _ { t } / \\sigma _ { t } ) }",
"type": "inline_equation"
},
{
"bbox": [
317,
581,
447,
594
],
"score": 1.0,
"content": "(one half of the log-SNR), then",
"type": "text"
},
{
"bbox": [
447,
582,
457,
592
],
"score": 0.89,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
458,
581,
505,
594
],
"score": 1.0,
"content": "is a strictly",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
591,
505,
605
],
"spans": [
{
"bbox": [
105,
591,
196,
605
],
"score": 1.0,
"content": "decreasing function of",
"type": "text"
},
{
"bbox": [
197,
593,
201,
602
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
202,
591,
487,
605
],
"score": 1.0,
"content": "(due to the definition of DPMs as discussed in Sec. 2.1). We can rewrite",
"type": "text"
},
{
"bbox": [
488,
592,
505,
604
],
"score": 0.91,
"content": "g ( t )",
"type": "inline_equation"
}
],
"index": 41
},
{
"bbox": [
105,
603,
166,
616
],
"spans": [
{
"bbox": [
105,
603,
166,
616
],
"score": 1.0,
"content": "in Eq. (2.3) as",
"type": "text"
}
],
"index": 42
}
],
"index": 40.5
},
{
"type": "interline_equation",
"bbox": [
156,
615,
455,
644
],
"lines": [
{
"bbox": [
156,
615,
455,
644
],
"spans": [
{
"bbox": [
156,
615,
455,
644
],
"score": 0.92,
"content": "g ^ { 2 } ( t ) = \\frac { \\mathrm { d } \\sigma _ { t } ^ { 2 } } { \\mathrm { d } t } - 2 \\frac { \\mathrm { d } \\log { \\alpha _ { t } } } { \\mathrm { d } t } \\sigma _ { t } ^ { 2 } = 2 \\sigma _ { t } ^ { 2 } \\left( \\frac { \\mathrm { d } \\log { \\sigma _ { t } } } { \\mathrm { d } t } - \\frac { \\mathrm { d } \\log { \\alpha _ { t } } } { \\mathrm { d } t } \\right) = - 2 \\sigma _ { t } ^ { 2 } \\frac { \\mathrm { d } \\lambda _ { t } } { \\mathrm { d } t } .",
"type": "interline_equation",
"image_path": "28428e28211e5384d76d1e8330c965590495d098f46b2f50a448a40982e1a8f0.jpg"
}
]
}
],
"index": 44,
"virtual_lines": [
{
"bbox": [
156,
615,
455,
624.6666666666666
],
"spans": [],
"index": 43
},
{
"bbox": [
156,
624.6666666666666,
455,
634.3333333333333
],
"spans": [],
"index": 44
},
{
"bbox": [
156,
634.3333333333333,
455,
643.9999999999999
],
"spans": [],
"index": 45
}
]
},
{
"type": "text",
"bbox": [
105,
644,
414,
657
],
"lines": [
{
"bbox": [
106,
643,
413,
659
],
"spans": [
{
"bbox": [
106,
643,
174,
659
],
"score": 1.0,
"content": "Combining with",
"type": "text"
},
{
"bbox": [
174,
645,
252,
657
],
"score": 0.93,
"content": "f ( t ) = \\mathrm { d } \\log \\alpha _ { t } / \\mathrm { d } t",
"type": "inline_equation"
},
{
"bbox": [
252,
643,
413,
659
],
"score": 1.0,
"content": "in Eq. (2.3), we can rewrite Eq. (3.1) as",
"type": "text"
}
],
"index": 46
}
],
"index": 46
},
{
"type": "interline_equation",
"bbox": [
211,
658,
399,
686
],
"lines": [
{
"bbox": [
211,
658,
399,
686
],
"spans": [
{
"bbox": [
211,
658,
399,
686
],
"score": 0.94,
"content": "\\pmb { x } _ { t } = \\frac { \\alpha _ { t } } { \\alpha _ { s } } \\pmb { x } _ { s } - \\alpha _ { t } \\int _ { s } ^ { t } \\left( \\frac { \\mathrm { d } \\lambda _ { \\tau } } { \\mathrm { d } \\tau } \\right) \\frac { \\sigma _ { \\tau } } { \\alpha _ { \\tau } } \\pmb { \\epsilon } _ { \\theta } ( \\pmb { x } _ { \\tau } , \\tau ) \\mathrm { d } \\tau .",
"type": "interline_equation",
"image_path": "d804fb43f91c9fdbe50275bbfdd61568c9d23d13e7c4cea3f8e0751dee5a2bb6.jpg"
}
]
}
],
"index": 47.5,
"virtual_lines": [
{
"bbox": [
211,
658,
399,
672.0
],
"spans": [],
"index": 47
},
{
"bbox": [
211,
672.0,
399,
686.0
],
"spans": [],
"index": 48
}
]
},
{
"type": "text",
"bbox": [
106,
687,
505,
723
],
"lines": [
{
"bbox": [
105,
686,
505,
701
],
"spans": [
{
"bbox": [
105,
686,
121,
701
],
"score": 1.0,
"content": "As",
"type": "text"
},
{
"bbox": [
121,
687,
167,
699
],
"score": 0.91,
"content": "\\lambda ( t ) = \\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
168,
686,
318,
701
],
"score": 1.0,
"content": "is a strictly decreasing function of",
"type": "text"
},
{
"bbox": [
318,
689,
324,
698
],
"score": 0.78,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
324,
686,
439,
701
],
"score": 1.0,
"content": ", it has an inverse function",
"type": "text"
},
{
"bbox": [
439,
687,
460,
699
],
"score": 0.91,
"content": "t _ { \\lambda } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
461,
686,
505,
701
],
"score": 1.0,
"content": "satisfying",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
107,
695,
502,
715
],
"spans": [
{
"bbox": [
107,
699,
159,
711
],
"score": 0.91,
"content": "t = t _ { \\lambda } ( \\lambda ( t ) )",
"type": "inline_equation"
},
{
"bbox": [
159,
695,
313,
715
],
"score": 1.0,
"content": ". We further change the subscripts of",
"type": "text"
},
{
"bbox": [
313,
701,
321,
709
],
"score": 0.81,
"content": "_ { \\textbf { \\em x } }",
"type": "inline_equation"
},
{
"bbox": [
321,
695,
339,
715
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
340,
700,
350,
710
],
"score": 0.85,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
351,
695,
374,
715
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
374,
700,
379,
709
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
380,
695,
391,
715
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
392,
699,
398,
709
],
"score": 0.83,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
399,
695,
447,
715
],
"score": 1.0,
"content": "and denote",
"type": "text"
},
{
"bbox": [
447,
699,
502,
712
],
"score": 0.91,
"content": "\\hat { \\pmb x } _ { \\lambda } : = \\pmb x _ { t _ { \\lambda } ( \\lambda ) }",
"type": "inline_equation"
}
],
"index": 50
},
{
"bbox": [
106,
710,
492,
724
],
"spans": [
{
"bbox": [
106,
711,
232,
724
],
"score": 0.9,
"content": "\\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\hat { x } _ { \\lambda } , \\lambda ) : = \\epsilon _ { \\boldsymbol { \\theta } } ( x _ { t _ { \\lambda } ( \\lambda ) } , t _ { \\lambda } ( \\lambda ) )",
"type": "inline_equation"
},
{
"bbox": [
232,
710,
424,
724
],
"score": 1.0,
"content": ". Rewrite Eq. (3.3) by “change-of-variable” for",
"type": "text"
},
{
"bbox": [
424,
711,
431,
721
],
"score": 0.79,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
431,
710,
492,
724
],
"score": 1.0,
"content": ", then we have:",
"type": "text"
}
],
"index": 51
}
],
"index": 50
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
301,
741,
310,
752
],
"spans": [
{
"bbox": [
301,
741,
310,
752
],
"score": 1.0,
"content": "4",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
95
],
"lines": [
{
"bbox": [
106,
73,
505,
86
],
"spans": [
{
"bbox": [
106,
73,
245,
86
],
"score": 1.0,
"content": "where the marginal distribution of",
"type": "text"
},
{
"bbox": [
246,
74,
257,
84
],
"score": 0.86,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
257,
73,
286,
86
],
"score": 1.0,
"content": "is also",
"type": "text"
},
{
"bbox": [
286,
73,
313,
85
],
"score": 0.92,
"content": "q _ { t } ( \\pmb { x } _ { t } )",
"type": "inline_equation"
},
{
"bbox": [
313,
73,
505,
86
],
"score": 1.0,
"content": ". By replacing the score function with the noise",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
478,
97
],
"spans": [
{
"bbox": [
105,
83,
448,
97
],
"score": 1.0,
"content": "prediction model, Song et al. [3] defined the following parameterized ODE (diffusion",
"type": "text"
},
{
"bbox": [
448,
84,
470,
94
],
"score": 0.41,
"content": "O D E",
"type": "inline_equation"
},
{
"bbox": [
470,
83,
478,
97
],
"score": 1.0,
"content": "):",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
105,
73,
505,
97
]
},
{
"type": "interline_equation",
"bbox": [
170,
96,
441,
123
],
"lines": [
{
"bbox": [
170,
96,
441,
123
],
"spans": [
{
"bbox": [
170,
96,
441,
123
],
"score": 0.94,
"content": "\\frac { \\mathrm { d } \\pmb { x } _ { t } } { \\mathrm { d } t } = \\pmb { h } _ { \\theta } ( \\pmb { x } _ { t } , t ) : = f ( t ) \\pmb { x } _ { t } + \\frac { g ^ { 2 } ( t ) } { 2 \\sigma _ { t } } \\epsilon _ { \\theta } ( \\pmb { x } _ { t } , t ) , \\quad \\pmb { x } _ { T } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\tilde { \\sigma } ^ { 2 } \\mathbf { I } ) .",
"type": "interline_equation",
"image_path": "772c489895a83788baac27c0b2cc0da51d29a5ff2c72c378ea8301efa3cfe52d.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
170,
96,
441,
105.0
],
"spans": [],
"index": 2
},
{
"bbox": [
170,
105.0,
441,
114.0
],
"spans": [],
"index": 3
},
{
"bbox": [
170,
114.0,
441,
123.0
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
123,
506,
212
],
"lines": [
{
"bbox": [
106,
124,
506,
137
],
"spans": [
{
"bbox": [
106,
124,
299,
137
],
"score": 1.0,
"content": "Samples can be drawn by solving the ODE from",
"type": "text"
},
{
"bbox": [
299,
124,
308,
134
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
308,
124,
506,
137
],
"score": 1.0,
"content": "to 0. Comparing with SDEs, ODEs can be solved",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
135,
506,
147
],
"spans": [
{
"bbox": [
105,
135,
506,
147
],
"score": 1.0,
"content": "with larger step sizes as they have no randomness. Furthermore, we can take advantage of efficient",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
146,
505,
158
],
"spans": [
{
"bbox": [
105,
146,
505,
158
],
"score": 1.0,
"content": "numerical ODE solvers to accelerate the sampling. Song et al. [3] used the RK45 ODE solver [28]",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
156,
506,
169
],
"spans": [
{
"bbox": [
105,
156,
315,
169
],
"score": 1.0,
"content": "for the diffusion ODEs, which generates samples in",
"type": "text"
},
{
"bbox": [
315,
157,
337,
167
],
"score": 0.86,
"content": "\\sim 6 0",
"type": "inline_equation"
},
{
"bbox": [
338,
156,
506,
169
],
"score": 1.0,
"content": "function evaluations to reach comparable",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
167,
506,
182
],
"spans": [
{
"bbox": [
105,
167,
506,
182
],
"score": 1.0,
"content": "quality with a 1000-step SDE solver for Eq. (2.5) on the CIFAR-10 dataset [29]. However, existing",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
178,
506,
192
],
"spans": [
{
"bbox": [
105,
178,
457,
192
],
"score": 1.0,
"content": "general-purpose ODE solvers still cannot generate satisfactory samples in the few-step",
"type": "text"
},
{
"bbox": [
457,
179,
479,
189
],
"score": 0.8,
"content": "\\sim 1 0",
"type": "inline_equation"
},
{
"bbox": [
479,
178,
506,
192
],
"score": 1.0,
"content": "steps)",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
190,
506,
201
],
"spans": [
{
"bbox": [
106,
190,
506,
201
],
"score": 1.0,
"content": "sampling regime. To the best of our knowledge, there is still a lack of training-free samplers for",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
200,
491,
213
],
"spans": [
{
"bbox": [
106,
200,
491,
213
],
"score": 1.0,
"content": "DPMs in the few-step sampling regime, and the sampling speed of DPMs is still a critical issue.",
"type": "text"
}
],
"index": 12
}
],
"index": 8.5,
"bbox_fs": [
105,
124,
506,
213
]
},
{
"type": "title",
"bbox": [
106,
226,
353,
240
],
"lines": [
{
"bbox": [
104,
225,
354,
242
],
"spans": [
{
"bbox": [
104,
225,
354,
242
],
"score": 1.0,
"content": "3 Customized Fast Solvers for Diffusion ODEs",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
249,
505,
316
],
"lines": [
{
"bbox": [
105,
250,
505,
263
],
"spans": [
{
"bbox": [
105,
250,
505,
263
],
"score": 1.0,
"content": "As highlighted in Sec. 2.2, discretizing SDEs is generally difficult in high dimensions [27, Chap. 11]",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
262,
505,
273
],
"spans": [
{
"bbox": [
106,
262,
505,
273
],
"score": 1.0,
"content": "and it is hard to converge within few steps. In contrast, ODEs are easier to solve, yielding a potential",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
272,
505,
284
],
"spans": [
{
"bbox": [
106,
272,
505,
284
],
"score": 1.0,
"content": "for fast samplers. However, as mentioned in Sec. 2.2, the general black-box ODE solver used in",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
283,
506,
295
],
"spans": [
{
"bbox": [
105,
283,
506,
295
],
"score": 1.0,
"content": "previous work [3] empirically fails to converge in few steps. This motivates us to design a dedicated",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
293,
505,
306
],
"score": 1.0,
"content": "solver for diffusion ODEs to enable fast and high-quality few-step sampling. We start with a detailed",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
305,
334,
316
],
"spans": [
{
"bbox": [
106,
305,
334,
316
],
"score": 1.0,
"content": "investigation of the specific structure of diffusion ODEs.",
"type": "text"
}
],
"index": 19
}
],
"index": 16.5,
"bbox_fs": [
105,
250,
506,
316
]
},
{
"type": "title",
"bbox": [
107,
327,
390,
340
],
"lines": [
{
"bbox": [
104,
327,
392,
342
],
"spans": [
{
"bbox": [
104,
327,
392,
342
],
"score": 1.0,
"content": "3.1 Simplified Formulation of Exact Solutions of Diffusion ODEs",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
108,
348,
504,
382
],
"lines": [
{
"bbox": [
106,
348,
504,
361
],
"spans": [
{
"bbox": [
106,
348,
338,
361
],
"score": 1.0,
"content": "The key insight of this work is that given an initial value",
"type": "text"
},
{
"bbox": [
338,
351,
351,
360
],
"score": 0.86,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
351,
348,
382,
361
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
383,
349,
407,
359
],
"score": 0.89,
"content": "s > 0",
"type": "inline_equation"
},
{
"bbox": [
407,
348,
461,
361
],
"score": 1.0,
"content": ", the solution",
"type": "text"
},
{
"bbox": [
461,
351,
473,
360
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
473,
348,
504,
361
],
"score": 1.0,
"content": "at each",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
359,
505,
371
],
"spans": [
{
"bbox": [
106,
359,
127,
371
],
"score": 1.0,
"content": "time",
"type": "text"
},
{
"bbox": [
127,
360,
151,
370
],
"score": 0.89,
"content": "t < s",
"type": "inline_equation"
},
{
"bbox": [
152,
359,
505,
371
],
"score": 1.0,
"content": "of diffusion ODEs in Eq. (2.7) can be simplified into a very special exact formulation",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
370,
264,
383
],
"spans": [
{
"bbox": [
106,
370,
264,
383
],
"score": 1.0,
"content": "which can be efficiently approximated.",
"type": "text"
}
],
"index": 23
}
],
"index": 22,
"bbox_fs": [
106,
348,
505,
383
]
},
{
"type": "text",
"bbox": [
106,
386,
505,
480
],
"lines": [
{
"bbox": [
105,
386,
505,
399
],
"spans": [
{
"bbox": [
105,
386,
317,
399
],
"score": 1.0,
"content": "Our first key observation is that a part of the solution",
"type": "text"
},
{
"bbox": [
317,
389,
328,
398
],
"score": 0.84,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
329,
386,
505,
399
],
"score": 1.0,
"content": "can be exactly computed by considering the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
104,
397,
507,
411
],
"spans": [
{
"bbox": [
104,
397,
507,
411
],
"score": 1.0,
"content": "particular structure of diffusion ODEs. The r.h.s. of diffusion ODEs in Eq. (2.7) consists of two parts:",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
408,
506,
426
],
"spans": [
{
"bbox": [
105,
409,
140,
426
],
"score": 1.0,
"content": "the part",
"type": "text"
},
{
"bbox": [
140,
411,
169,
424
],
"score": 0.93,
"content": "f ( t ) x _ { t }",
"type": "inline_equation"
},
{
"bbox": [
169,
409,
260,
426
],
"score": 1.0,
"content": "is a linear function of",
"type": "text"
},
{
"bbox": [
260,
414,
271,
423
],
"score": 0.85,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
272,
409,
350,
426
],
"score": 1.0,
"content": ", and the other part",
"type": "text"
},
{
"bbox": [
351,
408,
406,
426
],
"score": 0.93,
"content": "\\frac { g ^ { 2 } ( t ) } { 2 \\sigma _ { t } } \\epsilon _ { \\theta } ( \\pmb { x } _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
406,
409,
506,
426
],
"score": 1.0,
"content": "is generally a nonlinear",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
425,
505,
438
],
"spans": [
{
"bbox": [
106,
425,
151,
438
],
"score": 1.0,
"content": "function of",
"type": "text"
},
{
"bbox": [
152,
426,
163,
436
],
"score": 0.87,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
163,
425,
283,
438
],
"score": 1.0,
"content": "because of the neural network",
"type": "text"
},
{
"bbox": [
283,
425,
319,
437
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
320,
425,
505,
438
],
"score": 1.0,
"content": ". This type of ODE is referred to as semi-linear",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
435,
506,
448
],
"spans": [
{
"bbox": [
105,
435,
506,
448
],
"score": 1.0,
"content": "ODE. The black-box ODE solvers adopted by previous work [3] are ignorant of this semi-linear",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
446,
505,
459
],
"spans": [
{
"bbox": [
105,
447,
232,
459
],
"score": 1.0,
"content": "structure as they take the whole",
"type": "text"
},
{
"bbox": [
233,
446,
271,
459
],
"score": 0.93,
"content": "h _ { \\theta } ( x _ { t } , \\bar t ) ",
"type": "inline_equation"
},
{
"bbox": [
271,
447,
505,
459
],
"score": 1.0,
"content": "in Eq. (2.7) as the input, which causes discretization errors",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
458,
505,
469
],
"spans": [
{
"bbox": [
106,
458,
482,
469
],
"score": 1.0,
"content": "of both the linear and nonlinear term. We note that for semi-linear ODEs, the solution at time",
"type": "text"
},
{
"bbox": [
483,
459,
488,
467
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
488,
458,
505,
469
],
"score": 1.0,
"content": "can",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
469,
379,
480
],
"spans": [
{
"bbox": [
106,
469,
379,
480
],
"score": 1.0,
"content": "be exactly formulated by the “variation of constants” formula [30]:",
"type": "text"
}
],
"index": 31
}
],
"index": 27.5,
"bbox_fs": [
104,
386,
507,
480
]
},
{
"type": "interline_equation",
"bbox": [
188,
481,
423,
510
],
"lines": [
{
"bbox": [
188,
481,
423,
510
],
"spans": [
{
"bbox": [
188,
481,
423,
510
],
"score": 0.93,
"content": "\\pmb { x } _ { t } = e ^ { \\int _ { s } ^ { t } f ( \\tau ) \\mathrm { d } \\tau } \\pmb { x } _ { s } + \\int _ { s } ^ { t } \\left( e ^ { \\int _ { \\tau } ^ { t } f ( r ) \\mathrm { d } r } \\frac { g ^ { 2 } \\big ( \\tau \\big ) } { 2 \\sigma _ { \\tau } } \\epsilon _ { \\theta } ( \\pmb { x } _ { \\tau } , \\tau ) \\right) \\mathrm { d } \\tau .",
"type": "interline_equation",
"image_path": "6f1153f8f759866e204632c40d87e1d736b7589259f6710fccd3282a51e399e5.jpg"
}
]
}
],
"index": 32.5,
"virtual_lines": [
{
"bbox": [
188,
481,
423,
495.5
],
"spans": [],
"index": 32
},
{
"bbox": [
188,
495.5,
423,
510.0
],
"spans": [],
"index": 33
}
]
},
{
"type": "text",
"bbox": [
106,
510,
505,
565
],
"lines": [
{
"bbox": [
106,
510,
505,
523
],
"spans": [
{
"bbox": [
106,
510,
505,
523
],
"score": 1.0,
"content": "This formulation decouples the linear part and the nonlinear part. In contrast to black-box ODE",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
522,
504,
532
],
"spans": [
{
"bbox": [
106,
522,
504,
532
],
"score": 1.0,
"content": "solvers, the linear part is now exactly computed, which eliminates the approximation error of the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
532,
504,
544
],
"spans": [
{
"bbox": [
106,
532,
504,
544
],
"score": 1.0,
"content": "linear term. However, the integral of the nonlinear part is still complicated because it couples the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
542,
506,
556
],
"spans": [
{
"bbox": [
105,
542,
273,
556
],
"score": 1.0,
"content": "coefficients about the noise schedule (i.e.,",
"type": "text"
},
{
"bbox": [
273,
543,
334,
555
],
"score": 0.95,
"content": "f ( \\tau ) , g ( \\tau ) \\bar { , } \\sigma _ { \\tau } )",
"type": "inline_equation"
},
{
"bbox": [
335,
542,
464,
556
],
"score": 1.0,
"content": "and the complex neural network",
"type": "text"
},
{
"bbox": [
464,
544,
474,
554
],
"score": 0.85,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
475,
542,
506,
556
],
"score": 1.0,
"content": ", which",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
554,
217,
567
],
"spans": [
{
"bbox": [
105,
554,
217,
567
],
"score": 1.0,
"content": "is still hard to approximate.",
"type": "text"
}
],
"index": 38
}
],
"index": 36,
"bbox_fs": [
105,
510,
506,
567
]
},
{
"type": "text",
"bbox": [
106,
570,
505,
614
],
"lines": [
{
"bbox": [
105,
569,
505,
583
],
"spans": [
{
"bbox": [
105,
569,
505,
583
],
"score": 1.0,
"content": "Our second key observation is that the integral of the nonlinear part can be greatly simplified by",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
581,
505,
594
],
"spans": [
{
"bbox": [
105,
581,
246,
594
],
"score": 1.0,
"content": "introducing a special variable. Let",
"type": "text"
},
{
"bbox": [
247,
581,
317,
593
],
"score": 0.92,
"content": "\\lambda _ { t } : = \\log \\bar { ( \\alpha _ { t } / \\sigma _ { t } ) }",
"type": "inline_equation"
},
{
"bbox": [
317,
581,
447,
594
],
"score": 1.0,
"content": "(one half of the log-SNR), then",
"type": "text"
},
{
"bbox": [
447,
582,
457,
592
],
"score": 0.89,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
458,
581,
505,
594
],
"score": 1.0,
"content": "is a strictly",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
591,
505,
605
],
"spans": [
{
"bbox": [
105,
591,
196,
605
],
"score": 1.0,
"content": "decreasing function of",
"type": "text"
},
{
"bbox": [
197,
593,
201,
602
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
202,
591,
487,
605
],
"score": 1.0,
"content": "(due to the definition of DPMs as discussed in Sec. 2.1). We can rewrite",
"type": "text"
},
{
"bbox": [
488,
592,
505,
604
],
"score": 0.91,
"content": "g ( t )",
"type": "inline_equation"
}
],
"index": 41
},
{
"bbox": [
105,
603,
166,
616
],
"spans": [
{
"bbox": [
105,
603,
166,
616
],
"score": 1.0,
"content": "in Eq. (2.3) as",
"type": "text"
}
],
"index": 42
}
],
"index": 40.5,
"bbox_fs": [
105,
569,
505,
616
]
},
{
"type": "interline_equation",
"bbox": [
156,
615,
455,
644
],
"lines": [
{
"bbox": [
156,
615,
455,
644
],
"spans": [
{
"bbox": [
156,
615,
455,
644
],
"score": 0.92,
"content": "g ^ { 2 } ( t ) = \\frac { \\mathrm { d } \\sigma _ { t } ^ { 2 } } { \\mathrm { d } t } - 2 \\frac { \\mathrm { d } \\log { \\alpha _ { t } } } { \\mathrm { d } t } \\sigma _ { t } ^ { 2 } = 2 \\sigma _ { t } ^ { 2 } \\left( \\frac { \\mathrm { d } \\log { \\sigma _ { t } } } { \\mathrm { d } t } - \\frac { \\mathrm { d } \\log { \\alpha _ { t } } } { \\mathrm { d } t } \\right) = - 2 \\sigma _ { t } ^ { 2 } \\frac { \\mathrm { d } \\lambda _ { t } } { \\mathrm { d } t } .",
"type": "interline_equation",
"image_path": "28428e28211e5384d76d1e8330c965590495d098f46b2f50a448a40982e1a8f0.jpg"
}
]
}
],
"index": 44,
"virtual_lines": [
{
"bbox": [
156,
615,
455,
624.6666666666666
],
"spans": [],
"index": 43
},
{
"bbox": [
156,
624.6666666666666,
455,
634.3333333333333
],
"spans": [],
"index": 44
},
{
"bbox": [
156,
634.3333333333333,
455,
643.9999999999999
],
"spans": [],
"index": 45
}
]
},
{
"type": "text",
"bbox": [
105,
644,
414,
657
],
"lines": [
{
"bbox": [
106,
643,
413,
659
],
"spans": [
{
"bbox": [
106,
643,
174,
659
],
"score": 1.0,
"content": "Combining with",
"type": "text"
},
{
"bbox": [
174,
645,
252,
657
],
"score": 0.93,
"content": "f ( t ) = \\mathrm { d } \\log \\alpha _ { t } / \\mathrm { d } t",
"type": "inline_equation"
},
{
"bbox": [
252,
643,
413,
659
],
"score": 1.0,
"content": "in Eq. (2.3), we can rewrite Eq. (3.1) as",
"type": "text"
}
],
"index": 46
}
],
"index": 46,
"bbox_fs": [
106,
643,
413,
659
]
},
{
"type": "interline_equation",
"bbox": [
211,
658,
399,
686
],
"lines": [
{
"bbox": [
211,
658,
399,
686
],
"spans": [
{
"bbox": [
211,
658,
399,
686
],
"score": 0.94,
"content": "\\pmb { x } _ { t } = \\frac { \\alpha _ { t } } { \\alpha _ { s } } \\pmb { x } _ { s } - \\alpha _ { t } \\int _ { s } ^ { t } \\left( \\frac { \\mathrm { d } \\lambda _ { \\tau } } { \\mathrm { d } \\tau } \\right) \\frac { \\sigma _ { \\tau } } { \\alpha _ { \\tau } } \\pmb { \\epsilon } _ { \\theta } ( \\pmb { x } _ { \\tau } , \\tau ) \\mathrm { d } \\tau .",
"type": "interline_equation",
"image_path": "d804fb43f91c9fdbe50275bbfdd61568c9d23d13e7c4cea3f8e0751dee5a2bb6.jpg"
}
]
}
],
"index": 47.5,
"virtual_lines": [
{
"bbox": [
211,
658,
399,
672.0
],
"spans": [],
"index": 47
},
{
"bbox": [
211,
672.0,
399,
686.0
],
"spans": [],
"index": 48
}
]
},
{
"type": "text",
"bbox": [
106,
687,
505,
723
],
"lines": [
{
"bbox": [
105,
686,
505,
701
],
"spans": [
{
"bbox": [
105,
686,
121,
701
],
"score": 1.0,
"content": "As",
"type": "text"
},
{
"bbox": [
121,
687,
167,
699
],
"score": 0.91,
"content": "\\lambda ( t ) = \\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
168,
686,
318,
701
],
"score": 1.0,
"content": "is a strictly decreasing function of",
"type": "text"
},
{
"bbox": [
318,
689,
324,
698
],
"score": 0.78,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
324,
686,
439,
701
],
"score": 1.0,
"content": ", it has an inverse function",
"type": "text"
},
{
"bbox": [
439,
687,
460,
699
],
"score": 0.91,
"content": "t _ { \\lambda } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
461,
686,
505,
701
],
"score": 1.0,
"content": "satisfying",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
107,
695,
502,
715
],
"spans": [
{
"bbox": [
107,
699,
159,
711
],
"score": 0.91,
"content": "t = t _ { \\lambda } ( \\lambda ( t ) )",
"type": "inline_equation"
},
{
"bbox": [
159,
695,
313,
715
],
"score": 1.0,
"content": ". We further change the subscripts of",
"type": "text"
},
{
"bbox": [
313,
701,
321,
709
],
"score": 0.81,
"content": "_ { \\textbf { \\em x } }",
"type": "inline_equation"
},
{
"bbox": [
321,
695,
339,
715
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
340,
700,
350,
710
],
"score": 0.85,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
351,
695,
374,
715
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
374,
700,
379,
709
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
380,
695,
391,
715
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
392,
699,
398,
709
],
"score": 0.83,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
399,
695,
447,
715
],
"score": 1.0,
"content": "and denote",
"type": "text"
},
{
"bbox": [
447,
699,
502,
712
],
"score": 0.91,
"content": "\\hat { \\pmb x } _ { \\lambda } : = \\pmb x _ { t _ { \\lambda } ( \\lambda ) }",
"type": "inline_equation"
}
],
"index": 50
},
{
"bbox": [
106,
710,
492,
724
],
"spans": [
{
"bbox": [
106,
711,
232,
724
],
"score": 0.9,
"content": "\\hat { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\hat { x } _ { \\lambda } , \\lambda ) : = \\epsilon _ { \\boldsymbol { \\theta } } ( x _ { t _ { \\lambda } ( \\lambda ) } , t _ { \\lambda } ( \\lambda ) )",
"type": "inline_equation"
},
{
"bbox": [
232,
710,
424,
724
],
"score": 1.0,
"content": ". Rewrite Eq. (3.3) by “change-of-variable” for",
"type": "text"
},
{
"bbox": [
424,
711,
431,
721
],
"score": 0.79,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
431,
710,
492,
724
],
"score": 1.0,
"content": ", then we have:",
"type": "text"
}
],
"index": 51
}
],
"index": 50,
"bbox_fs": [
105,
686,
505,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
96
],
"lines": [
{
"bbox": [
106,
72,
505,
85
],
"spans": [
{
"bbox": [
106,
72,
416,
85
],
"score": 1.0,
"content": "Proposition 3.1 (Exact solution of diffusion ODEs). Given an initial value",
"type": "text"
},
{
"bbox": [
417,
75,
429,
84
],
"score": 0.85,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
429,
72,
461,
85
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
461,
73,
486,
83
],
"score": 0.88,
"content": "s > 0",
"type": "inline_equation"
},
{
"bbox": [
486,
72,
505,
85
],
"score": 1.0,
"content": ", the",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
83,
356,
96
],
"spans": [
{
"bbox": [
106,
83,
141,
96
],
"score": 1.0,
"content": "solution",
"type": "text"
},
{
"bbox": [
141,
86,
152,
95
],
"score": 0.82,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
153,
83,
183,
96
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
184,
83,
220,
96
],
"score": 0.93,
"content": "t \\in [ 0 , s ]",
"type": "inline_equation"
},
{
"bbox": [
220,
83,
356,
96
],
"score": 1.0,
"content": "of diffusion ODEs in Eq. (2.7) is:",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "interline_equation",
"bbox": [
224,
99,
387,
129
],
"lines": [
{
"bbox": [
224,
99,
387,
129
],
"spans": [
{
"bbox": [
224,
99,
387,
129
],
"score": 0.95,
"content": "\\pmb { x } _ { t } = \\frac { \\alpha _ { t } } { \\alpha _ { s } } \\pmb { x } _ { s } - \\alpha _ { t } \\int _ { \\lambda _ { s } } ^ { \\lambda _ { t } } e ^ { - \\lambda } \\hat { \\pmb { \\epsilon } } _ { \\theta } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda .",
"type": "interline_equation",
"image_path": "9abcfd7effacf68fb53c21394f34e41f5f607171cf80a13282a4bf3971c34831.jpg"
}
]
}
],
"index": 2.5,
"virtual_lines": [
{
"bbox": [
224,
99,
387,
114.0
],
"spans": [],
"index": 2
},
{
"bbox": [
224,
114.0,
387,
129.0
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
106,
138,
505,
172
],
"lines": [
{
"bbox": [
104,
137,
506,
153
],
"spans": [
{
"bbox": [
104,
137,
191,
153
],
"score": 1.0,
"content": "We call the integral",
"type": "text"
},
{
"bbox": [
192,
138,
267,
151
],
"score": 0.93,
"content": "\\begin{array} { r } { \\int e ^ { - \\lambda } \\hat { \\epsilon } _ { \\theta } ( \\hat { \\pmb x } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
267,
137,
430,
153
],
"score": 1.0,
"content": "the exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
431,
139,
441,
150
],
"score": 0.87,
"content": "\\scriptstyle { \\hat { \\epsilon } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
441,
137,
506,
153
],
"score": 1.0,
"content": ", which is very",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
149,
506,
162
],
"spans": [
{
"bbox": [
105,
149,
506,
162
],
"score": 1.0,
"content": "special and highly related to the exponential integrators in the literature of ODE solvers [25]. To the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
160,
496,
173
],
"spans": [
{
"bbox": [
106,
160,
496,
173
],
"score": 1.0,
"content": "best of our knowledge, such formulation has not been revealed in prior work of diffusion models.",
"type": "text"
}
],
"index": 6
}
],
"index": 5
},
{
"type": "text",
"bbox": [
106,
177,
505,
232
],
"lines": [
{
"bbox": [
106,
176,
506,
190
],
"spans": [
{
"bbox": [
106,
176,
506,
190
],
"score": 1.0,
"content": "Eq. (3.4) provides a new perspective for approximating the solutions of diffusion ODEs. Specifically,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
188,
506,
201
],
"spans": [
{
"bbox": [
106,
188,
132,
201
],
"score": 1.0,
"content": "given",
"type": "text"
},
{
"bbox": [
132,
190,
144,
199
],
"score": 0.86,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
144,
188,
177,
201
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
178,
190,
183,
198
],
"score": 0.68,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
184,
188,
432,
201
],
"score": 1.0,
"content": ", According to Eq. (3.4), approximating the solution at time",
"type": "text"
},
{
"bbox": [
433,
189,
438,
198
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
438,
188,
506,
201
],
"score": 1.0,
"content": "is equivalent to",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
198,
505,
212
],
"spans": [
{
"bbox": [
106,
198,
360,
212
],
"score": 1.0,
"content": "directly approximating the exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
361,
199,
371,
210
],
"score": 0.88,
"content": "\\hat { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
371,
198,
395,
212
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
395,
199,
407,
210
],
"score": 0.88,
"content": "\\lambda _ { s }",
"type": "inline_equation"
},
{
"bbox": [
407,
198,
419,
212
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
419,
199,
429,
210
],
"score": 0.87,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
430,
198,
505,
212
],
"score": 1.0,
"content": ", which avoids the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
209,
506,
223
],
"spans": [
{
"bbox": [
105,
209,
506,
223
],
"score": 1.0,
"content": "error of the linear terms and is well-studied in the literature of exponential integrators [25, 31]. Based",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
220,
492,
234
],
"spans": [
{
"bbox": [
106,
220,
492,
234
],
"score": 1.0,
"content": "on this insight, we propose fast solvers for diffusion ODEs, as detailed in the following sections.",
"type": "text"
}
],
"index": 11
}
],
"index": 9
},
{
"type": "title",
"bbox": [
106,
244,
300,
257
],
"lines": [
{
"bbox": [
105,
244,
301,
259
],
"spans": [
{
"bbox": [
105,
244,
301,
259
],
"score": 1.0,
"content": "3.2 High-Order Solvers for Diffusion ODEs",
"type": "text"
}
],
"index": 12
}
],
"index": 12
},
{
"type": "text",
"bbox": [
106,
264,
505,
298
],
"lines": [
{
"bbox": [
105,
264,
505,
278
],
"spans": [
{
"bbox": [
105,
264,
505,
278
],
"score": 1.0,
"content": "In this section, we propose high-order solvers for diffusion ODEs with convergence order guarantee",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
276,
505,
288
],
"spans": [
{
"bbox": [
106,
276,
505,
288
],
"score": 1.0,
"content": "by leveraging our proposed solution formulation Eq. (3.4). The proposed solvers and analysis are",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
287,
467,
299
],
"spans": [
{
"bbox": [
105,
287,
467,
299
],
"score": 1.0,
"content": "highly motivated by the methods of exponential integrators [25, 31] in the ODE literature.",
"type": "text"
}
],
"index": 15
}
],
"index": 14
},
{
"type": "text",
"bbox": [
106,
302,
505,
348
],
"lines": [
{
"bbox": [
103,
300,
507,
318
],
"spans": [
{
"bbox": [
103,
300,
249,
318
],
"score": 1.0,
"content": "Specifically, given an initial value",
"type": "text"
},
{
"bbox": [
250,
305,
264,
314
],
"score": 0.85,
"content": "\\mathbf { \\nabla } _ { \\mathbf { x } _ { T } }",
"type": "inline_equation"
},
{
"bbox": [
264,
300,
297,
318
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
298,
304,
306,
313
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
307,
300,
326,
318
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
326,
303,
356,
314
],
"score": 0.89,
"content": "M + 1",
"type": "inline_equation"
},
{
"bbox": [
356,
300,
402,
318
],
"score": 1.0,
"content": "time steps",
"type": "text"
},
{
"bbox": [
403,
302,
434,
315
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
434,
300,
507,
318
],
"score": 1.0,
"content": "decreasing from",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
313,
505,
327
],
"spans": [
{
"bbox": [
106,
314,
136,
325
],
"score": 0.9,
"content": "t _ { 0 } = T",
"type": "inline_equation"
},
{
"bbox": [
137,
313,
147,
327
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
147,
315,
179,
325
],
"score": 0.9,
"content": "t _ { M } = 0",
"type": "inline_equation"
},
{
"bbox": [
180,
313,
199,
327
],
"score": 1.0,
"content": ". Let",
"type": "text"
},
{
"bbox": [
199,
315,
240,
325
],
"score": 0.89,
"content": "\\tilde { \\mathbf { x } } _ { t _ { 0 } } = \\mathbf { x } _ { T }",
"type": "inline_equation"
},
{
"bbox": [
240,
313,
419,
327
],
"score": 1.0,
"content": "be the initial value. The proposed solvers use",
"type": "text"
},
{
"bbox": [
419,
315,
431,
324
],
"score": 0.81,
"content": "M",
"type": "inline_equation"
},
{
"bbox": [
431,
313,
505,
327
],
"score": 1.0,
"content": "steps to iteratively",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
102,
320,
510,
342
],
"spans": [
{
"bbox": [
102,
320,
190,
342
],
"score": 1.0,
"content": "compute a sequence",
"type": "text"
},
{
"bbox": [
190,
325,
227,
337
],
"score": 0.92,
"content": "\\{ \\tilde { { \\pmb { x } } } _ { t _ { i } } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
228,
320,
416,
342
],
"score": 1.0,
"content": "to approximate the true solutions at time steps",
"type": "text"
},
{
"bbox": [
417,
324,
448,
337
],
"score": 0.92,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
448,
320,
510,
342
],
"score": 1.0,
"content": ". In particular,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
335,
348,
349
],
"spans": [
{
"bbox": [
105,
335,
165,
349
],
"score": 1.0,
"content": "the last iterate",
"type": "text"
},
{
"bbox": [
165,
336,
183,
348
],
"score": 0.91,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { M } }",
"type": "inline_equation"
},
{
"bbox": [
183,
335,
348,
349
],
"score": 1.0,
"content": "approximates the true solution at time 0.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
107,
352,
505,
386
],
"lines": [
{
"bbox": [
105,
352,
505,
365
],
"spans": [
{
"bbox": [
105,
352,
314,
365
],
"score": 1.0,
"content": "In order to reduce the approximation error between",
"type": "text"
},
{
"bbox": [
315,
353,
333,
364
],
"score": 0.91,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } }",
"type": "inline_equation"
},
{
"bbox": [
333,
352,
505,
365
],
"score": 1.0,
"content": "and the true solution at time 0, we need to",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
360,
503,
379
],
"spans": [
{
"bbox": [
104,
360,
264,
379
],
"score": 1.0,
"content": "reduce the approximation error for each",
"type": "text"
},
{
"bbox": [
264,
364,
279,
375
],
"score": 0.9,
"content": "\\tilde { \\mathbf { x } } _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
279,
360,
480,
379
],
"score": 1.0,
"content": "at every step [30]. Starting with the previous value",
"type": "text"
},
{
"bbox": [
480,
363,
503,
376
],
"score": 0.91,
"content": "\\tilde { \\pmb { x } } _ { t _ { i - 1 } }",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
105,
374,
447,
388
],
"spans": [
{
"bbox": [
105,
374,
136,
388
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
136,
375,
154,
386
],
"score": 0.89,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
155,
374,
323,
388
],
"score": 1.0,
"content": ", according to Eq. (3.4), the exact solution",
"type": "text"
},
{
"bbox": [
324,
376,
360,
387
],
"score": 0.92,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
360,
374,
392,
388
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
392,
375,
400,
385
],
"score": 0.85,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
400,
374,
447,
388
],
"score": 1.0,
"content": "is given by",
"type": "text"
}
],
"index": 22
}
],
"index": 21
},
{
"type": "interline_equation",
"bbox": [
194,
390,
416,
422
],
"lines": [
{
"bbox": [
194,
390,
416,
422
],
"spans": [
{
"bbox": [
194,
390,
416,
422
],
"score": 0.93,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\hat { \\pmb { \\epsilon } } _ { \\theta } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda .",
"type": "interline_equation",
"image_path": "e941fd95cbf3e1c496583b82dd623ff65158438a8e8529e43b6a17ad2855df95.jpg"
}
]
}
],
"index": 23.5,
"virtual_lines": [
{
"bbox": [
194,
390,
416,
406.0
],
"spans": [],
"index": 23
},
{
"bbox": [
194,
406.0,
416,
422.0
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
106,
425,
505,
478
],
"lines": [
{
"bbox": [
104,
423,
506,
441
],
"spans": [
{
"bbox": [
104,
423,
245,
441
],
"score": 1.0,
"content": "Therefore, to compute the value",
"type": "text"
},
{
"bbox": [
246,
426,
260,
438
],
"score": 0.9,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
260,
423,
342,
441
],
"score": 1.0,
"content": "for approximating",
"type": "text"
},
{
"bbox": [
342,
428,
379,
438
],
"score": 0.92,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
379,
423,
506,
441
],
"score": 1.0,
"content": ", we need to approximate the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
436,
505,
455
],
"spans": [
{
"bbox": [
104,
436,
241,
455
],
"score": 1.0,
"content": "exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
241,
441,
252,
452
],
"score": 0.87,
"content": "\\hat { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
252,
436,
273,
455
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
274,
441,
295,
453
],
"score": 0.92,
"content": "\\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
295,
436,
306,
455
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
307,
441,
320,
452
],
"score": 0.89,
"content": "\\lambda _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
320,
436,
356,
455
],
"score": 1.0,
"content": ". Denote",
"type": "text"
},
{
"bbox": [
357,
440,
424,
453
],
"score": 0.92,
"content": "h _ { i } : = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
425,
436,
444,
455
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
444,
437,
505,
453
],
"score": 0.91,
"content": "\\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { { \\mathbf x } } _ { \\lambda } , \\lambda ) \\mathrel { \\mathop : } =",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
104,
450,
507,
469
],
"spans": [
{
"bbox": [
104,
450,
507,
469
],
"score": 1.0,
"content": "dnϵˆθ(xˆλ,λ)dλn as the n-th order total derivative of ϵˆθ(xˆλ, λ) w.r.t. λ. For k ≥ 1, the (k − 1)-th order",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
464,
304,
479
],
"spans": [
{
"bbox": [
105,
464,
188,
479
],
"score": 1.0,
"content": "Taylor expansion of",
"type": "text"
},
{
"bbox": [
188,
465,
228,
478
],
"score": 0.93,
"content": "\\hat { \\epsilon } _ { \\theta } ( \\hat { \\pmb x } _ { \\lambda } , \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
228,
464,
252,
479
],
"score": 1.0,
"content": "w.r.t.",
"type": "text"
},
{
"bbox": [
252,
466,
259,
475
],
"score": 0.79,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
260,
464,
270,
479
],
"score": 1.0,
"content": "at",
"type": "text"
},
{
"bbox": [
271,
466,
292,
478
],
"score": 0.92,
"content": "\\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
292,
464,
304,
479
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5
},
{
"type": "interline_equation",
"bbox": [
162,
482,
448,
516
],
"lines": [
{
"bbox": [
162,
482,
448,
516
],
"spans": [
{
"bbox": [
162,
482,
448,
516
],
"score": 0.94,
"content": "\\hat { \\epsilon } _ { \\theta } ( \\hat { x } _ { \\lambda } , \\lambda ) = \\sum _ { n = 0 } ^ { k - 1 } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { x } _ { \\lambda _ { t _ { i - 1 } } } , \\lambda _ { t _ { i - 1 } } ) + \\mathcal { O } ( ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { k } ) ,",
"type": "interline_equation",
"image_path": "976243e2861066edc99111ffc150a8141dda0494f026576eb3e22e90ae405d03.jpg"
}
]
}
],
"index": 30,
"virtual_lines": [
{
"bbox": [
162,
482,
448,
493.3333333333333
],
"spans": [],
"index": 29
},
{
"bbox": [
162,
493.3333333333333,
448,
504.66666666666663
],
"spans": [],
"index": 30
},
{
"bbox": [
162,
504.66666666666663,
448,
516.0
],
"spans": [],
"index": 31
}
]
},
{
"type": "text",
"bbox": [
106,
519,
351,
531
],
"lines": [
{
"bbox": [
106,
518,
351,
532
],
"spans": [
{
"bbox": [
106,
518,
351,
532
],
"score": 1.0,
"content": "Substituting the above Taylor expansion into Eq. (3.5) yields",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "interline_equation",
"bbox": [
111,
534,
482,
569
],
"lines": [
{
"bbox": [
111,
534,
482,
569
],
"spans": [
{
"bbox": [
111,
534,
482,
569
],
"score": 0.94,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\sum _ { n = 0 } ^ { k - 1 } \\hat { \\pmb { \\epsilon } } _ { \\theta } ^ { ( n ) } ( \\hat { \\pmb { x } } _ { { \\pmb { \\lambda } } _ { t _ { i - 1 } } } , \\lambda _ { t _ { i - 1 } } ) \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\mathrm { d } \\lambda + \\mathcal { O } ( h _ { i } ^ { k + 1 } ) ,",
"type": "interline_equation",
"image_path": "aa02a3d9e9670da70428ff51d8dccbfd33121e4f699e839306505e69a37a124c.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
111,
534,
482,
545.6666666666666
],
"spans": [],
"index": 33
},
{
"bbox": [
111,
545.6666666666666,
482,
557.3333333333333
],
"spans": [],
"index": 34
},
{
"bbox": [
111,
557.3333333333333,
482,
568.9999999999999
],
"spans": [],
"index": 35
}
]
},
{
"type": "text",
"bbox": [
106,
572,
506,
660
],
"lines": [
{
"bbox": [
102,
569,
509,
593
],
"spans": [
{
"bbox": [
102,
569,
179,
593
],
"score": 1.0,
"content": "where the integral",
"type": "text"
},
{
"bbox": [
180,
572,
261,
590
],
"score": 0.93,
"content": "\\begin{array} { r } { \\int e ^ { - \\lambda } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\mathrm { d } \\lambda } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
261,
569,
473,
593
],
"score": 1.0,
"content": "can be analytically computed by repeatedly applying",
"type": "text"
},
{
"bbox": [
473,
579,
480,
587
],
"score": 0.75,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
480,
569,
509,
593
],
"score": 1.0,
"content": "times",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
585,
507,
603
],
"spans": [
{
"bbox": [
105,
585,
396,
603
],
"score": 1.0,
"content": "of integration-by-parts (see Appendix B.2). Therefore, to approximate",
"type": "text"
},
{
"bbox": [
396,
590,
433,
601
],
"score": 0.91,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
433,
585,
507,
603
],
"score": 1.0,
"content": ", we only need to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
103,
599,
507,
617
],
"spans": [
{
"bbox": [
103,
599,
176,
617
],
"score": 1.0,
"content": "approximate the",
"type": "text"
},
{
"bbox": [
176,
604,
183,
613
],
"score": 0.77,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
184,
599,
290,
617
],
"score": 1.0,
"content": "-th order total derivatives",
"type": "text"
},
{
"bbox": [
290,
600,
337,
615
],
"score": 0.94,
"content": "\\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
337,
599,
354,
617
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
354,
603,
401,
614
],
"score": 0.9,
"content": "n \\leq k - 1",
"type": "inline_equation"
},
{
"bbox": [
402,
599,
507,
617
],
"score": 1.0,
"content": ", which is a well-studied",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
612,
507,
630
],
"spans": [
{
"bbox": [
104,
612,
332,
630
],
"score": 1.0,
"content": "problem in the ODE literature [31, 32]. By dropping the",
"type": "text"
},
{
"bbox": [
332,
614,
369,
628
],
"score": 0.93,
"content": "\\mathcal { O } ( h _ { i } ^ { k + 1 } )",
"type": "inline_equation"
},
{
"bbox": [
370,
612,
507,
630
],
"score": 1.0,
"content": "error term and approximating the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
626,
506,
640
],
"spans": [
{
"bbox": [
105,
626,
125,
640
],
"score": 1.0,
"content": "first",
"type": "text"
},
{
"bbox": [
126,
628,
156,
639
],
"score": 0.92,
"content": "( k - 1 )",
"type": "inline_equation"
},
{
"bbox": [
156,
626,
461,
640
],
"score": 1.0,
"content": "-th total derivatives with the “stiff order conditions” [31, 32], we can derive",
"type": "text"
},
{
"bbox": [
461,
627,
468,
637
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
468,
626,
506,
640
],
"score": 1.0,
"content": "-th-order",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
638,
504,
649
],
"spans": [
{
"bbox": [
106,
638,
497,
649
],
"score": 1.0,
"content": "ODE solvers for diffusion ODEs. We name such solvers as DPM-Solver overall, and DPM-Solver-",
"type": "text"
},
{
"bbox": [
497,
638,
504,
648
],
"score": 0.77,
"content": "k",
"type": "inline_equation"
}
],
"index": 41
},
{
"bbox": [
105,
647,
478,
662
],
"spans": [
{
"bbox": [
105,
647,
184,
662
],
"score": 1.0,
"content": "for a specific order",
"type": "text"
},
{
"bbox": [
184,
649,
191,
658
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
191,
647,
250,
662
],
"score": 1.0,
"content": ". Here we take",
"type": "text"
},
{
"bbox": [
250,
649,
275,
658
],
"score": 0.89,
"content": "k = 1",
"type": "inline_equation"
},
{
"bbox": [
275,
647,
478,
662
],
"score": 1.0,
"content": "for demonstration. In this case, Eq. (3.6) becomes",
"type": "text"
}
],
"index": 42
}
],
"index": 39
},
{
"type": "interline_equation",
"bbox": [
165,
663,
446,
722
],
"lines": [
{
"bbox": [
165,
663,
446,
722
],
"spans": [
{
"bbox": [
165,
663,
446,
722
],
"score": 0.93,
"content": "\\begin{array} { l } { \\displaystyle { \\boldsymbol { x } } _ { t _ { i - 1 } \\to t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\epsilon _ { \\theta } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\mathrm { d } \\lambda + \\mathcal { O } ( h _ { i } ^ { 2 } ) } \\\\ { \\displaystyle = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) + \\mathcal { O } ( h _ { i } ^ { 2 } ) . } \\end{array}",
"type": "interline_equation",
"image_path": "e1a7a77f08fcf2344ebba70e31c391e6092d3781368a0af49fae27dbeb5b255e.jpg"
}
]
}
],
"index": 44,
"virtual_lines": [
{
"bbox": [
165,
663,
446,
682.6666666666666
],
"spans": [],
"index": 43
},
{
"bbox": [
165,
682.6666666666666,
446,
702.3333333333333
],
"spans": [],
"index": 44
},
{
"bbox": [
165,
702.3333333333333,
446,
721.9999999999999
],
"spans": [],
"index": 45
}
]
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
302,
741,
309,
752
],
"spans": [
{
"bbox": [
302,
741,
309,
752
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
96
],
"lines": [
{
"bbox": [
106,
72,
505,
85
],
"spans": [
{
"bbox": [
106,
72,
416,
85
],
"score": 1.0,
"content": "Proposition 3.1 (Exact solution of diffusion ODEs). Given an initial value",
"type": "text"
},
{
"bbox": [
417,
75,
429,
84
],
"score": 0.85,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
429,
72,
461,
85
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
461,
73,
486,
83
],
"score": 0.88,
"content": "s > 0",
"type": "inline_equation"
},
{
"bbox": [
486,
72,
505,
85
],
"score": 1.0,
"content": ", the",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
83,
356,
96
],
"spans": [
{
"bbox": [
106,
83,
141,
96
],
"score": 1.0,
"content": "solution",
"type": "text"
},
{
"bbox": [
141,
86,
152,
95
],
"score": 0.82,
"content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }",
"type": "inline_equation"
},
{
"bbox": [
153,
83,
183,
96
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
184,
83,
220,
96
],
"score": 0.93,
"content": "t \\in [ 0 , s ]",
"type": "inline_equation"
},
{
"bbox": [
220,
83,
356,
96
],
"score": 1.0,
"content": "of diffusion ODEs in Eq. (2.7) is:",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
106,
72,
505,
96
]
},
{
"type": "interline_equation",
"bbox": [
224,
99,
387,
129
],
"lines": [
{
"bbox": [
224,
99,
387,
129
],
"spans": [
{
"bbox": [
224,
99,
387,
129
],
"score": 0.95,
"content": "\\pmb { x } _ { t } = \\frac { \\alpha _ { t } } { \\alpha _ { s } } \\pmb { x } _ { s } - \\alpha _ { t } \\int _ { \\lambda _ { s } } ^ { \\lambda _ { t } } e ^ { - \\lambda } \\hat { \\pmb { \\epsilon } } _ { \\theta } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda .",
"type": "interline_equation",
"image_path": "9abcfd7effacf68fb53c21394f34e41f5f607171cf80a13282a4bf3971c34831.jpg"
}
]
}
],
"index": 2.5,
"virtual_lines": [
{
"bbox": [
224,
99,
387,
114.0
],
"spans": [],
"index": 2
},
{
"bbox": [
224,
114.0,
387,
129.0
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
106,
138,
505,
172
],
"lines": [
{
"bbox": [
104,
137,
506,
153
],
"spans": [
{
"bbox": [
104,
137,
191,
153
],
"score": 1.0,
"content": "We call the integral",
"type": "text"
},
{
"bbox": [
192,
138,
267,
151
],
"score": 0.93,
"content": "\\begin{array} { r } { \\int e ^ { - \\lambda } \\hat { \\epsilon } _ { \\theta } ( \\hat { \\pmb x } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
267,
137,
430,
153
],
"score": 1.0,
"content": "the exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
431,
139,
441,
150
],
"score": 0.87,
"content": "\\scriptstyle { \\hat { \\epsilon } } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
441,
137,
506,
153
],
"score": 1.0,
"content": ", which is very",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
149,
506,
162
],
"spans": [
{
"bbox": [
105,
149,
506,
162
],
"score": 1.0,
"content": "special and highly related to the exponential integrators in the literature of ODE solvers [25]. To the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
160,
496,
173
],
"spans": [
{
"bbox": [
106,
160,
496,
173
],
"score": 1.0,
"content": "best of our knowledge, such formulation has not been revealed in prior work of diffusion models.",
"type": "text"
}
],
"index": 6
}
],
"index": 5,
"bbox_fs": [
104,
137,
506,
173
]
},
{
"type": "text",
"bbox": [
106,
177,
505,
232
],
"lines": [
{
"bbox": [
106,
176,
506,
190
],
"spans": [
{
"bbox": [
106,
176,
506,
190
],
"score": 1.0,
"content": "Eq. (3.4) provides a new perspective for approximating the solutions of diffusion ODEs. Specifically,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
188,
506,
201
],
"spans": [
{
"bbox": [
106,
188,
132,
201
],
"score": 1.0,
"content": "given",
"type": "text"
},
{
"bbox": [
132,
190,
144,
199
],
"score": 0.86,
"content": "\\mathbf { \\delta } _ { \\mathbf { \\mathcal { X } } _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
144,
188,
177,
201
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
178,
190,
183,
198
],
"score": 0.68,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
184,
188,
432,
201
],
"score": 1.0,
"content": ", According to Eq. (3.4), approximating the solution at time",
"type": "text"
},
{
"bbox": [
433,
189,
438,
198
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
438,
188,
506,
201
],
"score": 1.0,
"content": "is equivalent to",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
198,
505,
212
],
"spans": [
{
"bbox": [
106,
198,
360,
212
],
"score": 1.0,
"content": "directly approximating the exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
361,
199,
371,
210
],
"score": 0.88,
"content": "\\hat { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
371,
198,
395,
212
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
395,
199,
407,
210
],
"score": 0.88,
"content": "\\lambda _ { s }",
"type": "inline_equation"
},
{
"bbox": [
407,
198,
419,
212
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
419,
199,
429,
210
],
"score": 0.87,
"content": "\\lambda _ { t }",
"type": "inline_equation"
},
{
"bbox": [
430,
198,
505,
212
],
"score": 1.0,
"content": ", which avoids the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
209,
506,
223
],
"spans": [
{
"bbox": [
105,
209,
506,
223
],
"score": 1.0,
"content": "error of the linear terms and is well-studied in the literature of exponential integrators [25, 31]. Based",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
220,
492,
234
],
"spans": [
{
"bbox": [
106,
220,
492,
234
],
"score": 1.0,
"content": "on this insight, we propose fast solvers for diffusion ODEs, as detailed in the following sections.",
"type": "text"
}
],
"index": 11
}
],
"index": 9,
"bbox_fs": [
105,
176,
506,
234
]
},
{
"type": "title",
"bbox": [
106,
244,
300,
257
],
"lines": [
{
"bbox": [
105,
244,
301,
259
],
"spans": [
{
"bbox": [
105,
244,
301,
259
],
"score": 1.0,
"content": "3.2 High-Order Solvers for Diffusion ODEs",
"type": "text"
}
],
"index": 12
}
],
"index": 12
},
{
"type": "text",
"bbox": [
106,
264,
505,
298
],
"lines": [
{
"bbox": [
105,
264,
505,
278
],
"spans": [
{
"bbox": [
105,
264,
505,
278
],
"score": 1.0,
"content": "In this section, we propose high-order solvers for diffusion ODEs with convergence order guarantee",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
276,
505,
288
],
"spans": [
{
"bbox": [
106,
276,
505,
288
],
"score": 1.0,
"content": "by leveraging our proposed solution formulation Eq. (3.4). The proposed solvers and analysis are",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
287,
467,
299
],
"spans": [
{
"bbox": [
105,
287,
467,
299
],
"score": 1.0,
"content": "highly motivated by the methods of exponential integrators [25, 31] in the ODE literature.",
"type": "text"
}
],
"index": 15
}
],
"index": 14,
"bbox_fs": [
105,
264,
505,
299
]
},
{
"type": "text",
"bbox": [
106,
302,
505,
348
],
"lines": [
{
"bbox": [
103,
300,
507,
318
],
"spans": [
{
"bbox": [
103,
300,
249,
318
],
"score": 1.0,
"content": "Specifically, given an initial value",
"type": "text"
},
{
"bbox": [
250,
305,
264,
314
],
"score": 0.85,
"content": "\\mathbf { \\nabla } _ { \\mathbf { x } _ { T } }",
"type": "inline_equation"
},
{
"bbox": [
264,
300,
297,
318
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
298,
304,
306,
313
],
"score": 0.81,
"content": "T",
"type": "inline_equation"
},
{
"bbox": [
307,
300,
326,
318
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
326,
303,
356,
314
],
"score": 0.89,
"content": "M + 1",
"type": "inline_equation"
},
{
"bbox": [
356,
300,
402,
318
],
"score": 1.0,
"content": "time steps",
"type": "text"
},
{
"bbox": [
403,
302,
434,
315
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
434,
300,
507,
318
],
"score": 1.0,
"content": "decreasing from",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
313,
505,
327
],
"spans": [
{
"bbox": [
106,
314,
136,
325
],
"score": 0.9,
"content": "t _ { 0 } = T",
"type": "inline_equation"
},
{
"bbox": [
137,
313,
147,
327
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
147,
315,
179,
325
],
"score": 0.9,
"content": "t _ { M } = 0",
"type": "inline_equation"
},
{
"bbox": [
180,
313,
199,
327
],
"score": 1.0,
"content": ". Let",
"type": "text"
},
{
"bbox": [
199,
315,
240,
325
],
"score": 0.89,
"content": "\\tilde { \\mathbf { x } } _ { t _ { 0 } } = \\mathbf { x } _ { T }",
"type": "inline_equation"
},
{
"bbox": [
240,
313,
419,
327
],
"score": 1.0,
"content": "be the initial value. The proposed solvers use",
"type": "text"
},
{
"bbox": [
419,
315,
431,
324
],
"score": 0.81,
"content": "M",
"type": "inline_equation"
},
{
"bbox": [
431,
313,
505,
327
],
"score": 1.0,
"content": "steps to iteratively",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
102,
320,
510,
342
],
"spans": [
{
"bbox": [
102,
320,
190,
342
],
"score": 1.0,
"content": "compute a sequence",
"type": "text"
},
{
"bbox": [
190,
325,
227,
337
],
"score": 0.92,
"content": "\\{ \\tilde { { \\pmb { x } } } _ { t _ { i } } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
228,
320,
416,
342
],
"score": 1.0,
"content": "to approximate the true solutions at time steps",
"type": "text"
},
{
"bbox": [
417,
324,
448,
337
],
"score": 0.92,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
448,
320,
510,
342
],
"score": 1.0,
"content": ". In particular,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
335,
348,
349
],
"spans": [
{
"bbox": [
105,
335,
165,
349
],
"score": 1.0,
"content": "the last iterate",
"type": "text"
},
{
"bbox": [
165,
336,
183,
348
],
"score": 0.91,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { M } }",
"type": "inline_equation"
},
{
"bbox": [
183,
335,
348,
349
],
"score": 1.0,
"content": "approximates the true solution at time 0.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5,
"bbox_fs": [
102,
300,
510,
349
]
},
{
"type": "text",
"bbox": [
107,
352,
505,
386
],
"lines": [
{
"bbox": [
105,
352,
505,
365
],
"spans": [
{
"bbox": [
105,
352,
314,
365
],
"score": 1.0,
"content": "In order to reduce the approximation error between",
"type": "text"
},
{
"bbox": [
315,
353,
333,
364
],
"score": 0.91,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } }",
"type": "inline_equation"
},
{
"bbox": [
333,
352,
505,
365
],
"score": 1.0,
"content": "and the true solution at time 0, we need to",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
360,
503,
379
],
"spans": [
{
"bbox": [
104,
360,
264,
379
],
"score": 1.0,
"content": "reduce the approximation error for each",
"type": "text"
},
{
"bbox": [
264,
364,
279,
375
],
"score": 0.9,
"content": "\\tilde { \\mathbf { x } } _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
279,
360,
480,
379
],
"score": 1.0,
"content": "at every step [30]. Starting with the previous value",
"type": "text"
},
{
"bbox": [
480,
363,
503,
376
],
"score": 0.91,
"content": "\\tilde { \\pmb { x } } _ { t _ { i - 1 } }",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
105,
374,
447,
388
],
"spans": [
{
"bbox": [
105,
374,
136,
388
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
136,
375,
154,
386
],
"score": 0.89,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
155,
374,
323,
388
],
"score": 1.0,
"content": ", according to Eq. (3.4), the exact solution",
"type": "text"
},
{
"bbox": [
324,
376,
360,
387
],
"score": 0.92,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
360,
374,
392,
388
],
"score": 1.0,
"content": "at time",
"type": "text"
},
{
"bbox": [
392,
375,
400,
385
],
"score": 0.85,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
400,
374,
447,
388
],
"score": 1.0,
"content": "is given by",
"type": "text"
}
],
"index": 22
}
],
"index": 21,
"bbox_fs": [
104,
352,
505,
388
]
},
{
"type": "interline_equation",
"bbox": [
194,
390,
416,
422
],
"lines": [
{
"bbox": [
194,
390,
416,
422
],
"spans": [
{
"bbox": [
194,
390,
416,
422
],
"score": 0.93,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\hat { \\pmb { \\epsilon } } _ { \\theta } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda ) \\mathrm { d } \\lambda .",
"type": "interline_equation",
"image_path": "e941fd95cbf3e1c496583b82dd623ff65158438a8e8529e43b6a17ad2855df95.jpg"
}
]
}
],
"index": 23.5,
"virtual_lines": [
{
"bbox": [
194,
390,
416,
406.0
],
"spans": [],
"index": 23
},
{
"bbox": [
194,
406.0,
416,
422.0
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
106,
425,
505,
478
],
"lines": [
{
"bbox": [
104,
423,
506,
441
],
"spans": [
{
"bbox": [
104,
423,
245,
441
],
"score": 1.0,
"content": "Therefore, to compute the value",
"type": "text"
},
{
"bbox": [
246,
426,
260,
438
],
"score": 0.9,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
260,
423,
342,
441
],
"score": 1.0,
"content": "for approximating",
"type": "text"
},
{
"bbox": [
342,
428,
379,
438
],
"score": 0.92,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
379,
423,
506,
441
],
"score": 1.0,
"content": ", we need to approximate the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
436,
505,
455
],
"spans": [
{
"bbox": [
104,
436,
241,
455
],
"score": 1.0,
"content": "exponentially weighted integral of",
"type": "text"
},
{
"bbox": [
241,
441,
252,
452
],
"score": 0.87,
"content": "\\hat { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
252,
436,
273,
455
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
274,
441,
295,
453
],
"score": 0.92,
"content": "\\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
295,
436,
306,
455
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
307,
441,
320,
452
],
"score": 0.89,
"content": "\\lambda _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
320,
436,
356,
455
],
"score": 1.0,
"content": ". Denote",
"type": "text"
},
{
"bbox": [
357,
440,
424,
453
],
"score": 0.92,
"content": "h _ { i } : = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
425,
436,
444,
455
],
"score": 1.0,
"content": ", and",
"type": "text"
},
{
"bbox": [
444,
437,
505,
453
],
"score": 0.91,
"content": "\\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { { \\mathbf x } } _ { \\lambda } , \\lambda ) \\mathrel { \\mathop : } =",
"type": "inline_equation"
}
],
"index": 26
},
{
"bbox": [
104,
450,
507,
469
],
"spans": [
{
"bbox": [
104,
450,
507,
469
],
"score": 1.0,
"content": "dnϵˆθ(xˆλ,λ)dλn as the n-th order total derivative of ϵˆθ(xˆλ, λ) w.r.t. λ. For k ≥ 1, the (k − 1)-th order",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
464,
304,
479
],
"spans": [
{
"bbox": [
105,
464,
188,
479
],
"score": 1.0,
"content": "Taylor expansion of",
"type": "text"
},
{
"bbox": [
188,
465,
228,
478
],
"score": 0.93,
"content": "\\hat { \\epsilon } _ { \\theta } ( \\hat { \\pmb x } _ { \\lambda } , \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
228,
464,
252,
479
],
"score": 1.0,
"content": "w.r.t.",
"type": "text"
},
{
"bbox": [
252,
466,
259,
475
],
"score": 0.79,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
260,
464,
270,
479
],
"score": 1.0,
"content": "at",
"type": "text"
},
{
"bbox": [
271,
466,
292,
478
],
"score": 0.92,
"content": "\\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
292,
464,
304,
479
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5,
"bbox_fs": [
104,
423,
507,
479
]
},
{
"type": "interline_equation",
"bbox": [
162,
482,
448,
516
],
"lines": [
{
"bbox": [
162,
482,
448,
516
],
"spans": [
{
"bbox": [
162,
482,
448,
516
],
"score": 0.94,
"content": "\\hat { \\epsilon } _ { \\theta } ( \\hat { x } _ { \\lambda } , \\lambda ) = \\sum _ { n = 0 } ^ { k - 1 } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { x } _ { \\lambda _ { t _ { i - 1 } } } , \\lambda _ { t _ { i - 1 } } ) + \\mathcal { O } ( ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { k } ) ,",
"type": "interline_equation",
"image_path": "976243e2861066edc99111ffc150a8141dda0494f026576eb3e22e90ae405d03.jpg"
}
]
}
],
"index": 30,
"virtual_lines": [
{
"bbox": [
162,
482,
448,
493.3333333333333
],
"spans": [],
"index": 29
},
{
"bbox": [
162,
493.3333333333333,
448,
504.66666666666663
],
"spans": [],
"index": 30
},
{
"bbox": [
162,
504.66666666666663,
448,
516.0
],
"spans": [],
"index": 31
}
]
},
{
"type": "text",
"bbox": [
106,
519,
351,
531
],
"lines": [
{
"bbox": [
106,
518,
351,
532
],
"spans": [
{
"bbox": [
106,
518,
351,
532
],
"score": 1.0,
"content": "Substituting the above Taylor expansion into Eq. (3.5) yields",
"type": "text"
}
],
"index": 32
}
],
"index": 32,
"bbox_fs": [
106,
518,
351,
532
]
},
{
"type": "interline_equation",
"bbox": [
111,
534,
482,
569
],
"lines": [
{
"bbox": [
111,
534,
482,
569
],
"spans": [
{
"bbox": [
111,
534,
482,
569
],
"score": 0.94,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\sum _ { n = 0 } ^ { k - 1 } \\hat { \\pmb { \\epsilon } } _ { \\theta } ^ { ( n ) } ( \\hat { \\pmb { x } } _ { { \\pmb { \\lambda } } _ { t _ { i - 1 } } } , \\lambda _ { t _ { i - 1 } } ) \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\mathrm { d } \\lambda + \\mathcal { O } ( h _ { i } ^ { k + 1 } ) ,",
"type": "interline_equation",
"image_path": "aa02a3d9e9670da70428ff51d8dccbfd33121e4f699e839306505e69a37a124c.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
111,
534,
482,
545.6666666666666
],
"spans": [],
"index": 33
},
{
"bbox": [
111,
545.6666666666666,
482,
557.3333333333333
],
"spans": [],
"index": 34
},
{
"bbox": [
111,
557.3333333333333,
482,
568.9999999999999
],
"spans": [],
"index": 35
}
]
},
{
"type": "text",
"bbox": [
106,
572,
506,
660
],
"lines": [
{
"bbox": [
102,
569,
509,
593
],
"spans": [
{
"bbox": [
102,
569,
179,
593
],
"score": 1.0,
"content": "where the integral",
"type": "text"
},
{
"bbox": [
180,
572,
261,
590
],
"score": 0.93,
"content": "\\begin{array} { r } { \\int e ^ { - \\lambda } \\frac { ( \\lambda - \\lambda _ { t _ { i - 1 } } ) ^ { n } } { n ! } \\mathrm { d } \\lambda } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
261,
569,
473,
593
],
"score": 1.0,
"content": "can be analytically computed by repeatedly applying",
"type": "text"
},
{
"bbox": [
473,
579,
480,
587
],
"score": 0.75,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
480,
569,
509,
593
],
"score": 1.0,
"content": "times",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
585,
507,
603
],
"spans": [
{
"bbox": [
105,
585,
396,
603
],
"score": 1.0,
"content": "of integration-by-parts (see Appendix B.2). Therefore, to approximate",
"type": "text"
},
{
"bbox": [
396,
590,
433,
601
],
"score": 0.91,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
433,
585,
507,
603
],
"score": 1.0,
"content": ", we only need to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
103,
599,
507,
617
],
"spans": [
{
"bbox": [
103,
599,
176,
617
],
"score": 1.0,
"content": "approximate the",
"type": "text"
},
{
"bbox": [
176,
604,
183,
613
],
"score": 0.77,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
184,
599,
290,
617
],
"score": 1.0,
"content": "-th order total derivatives",
"type": "text"
},
{
"bbox": [
290,
600,
337,
615
],
"score": 0.94,
"content": "\\hat { \\epsilon } _ { \\theta } ^ { ( n ) } ( \\hat { \\pmb { x } } _ { \\lambda } , \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
337,
599,
354,
617
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
354,
603,
401,
614
],
"score": 0.9,
"content": "n \\leq k - 1",
"type": "inline_equation"
},
{
"bbox": [
402,
599,
507,
617
],
"score": 1.0,
"content": ", which is a well-studied",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
612,
507,
630
],
"spans": [
{
"bbox": [
104,
612,
332,
630
],
"score": 1.0,
"content": "problem in the ODE literature [31, 32]. By dropping the",
"type": "text"
},
{
"bbox": [
332,
614,
369,
628
],
"score": 0.93,
"content": "\\mathcal { O } ( h _ { i } ^ { k + 1 } )",
"type": "inline_equation"
},
{
"bbox": [
370,
612,
507,
630
],
"score": 1.0,
"content": "error term and approximating the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
626,
506,
640
],
"spans": [
{
"bbox": [
105,
626,
125,
640
],
"score": 1.0,
"content": "first",
"type": "text"
},
{
"bbox": [
126,
628,
156,
639
],
"score": 0.92,
"content": "( k - 1 )",
"type": "inline_equation"
},
{
"bbox": [
156,
626,
461,
640
],
"score": 1.0,
"content": "-th total derivatives with the “stiff order conditions” [31, 32], we can derive",
"type": "text"
},
{
"bbox": [
461,
627,
468,
637
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
468,
626,
506,
640
],
"score": 1.0,
"content": "-th-order",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
638,
504,
649
],
"spans": [
{
"bbox": [
106,
638,
497,
649
],
"score": 1.0,
"content": "ODE solvers for diffusion ODEs. We name such solvers as DPM-Solver overall, and DPM-Solver-",
"type": "text"
},
{
"bbox": [
497,
638,
504,
648
],
"score": 0.77,
"content": "k",
"type": "inline_equation"
}
],
"index": 41
},
{
"bbox": [
105,
647,
478,
662
],
"spans": [
{
"bbox": [
105,
647,
184,
662
],
"score": 1.0,
"content": "for a specific order",
"type": "text"
},
{
"bbox": [
184,
649,
191,
658
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
191,
647,
250,
662
],
"score": 1.0,
"content": ". Here we take",
"type": "text"
},
{
"bbox": [
250,
649,
275,
658
],
"score": 0.89,
"content": "k = 1",
"type": "inline_equation"
},
{
"bbox": [
275,
647,
478,
662
],
"score": 1.0,
"content": "for demonstration. In this case, Eq. (3.6) becomes",
"type": "text"
}
],
"index": 42
}
],
"index": 39,
"bbox_fs": [
102,
569,
509,
662
]
},
{
"type": "interline_equation",
"bbox": [
165,
663,
446,
722
],
"lines": [
{
"bbox": [
165,
663,
446,
722
],
"spans": [
{
"bbox": [
165,
663,
446,
722
],
"score": 0.93,
"content": "\\begin{array} { l } { \\displaystyle { \\boldsymbol { x } } _ { t _ { i - 1 } \\to t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\epsilon _ { \\theta } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) \\int _ { \\lambda _ { t _ { i - 1 } } } ^ { \\lambda _ { t _ { i } } } e ^ { - \\lambda } \\mathrm { d } \\lambda + \\mathcal { O } ( h _ { i } ^ { 2 } ) } \\\\ { \\displaystyle = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) + \\mathcal { O } ( h _ { i } ^ { 2 } ) . } \\end{array}",
"type": "interline_equation",
"image_path": "e1a7a77f08fcf2344ebba70e31c391e6092d3781368a0af49fae27dbeb5b255e.jpg"
}
]
}
],
"index": 44,
"virtual_lines": [
{
"bbox": [
165,
663,
446,
682.6666666666666
],
"spans": [],
"index": 43
},
{
"bbox": [
165,
682.6666666666666,
446,
702.3333333333333
],
"spans": [],
"index": 44
},
{
"bbox": [
165,
702.3333333333333,
446,
721.9999999999999
],
"spans": [],
"index": 45
}
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
95
],
"lines": [
{
"bbox": [
105,
71,
505,
87
],
"spans": [
{
"bbox": [
105,
71,
266,
87
],
"score": 1.0,
"content": "By dropping the high-order error term",
"type": "text"
},
{
"bbox": [
267,
72,
293,
85
],
"score": 0.92,
"content": "\\mathcal { O } ( h _ { i } ^ { 2 } )",
"type": "inline_equation"
},
{
"bbox": [
294,
71,
448,
87
],
"score": 1.0,
"content": ", we can obtain an approximation for",
"type": "text"
},
{
"bbox": [
448,
75,
485,
85
],
"score": 0.89,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
485,
71,
505,
87
],
"score": 1.0,
"content": ". As",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
107,
80,
464,
99
],
"spans": [
{
"bbox": [
107,
84,
132,
94
],
"score": 0.88,
"content": "k = 1",
"type": "inline_equation"
},
{
"bbox": [
132,
80,
283,
99
],
"score": 1.0,
"content": "here, we call this solver DPM-Solver-",
"type": "text"
},
{
"bbox": [
284,
85,
289,
93
],
"score": 0.69,
"content": "^ { l }",
"type": "inline_equation"
},
{
"bbox": [
289,
80,
464,
99
],
"score": 1.0,
"content": ", and the detailed algorithm is as following.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
104,
99,
505,
124
],
"lines": [
{
"bbox": [
104,
95,
504,
117
],
"spans": [
{
"bbox": [
104,
95,
268,
117
],
"score": 1.0,
"content": "DPM-Solver-1. Given an initial value",
"type": "text"
},
{
"bbox": [
268,
102,
282,
111
],
"score": 0.87,
"content": "\\mathbf { \\nabla } _ { \\mathbf { x } _ { T } }",
"type": "inline_equation"
},
{
"bbox": [
283,
95,
300,
117
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
301,
100,
330,
111
],
"score": 0.89,
"content": "M + 1",
"type": "inline_equation"
},
{
"bbox": [
330,
95,
374,
117
],
"score": 1.0,
"content": "time steps",
"type": "text"
},
{
"bbox": [
374,
99,
405,
113
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
406,
95,
474,
117
],
"score": 1.0,
"content": "decreasing from",
"type": "text"
},
{
"bbox": [
474,
100,
504,
111
],
"score": 0.88,
"content": "t _ { 0 } = T",
"type": "inline_equation"
}
],
"index": 2
},
{
"bbox": [
104,
108,
486,
128
],
"spans": [
{
"bbox": [
104,
108,
117,
128
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
113,
149,
123
],
"score": 0.9,
"content": "t _ { M } = 0",
"type": "inline_equation"
},
{
"bbox": [
149,
108,
207,
128
],
"score": 1.0,
"content": ". Starting with",
"type": "text"
},
{
"bbox": [
208,
112,
248,
124
],
"score": 0.92,
"content": "\\tilde { \\mathbf { x } } _ { t _ { 0 } } = \\mathbf { x } _ { T }",
"type": "inline_equation"
},
{
"bbox": [
249,
108,
306,
128
],
"score": 1.0,
"content": ", the sequence",
"type": "text"
},
{
"bbox": [
306,
111,
344,
124
],
"score": 0.93,
"content": "\\{ \\tilde { { x } } _ { t _ { i } } \\} _ { i = 1 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
344,
108,
486,
128
],
"score": 1.0,
"content": "is computed iteratively as follows:",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "interline_equation",
"bbox": [
149,
129,
461,
154
],
"lines": [
{
"bbox": [
149,
129,
461,
154
],
"spans": [
{
"bbox": [
149,
129,
461,
154
],
"score": 0.92,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\boldsymbol { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) , \\quad \\mathrm { w h e r e ~ } h _ { i } = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } } .",
"type": "interline_equation",
"image_path": "8deaa6e7603de31da2fd31418b2d224dd8ec375a261d169ca354b6f5b2802791.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
149,
129,
461,
154
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
164,
505,
198
],
"lines": [
{
"bbox": [
106,
164,
505,
177
],
"spans": [
{
"bbox": [
106,
164,
123,
177
],
"score": 1.0,
"content": "For",
"type": "text"
},
{
"bbox": [
123,
165,
149,
176
],
"score": 0.9,
"content": "k \\geq 2",
"type": "inline_equation"
},
{
"bbox": [
150,
164,
249,
177
],
"score": 1.0,
"content": ", approximating the first",
"type": "text"
},
{
"bbox": [
249,
165,
256,
174
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
257,
164,
505,
177
],
"score": 1.0,
"content": "terms of the Taylor expansion needs additional intermediate",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
176,
505,
188
],
"spans": [
{
"bbox": [
106,
176,
167,
188
],
"score": 1.0,
"content": "points between",
"type": "text"
},
{
"bbox": [
167,
177,
173,
186
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
173,
176,
190,
188
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
190,
177,
196,
186
],
"score": 0.77,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
196,
176,
505,
188
],
"score": 1.0,
"content": "[31]. The derivation is more technical so we defer it to Appendix B. Below we",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
186,
503,
200
],
"spans": [
{
"bbox": [
105,
186,
200,
200
],
"score": 1.0,
"content": "propose algorithms for",
"type": "text"
},
{
"bbox": [
200,
187,
234,
198
],
"score": 0.91,
"content": "k = 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
235,
186,
503,
200
],
"score": 1.0,
"content": "and name them as DPM-Solver-2 and DPM-Solver-3, respectively.",
"type": "text"
}
],
"index": 7
}
],
"index": 6
},
{
"type": "title",
"bbox": [
107,
212,
222,
223
],
"lines": [
{
"bbox": [
106,
211,
224,
225
],
"spans": [
{
"bbox": [
106,
211,
224,
225
],
"score": 1.0,
"content": "Algorithm 1 DPM-Solver-2.",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
227,
333,
240
],
"lines": [
{
"bbox": [
105,
226,
332,
244
],
"spans": [
{
"bbox": [
105,
226,
198,
244
],
"score": 1.0,
"content": "Require: initial value",
"type": "text"
},
{
"bbox": [
199,
230,
212,
239
],
"score": 0.86,
"content": "\\mathbf { \\nabla } _ { \\mathbf { \\mathcal { X } } T }",
"type": "inline_equation"
},
{
"bbox": [
213,
226,
259,
244
],
"score": 1.0,
"content": ", time steps",
"type": "text"
},
{
"bbox": [
259,
227,
290,
241
],
"score": 0.9,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
291,
226,
321,
244
],
"score": 1.0,
"content": ", model",
"type": "text"
},
{
"bbox": [
321,
230,
332,
240
],
"score": 0.69,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "title",
"bbox": [
107,
366,
223,
377
],
"lines": [
{
"bbox": [
106,
365,
224,
379
],
"spans": [
{
"bbox": [
106,
365,
224,
379
],
"score": 1.0,
"content": "Algorithm 2 DPM-Solver-3.",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
382,
333,
395
],
"lines": [
{
"bbox": [
104,
378,
332,
399
],
"spans": [
{
"bbox": [
104,
378,
198,
399
],
"score": 1.0,
"content": "Require: initial value",
"type": "text"
},
{
"bbox": [
199,
385,
212,
394
],
"score": 0.84,
"content": "\\mathbf { \\nabla } _ { \\mathbf { \\mathcal { X } } T }",
"type": "inline_equation"
},
{
"bbox": [
213,
378,
259,
399
],
"score": 1.0,
"content": ", time steps",
"type": "text"
},
{
"bbox": [
259,
381,
290,
395
],
"score": 0.88,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
290,
378,
321,
399
],
"score": 1.0,
"content": ", model",
"type": "text"
},
{
"bbox": [
322,
385,
332,
394
],
"score": 0.67,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "interline_equation",
"bbox": [
136,
414,
483,
511
],
"lines": [
{
"bbox": [
136,
414,
483,
511
],
"spans": [
{
"bbox": [
136,
414,
483,
511
],
"score": 0.93,
"content": "\\begin{array} { r l } & { \\quad _ { s 2 i - 1 } _ { t _ { \\lambda } } ( \\overline { { \\lambda } } _ { t _ { i - 1 } } + r _ { 1 } h _ { i } ) , \\quad s _ { 2 i } t _ { \\lambda } ( \\lambda _ { t _ { i - 1 } } + r _ { 2 } h _ { i } ) } \\\\ & { u _ { 2 i - 1 } \\frac { \\alpha _ { s _ { 2 i - 1 } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { s _ { 2 i - 1 } } ( e ^ { r _ { 1 } h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { D _ { 2 i - 1 } \\epsilon _ { \\theta } ( u _ { 2 i - 1 } , s _ { 2 i - 1 } ) - \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { u _ { 2 i } \\frac { \\alpha _ { s _ { 2 i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { s _ { 2 i } } ( e ^ { r _ { 2 } h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) - \\frac { \\sigma _ { s _ { 2 i } } r _ { 2 } } { r _ { 1 } } ( \\frac { e ^ { r _ { 2 } h _ { i } } - 1 } { r _ { 2 } h _ { i } } - 1 ) D _ { 2 i - 1 } } \\\\ & { D _ { 2 i } \\epsilon _ { \\theta } ( u _ { 2 i } , s _ { 2 i } ) - \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { \\tilde { x } _ { t _ { i } } \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) - \\frac { \\sigma _ { t _ { i } } } { r _ { 2 } } ( \\frac { e ^ { h _ { i } } - 1 } { h } - 1 ) D _ { 2 i } } \\end{array}",
"type": "interline_equation",
"image_path": "ad4609832a6419c85fb2132d9d91d2c95beec92e10bd4907c45b4ed7d6c517e6.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
136,
414,
483,
446.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
136,
446.3333333333333,
483,
478.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
136,
478.66666666666663,
483,
510.99999999999994
],
"spans": [],
"index": 14
}
]
},
{
"type": "text",
"bbox": [
108,
519,
171,
531
],
"lines": [
{
"bbox": [
106,
515,
171,
534
],
"spans": [
{
"bbox": [
106,
515,
153,
534
],
"score": 1.0,
"content": "10: return",
"type": "text"
},
{
"bbox": [
153,
519,
171,
531
],
"score": 0.81,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } }",
"type": "inline_equation"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "text",
"bbox": [
106,
553,
506,
631
],
"lines": [
{
"bbox": [
106,
554,
506,
566
],
"spans": [
{
"bbox": [
106,
554,
130,
566
],
"score": 1.0,
"content": "Here,",
"type": "text"
},
{
"bbox": [
130,
554,
151,
566
],
"score": 0.9,
"content": "t _ { \\lambda } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
151,
554,
252,
566
],
"score": 1.0,
"content": "is the inverse function of",
"type": "text"
},
{
"bbox": [
252,
554,
270,
566
],
"score": 0.91,
"content": "\\lambda ( t )",
"type": "inline_equation"
},
{
"bbox": [
270,
554,
506,
566
],
"score": 1.0,
"content": ", which has an analytical formulation for the practical noise",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
565,
506,
577
],
"spans": [
{
"bbox": [
105,
565,
458,
577
],
"score": 1.0,
"content": "schedule used in [2, 16], as shown in Appendix D. The chosen intermediate points are",
"type": "text"
},
{
"bbox": [
459,
565,
489,
576
],
"score": 0.89,
"content": "( s _ { i } , \\pmb { u } _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
490,
565,
506,
577
],
"score": 1.0,
"content": "for",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
574,
506,
590
],
"spans": [
{
"bbox": [
104,
574,
188,
590
],
"score": 1.0,
"content": "DPM-Solver-2 and",
"type": "text"
},
{
"bbox": [
188,
576,
247,
588
],
"score": 0.92,
"content": "\\left( s _ { 2 i - 1 } , { \\pmb u } _ { 2 i - 1 } \\right)",
"type": "inline_equation"
},
{
"bbox": [
248,
574,
267,
590
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
268,
576,
306,
588
],
"score": 0.92,
"content": "( s _ { 2 i } , { \\pmb u } _ { 2 i } )",
"type": "inline_equation"
},
{
"bbox": [
307,
574,
506,
590
],
"score": 1.0,
"content": "for DPM-Solver-3. As shown in the algorithm,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
586,
505,
599
],
"spans": [
{
"bbox": [
105,
586,
161,
599
],
"score": 1.0,
"content": "DPM-Solver-",
"type": "text"
},
{
"bbox": [
161,
587,
168,
597
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
168,
586,
203,
599
],
"score": 1.0,
"content": "requires",
"type": "text"
},
{
"bbox": [
204,
588,
210,
597
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
211,
586,
343,
599
],
"score": 1.0,
"content": "function evaluations per step for",
"type": "text"
},
{
"bbox": [
344,
587,
387,
598
],
"score": 0.93,
"content": "k = 1 , 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
388,
586,
505,
599
],
"score": 1.0,
"content": ". Despite the more expensive",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
596,
506,
611
],
"spans": [
{
"bbox": [
105,
596,
215,
611
],
"score": 1.0,
"content": "steps, higher-order solvers",
"type": "text"
},
{
"bbox": [
216,
598,
253,
609
],
"score": 0.83,
"content": "( k = 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
253,
596,
506,
611
],
"score": 1.0,
"content": ") are usually more efficient since they require much fewer steps",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
609,
506,
621
],
"spans": [
{
"bbox": [
105,
609,
418,
621
],
"score": 1.0,
"content": "to converge, due to their higher convergence order. We show that DPM-Solver-",
"type": "text"
},
{
"bbox": [
418,
609,
425,
618
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
425,
609,
435,
621
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
435,
609,
442,
618
],
"score": 0.84,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
442,
609,
506,
621
],
"score": 1.0,
"content": "-th-order solver,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
619,
362,
632
],
"spans": [
{
"bbox": [
105,
619,
362,
632
],
"score": 1.0,
"content": "as stated in the following theorem. The proof is in Appendix B.",
"type": "text"
}
],
"index": 22
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
634,
505,
682
],
"lines": [
{
"bbox": [
105,
633,
506,
647
],
"spans": [
{
"bbox": [
105,
633,
218,
647
],
"score": 1.0,
"content": "Theorem 3.2 (DPM-Solver-",
"type": "text"
},
{
"bbox": [
219,
635,
225,
644
],
"score": 0.8,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
226,
633,
244,
647
],
"score": 1.0,
"content": "as a",
"type": "text"
},
{
"bbox": [
244,
635,
250,
644
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
251,
633,
354,
647
],
"score": 1.0,
"content": "-th-order solver). Assume",
"type": "text"
},
{
"bbox": [
354,
634,
390,
646
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
391,
633,
506,
647
],
"score": 1.0,
"content": "follows the regularity condi-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
646,
505,
658
],
"spans": [
{
"bbox": [
106,
646,
262,
658
],
"score": 1.0,
"content": "tions detailed in Appendix B.1, then for",
"type": "text"
},
{
"bbox": [
262,
646,
306,
657
],
"score": 0.78,
"content": "k = 1 , 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
306,
646,
362,
658
],
"score": 1.0,
"content": ", DPM-Solver-",
"type": "text"
},
{
"bbox": [
362,
647,
369,
655
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
369,
646,
386,
658
],
"score": 1.0,
"content": "is a",
"type": "text"
},
{
"bbox": [
386,
646,
393,
655
],
"score": 0.72,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
393,
646,
505,
658
],
"score": 1.0,
"content": "-th order solver for diffusion",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
655,
507,
671
],
"spans": [
{
"bbox": [
105,
655,
218,
671
],
"score": 1.0,
"content": "ODEs, i.e., for the sequence",
"type": "text"
},
{
"bbox": [
218,
655,
256,
668
],
"score": 0.92,
"content": "\\{ \\tilde { \\pmb { x } } _ { t _ { i } } \\} _ { i = 1 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
256,
655,
362,
671
],
"score": 1.0,
"content": "computed by DPM-Solver-",
"type": "text"
},
{
"bbox": [
362,
657,
368,
667
],
"score": 0.71,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
369,
655,
507,
671
],
"score": 1.0,
"content": ", the approximation error at time 0",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
667,
400,
683
],
"spans": [
{
"bbox": [
104,
667,
140,
683
],
"score": 1.0,
"content": "satisfies",
"type": "text"
},
{
"bbox": [
141,
668,
232,
680
],
"score": 0.93,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } } - \\pmb { x } _ { 0 } = \\mathcal { O } ( h _ { \\operatorname* { m a x } } ^ { k } )",
"type": "inline_equation"
},
{
"bbox": [
232,
667,
263,
683
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
263,
668,
395,
681
],
"score": 0.92,
"content": "h _ { m a x } = \\mathrm { m a x } _ { 1 \\leq i \\leq M } ( \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } } )",
"type": "inline_equation"
},
{
"bbox": [
395,
667,
400,
683
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 26
}
],
"index": 24.5
},
{
"type": "text",
"bbox": [
107,
688,
505,
722
],
"lines": [
{
"bbox": [
106,
688,
505,
701
],
"spans": [
{
"bbox": [
106,
688,
189,
701
],
"score": 1.0,
"content": "Finally, solvers with",
"type": "text"
},
{
"bbox": [
189,
689,
214,
700
],
"score": 0.91,
"content": "k \\geq 4",
"type": "inline_equation"
},
{
"bbox": [
214,
688,
505,
701
],
"score": 1.0,
"content": "need much more intermediate points as shown by previous work [31, 32]",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
700,
505,
712
],
"spans": [
{
"bbox": [
105,
700,
329,
712
],
"score": 1.0,
"content": "for exponential integrators. Therefore, we only consider",
"type": "text"
},
{
"bbox": [
329,
700,
335,
710
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
336,
700,
505,
712
],
"score": 1.0,
"content": "from 1 to 3 in this work, while leaving the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
710,
259,
723
],
"spans": [
{
"bbox": [
106,
710,
185,
723
],
"score": 1.0,
"content": "solvers with higher",
"type": "text"
},
{
"bbox": [
185,
711,
192,
721
],
"score": 0.81,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
192,
710,
259,
723
],
"score": 1.0,
"content": "for future study.",
"type": "text"
}
],
"index": 29
}
],
"index": 28
}
],
"page_idx": 5,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
302,
741,
309,
752
],
"spans": [
{
"bbox": [
302,
741,
309,
752
],
"score": 1.0,
"content": "6",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
72,
504,
95
],
"lines": [
{
"bbox": [
105,
71,
505,
87
],
"spans": [
{
"bbox": [
105,
71,
266,
87
],
"score": 1.0,
"content": "By dropping the high-order error term",
"type": "text"
},
{
"bbox": [
267,
72,
293,
85
],
"score": 0.92,
"content": "\\mathcal { O } ( h _ { i } ^ { 2 } )",
"type": "inline_equation"
},
{
"bbox": [
294,
71,
448,
87
],
"score": 1.0,
"content": ", we can obtain an approximation for",
"type": "text"
},
{
"bbox": [
448,
75,
485,
85
],
"score": 0.89,
"content": "\\pmb { x } _ { t _ { i - 1 } t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
485,
71,
505,
87
],
"score": 1.0,
"content": ". As",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
107,
80,
464,
99
],
"spans": [
{
"bbox": [
107,
84,
132,
94
],
"score": 0.88,
"content": "k = 1",
"type": "inline_equation"
},
{
"bbox": [
132,
80,
283,
99
],
"score": 1.0,
"content": "here, we call this solver DPM-Solver-",
"type": "text"
},
{
"bbox": [
284,
85,
289,
93
],
"score": 0.69,
"content": "^ { l }",
"type": "inline_equation"
},
{
"bbox": [
289,
80,
464,
99
],
"score": 1.0,
"content": ", and the detailed algorithm is as following.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
105,
71,
505,
99
]
},
{
"type": "text",
"bbox": [
104,
99,
505,
124
],
"lines": [
{
"bbox": [
104,
95,
504,
117
],
"spans": [
{
"bbox": [
104,
95,
268,
117
],
"score": 1.0,
"content": "DPM-Solver-1. Given an initial value",
"type": "text"
},
{
"bbox": [
268,
102,
282,
111
],
"score": 0.87,
"content": "\\mathbf { \\nabla } _ { \\mathbf { x } _ { T } }",
"type": "inline_equation"
},
{
"bbox": [
283,
95,
300,
117
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
301,
100,
330,
111
],
"score": 0.89,
"content": "M + 1",
"type": "inline_equation"
},
{
"bbox": [
330,
95,
374,
117
],
"score": 1.0,
"content": "time steps",
"type": "text"
},
{
"bbox": [
374,
99,
405,
113
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
406,
95,
474,
117
],
"score": 1.0,
"content": "decreasing from",
"type": "text"
},
{
"bbox": [
474,
100,
504,
111
],
"score": 0.88,
"content": "t _ { 0 } = T",
"type": "inline_equation"
}
],
"index": 2
},
{
"bbox": [
104,
108,
486,
128
],
"spans": [
{
"bbox": [
104,
108,
117,
128
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
113,
149,
123
],
"score": 0.9,
"content": "t _ { M } = 0",
"type": "inline_equation"
},
{
"bbox": [
149,
108,
207,
128
],
"score": 1.0,
"content": ". Starting with",
"type": "text"
},
{
"bbox": [
208,
112,
248,
124
],
"score": 0.92,
"content": "\\tilde { \\mathbf { x } } _ { t _ { 0 } } = \\mathbf { x } _ { T }",
"type": "inline_equation"
},
{
"bbox": [
249,
108,
306,
128
],
"score": 1.0,
"content": ", the sequence",
"type": "text"
},
{
"bbox": [
306,
111,
344,
124
],
"score": 0.93,
"content": "\\{ \\tilde { { x } } _ { t _ { i } } \\} _ { i = 1 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
344,
108,
486,
128
],
"score": 1.0,
"content": "is computed iteratively as follows:",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
104,
95,
504,
128
]
},
{
"type": "interline_equation",
"bbox": [
149,
129,
461,
154
],
"lines": [
{
"bbox": [
149,
129,
461,
154
],
"spans": [
{
"bbox": [
149,
129,
461,
154
],
"score": 0.92,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\boldsymbol { \\epsilon } _ { \\boldsymbol { \\theta } } ( \\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) , \\quad \\mathrm { w h e r e ~ } h _ { i } = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } } .",
"type": "interline_equation",
"image_path": "8deaa6e7603de31da2fd31418b2d224dd8ec375a261d169ca354b6f5b2802791.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
149,
129,
461,
154
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
164,
505,
198
],
"lines": [
{
"bbox": [
106,
164,
505,
177
],
"spans": [
{
"bbox": [
106,
164,
123,
177
],
"score": 1.0,
"content": "For",
"type": "text"
},
{
"bbox": [
123,
165,
149,
176
],
"score": 0.9,
"content": "k \\geq 2",
"type": "inline_equation"
},
{
"bbox": [
150,
164,
249,
177
],
"score": 1.0,
"content": ", approximating the first",
"type": "text"
},
{
"bbox": [
249,
165,
256,
174
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
257,
164,
505,
177
],
"score": 1.0,
"content": "terms of the Taylor expansion needs additional intermediate",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
176,
505,
188
],
"spans": [
{
"bbox": [
106,
176,
167,
188
],
"score": 1.0,
"content": "points between",
"type": "text"
},
{
"bbox": [
167,
177,
173,
186
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
173,
176,
190,
188
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
190,
177,
196,
186
],
"score": 0.77,
"content": "s",
"type": "inline_equation"
},
{
"bbox": [
196,
176,
505,
188
],
"score": 1.0,
"content": "[31]. The derivation is more technical so we defer it to Appendix B. Below we",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
186,
503,
200
],
"spans": [
{
"bbox": [
105,
186,
200,
200
],
"score": 1.0,
"content": "propose algorithms for",
"type": "text"
},
{
"bbox": [
200,
187,
234,
198
],
"score": 0.91,
"content": "k = 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
235,
186,
503,
200
],
"score": 1.0,
"content": "and name them as DPM-Solver-2 and DPM-Solver-3, respectively.",
"type": "text"
}
],
"index": 7
}
],
"index": 6,
"bbox_fs": [
105,
164,
505,
200
]
},
{
"type": "title",
"bbox": [
107,
212,
222,
223
],
"lines": [
{
"bbox": [
106,
211,
224,
225
],
"spans": [
{
"bbox": [
106,
211,
224,
225
],
"score": 1.0,
"content": "Algorithm 1 DPM-Solver-2.",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
227,
333,
240
],
"lines": [
{
"bbox": [
105,
226,
332,
244
],
"spans": [
{
"bbox": [
105,
226,
198,
244
],
"score": 1.0,
"content": "Require: initial value",
"type": "text"
},
{
"bbox": [
199,
230,
212,
239
],
"score": 0.86,
"content": "\\mathbf { \\nabla } _ { \\mathbf { \\mathcal { X } } T }",
"type": "inline_equation"
},
{
"bbox": [
213,
226,
259,
244
],
"score": 1.0,
"content": ", time steps",
"type": "text"
},
{
"bbox": [
259,
227,
290,
241
],
"score": 0.9,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
291,
226,
321,
244
],
"score": 1.0,
"content": ", model",
"type": "text"
},
{
"bbox": [
321,
230,
332,
240
],
"score": 0.69,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
}
],
"index": 9
}
],
"index": 9,
"bbox_fs": [
105,
226,
332,
244
]
},
{
"type": "title",
"bbox": [
107,
366,
223,
377
],
"lines": [
{
"bbox": [
106,
365,
224,
379
],
"spans": [
{
"bbox": [
106,
365,
224,
379
],
"score": 1.0,
"content": "Algorithm 2 DPM-Solver-3.",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
382,
333,
395
],
"lines": [
{
"bbox": [
104,
378,
332,
399
],
"spans": [
{
"bbox": [
104,
378,
198,
399
],
"score": 1.0,
"content": "Require: initial value",
"type": "text"
},
{
"bbox": [
199,
385,
212,
394
],
"score": 0.84,
"content": "\\mathbf { \\nabla } _ { \\mathbf { \\mathcal { X } } T }",
"type": "inline_equation"
},
{
"bbox": [
213,
378,
259,
399
],
"score": 1.0,
"content": ", time steps",
"type": "text"
},
{
"bbox": [
259,
381,
290,
395
],
"score": 0.88,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
290,
378,
321,
399
],
"score": 1.0,
"content": ", model",
"type": "text"
},
{
"bbox": [
322,
385,
332,
394
],
"score": 0.67,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
}
],
"index": 11
}
],
"index": 11,
"bbox_fs": [
104,
378,
332,
399
]
},
{
"type": "interline_equation",
"bbox": [
136,
414,
483,
511
],
"lines": [
{
"bbox": [
136,
414,
483,
511
],
"spans": [
{
"bbox": [
136,
414,
483,
511
],
"score": 0.93,
"content": "\\begin{array} { r l } & { \\quad _ { s 2 i - 1 } _ { t _ { \\lambda } } ( \\overline { { \\lambda } } _ { t _ { i - 1 } } + r _ { 1 } h _ { i } ) , \\quad s _ { 2 i } t _ { \\lambda } ( \\lambda _ { t _ { i - 1 } } + r _ { 2 } h _ { i } ) } \\\\ & { u _ { 2 i - 1 } \\frac { \\alpha _ { s _ { 2 i - 1 } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { s _ { 2 i - 1 } } ( e ^ { r _ { 1 } h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { D _ { 2 i - 1 } \\epsilon _ { \\theta } ( u _ { 2 i - 1 } , s _ { 2 i - 1 } ) - \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { u _ { 2 i } \\frac { \\alpha _ { s _ { 2 i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { s _ { 2 i } } ( e ^ { r _ { 2 } h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) - \\frac { \\sigma _ { s _ { 2 i } } r _ { 2 } } { r _ { 1 } } ( \\frac { e ^ { r _ { 2 } h _ { i } } - 1 } { r _ { 2 } h _ { i } } - 1 ) D _ { 2 i - 1 } } \\\\ & { D _ { 2 i } \\epsilon _ { \\theta } ( u _ { 2 i } , s _ { 2 i } ) - \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) } \\\\ & { \\tilde { x } _ { t _ { i } } \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { x } _ { t _ { i - 1 } } - \\sigma _ { t _ { i } } ( e ^ { h _ { i } } - 1 ) \\epsilon _ { \\theta } ( \\tilde { x } _ { t _ { i - 1 } } , t _ { i - 1 } ) - \\frac { \\sigma _ { t _ { i } } } { r _ { 2 } } ( \\frac { e ^ { h _ { i } } - 1 } { h } - 1 ) D _ { 2 i } } \\end{array}",
"type": "interline_equation",
"image_path": "ad4609832a6419c85fb2132d9d91d2c95beec92e10bd4907c45b4ed7d6c517e6.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
136,
414,
483,
446.3333333333333
],
"spans": [],
"index": 12
},
{
"bbox": [
136,
446.3333333333333,
483,
478.66666666666663
],
"spans": [],
"index": 13
},
{
"bbox": [
136,
478.66666666666663,
483,
510.99999999999994
],
"spans": [],
"index": 14
}
]
},
{
"type": "text",
"bbox": [
108,
519,
171,
531
],
"lines": [
{
"bbox": [
106,
515,
171,
534
],
"spans": [
{
"bbox": [
106,
515,
153,
534
],
"score": 1.0,
"content": "10: return",
"type": "text"
},
{
"bbox": [
153,
519,
171,
531
],
"score": 0.81,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } }",
"type": "inline_equation"
}
],
"index": 15
}
],
"index": 15,
"bbox_fs": [
106,
515,
171,
534
]
},
{
"type": "text",
"bbox": [
106,
553,
506,
631
],
"lines": [
{
"bbox": [
106,
554,
506,
566
],
"spans": [
{
"bbox": [
106,
554,
130,
566
],
"score": 1.0,
"content": "Here,",
"type": "text"
},
{
"bbox": [
130,
554,
151,
566
],
"score": 0.9,
"content": "t _ { \\lambda } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
151,
554,
252,
566
],
"score": 1.0,
"content": "is the inverse function of",
"type": "text"
},
{
"bbox": [
252,
554,
270,
566
],
"score": 0.91,
"content": "\\lambda ( t )",
"type": "inline_equation"
},
{
"bbox": [
270,
554,
506,
566
],
"score": 1.0,
"content": ", which has an analytical formulation for the practical noise",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
565,
506,
577
],
"spans": [
{
"bbox": [
105,
565,
458,
577
],
"score": 1.0,
"content": "schedule used in [2, 16], as shown in Appendix D. The chosen intermediate points are",
"type": "text"
},
{
"bbox": [
459,
565,
489,
576
],
"score": 0.89,
"content": "( s _ { i } , \\pmb { u } _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
490,
565,
506,
577
],
"score": 1.0,
"content": "for",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
574,
506,
590
],
"spans": [
{
"bbox": [
104,
574,
188,
590
],
"score": 1.0,
"content": "DPM-Solver-2 and",
"type": "text"
},
{
"bbox": [
188,
576,
247,
588
],
"score": 0.92,
"content": "\\left( s _ { 2 i - 1 } , { \\pmb u } _ { 2 i - 1 } \\right)",
"type": "inline_equation"
},
{
"bbox": [
248,
574,
267,
590
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
268,
576,
306,
588
],
"score": 0.92,
"content": "( s _ { 2 i } , { \\pmb u } _ { 2 i } )",
"type": "inline_equation"
},
{
"bbox": [
307,
574,
506,
590
],
"score": 1.0,
"content": "for DPM-Solver-3. As shown in the algorithm,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
586,
505,
599
],
"spans": [
{
"bbox": [
105,
586,
161,
599
],
"score": 1.0,
"content": "DPM-Solver-",
"type": "text"
},
{
"bbox": [
161,
587,
168,
597
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
168,
586,
203,
599
],
"score": 1.0,
"content": "requires",
"type": "text"
},
{
"bbox": [
204,
588,
210,
597
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
211,
586,
343,
599
],
"score": 1.0,
"content": "function evaluations per step for",
"type": "text"
},
{
"bbox": [
344,
587,
387,
598
],
"score": 0.93,
"content": "k = 1 , 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
388,
586,
505,
599
],
"score": 1.0,
"content": ". Despite the more expensive",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
596,
506,
611
],
"spans": [
{
"bbox": [
105,
596,
215,
611
],
"score": 1.0,
"content": "steps, higher-order solvers",
"type": "text"
},
{
"bbox": [
216,
598,
253,
609
],
"score": 0.83,
"content": "( k = 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
253,
596,
506,
611
],
"score": 1.0,
"content": ") are usually more efficient since they require much fewer steps",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
609,
506,
621
],
"spans": [
{
"bbox": [
105,
609,
418,
621
],
"score": 1.0,
"content": "to converge, due to their higher convergence order. We show that DPM-Solver-",
"type": "text"
},
{
"bbox": [
418,
609,
425,
618
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
425,
609,
435,
621
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
435,
609,
442,
618
],
"score": 0.84,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
442,
609,
506,
621
],
"score": 1.0,
"content": "-th-order solver,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
619,
362,
632
],
"spans": [
{
"bbox": [
105,
619,
362,
632
],
"score": 1.0,
"content": "as stated in the following theorem. The proof is in Appendix B.",
"type": "text"
}
],
"index": 22
}
],
"index": 19,
"bbox_fs": [
104,
554,
506,
632
]
},
{
"type": "text",
"bbox": [
106,
634,
505,
682
],
"lines": [
{
"bbox": [
105,
633,
506,
647
],
"spans": [
{
"bbox": [
105,
633,
218,
647
],
"score": 1.0,
"content": "Theorem 3.2 (DPM-Solver-",
"type": "text"
},
{
"bbox": [
219,
635,
225,
644
],
"score": 0.8,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
226,
633,
244,
647
],
"score": 1.0,
"content": "as a",
"type": "text"
},
{
"bbox": [
244,
635,
250,
644
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
251,
633,
354,
647
],
"score": 1.0,
"content": "-th-order solver). Assume",
"type": "text"
},
{
"bbox": [
354,
634,
390,
646
],
"score": 0.92,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
391,
633,
506,
647
],
"score": 1.0,
"content": "follows the regularity condi-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
646,
505,
658
],
"spans": [
{
"bbox": [
106,
646,
262,
658
],
"score": 1.0,
"content": "tions detailed in Appendix B.1, then for",
"type": "text"
},
{
"bbox": [
262,
646,
306,
657
],
"score": 0.78,
"content": "k = 1 , 2 , 3",
"type": "inline_equation"
},
{
"bbox": [
306,
646,
362,
658
],
"score": 1.0,
"content": ", DPM-Solver-",
"type": "text"
},
{
"bbox": [
362,
647,
369,
655
],
"score": 0.78,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
369,
646,
386,
658
],
"score": 1.0,
"content": "is a",
"type": "text"
},
{
"bbox": [
386,
646,
393,
655
],
"score": 0.72,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
393,
646,
505,
658
],
"score": 1.0,
"content": "-th order solver for diffusion",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
655,
507,
671
],
"spans": [
{
"bbox": [
105,
655,
218,
671
],
"score": 1.0,
"content": "ODEs, i.e., for the sequence",
"type": "text"
},
{
"bbox": [
218,
655,
256,
668
],
"score": 0.92,
"content": "\\{ \\tilde { \\pmb { x } } _ { t _ { i } } \\} _ { i = 1 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
256,
655,
362,
671
],
"score": 1.0,
"content": "computed by DPM-Solver-",
"type": "text"
},
{
"bbox": [
362,
657,
368,
667
],
"score": 0.71,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
369,
655,
507,
671
],
"score": 1.0,
"content": ", the approximation error at time 0",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
667,
400,
683
],
"spans": [
{
"bbox": [
104,
667,
140,
683
],
"score": 1.0,
"content": "satisfies",
"type": "text"
},
{
"bbox": [
141,
668,
232,
680
],
"score": 0.93,
"content": "\\tilde { \\pmb { x } } _ { t _ { M } } - \\pmb { x } _ { 0 } = \\mathcal { O } ( h _ { \\operatorname* { m a x } } ^ { k } )",
"type": "inline_equation"
},
{
"bbox": [
232,
667,
263,
683
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
263,
668,
395,
681
],
"score": 0.92,
"content": "h _ { m a x } = \\mathrm { m a x } _ { 1 \\leq i \\leq M } ( \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } } )",
"type": "inline_equation"
},
{
"bbox": [
395,
667,
400,
683
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 26
}
],
"index": 24.5,
"bbox_fs": [
104,
633,
507,
683
]
},
{
"type": "text",
"bbox": [
107,
688,
505,
722
],
"lines": [
{
"bbox": [
106,
688,
505,
701
],
"spans": [
{
"bbox": [
106,
688,
189,
701
],
"score": 1.0,
"content": "Finally, solvers with",
"type": "text"
},
{
"bbox": [
189,
689,
214,
700
],
"score": 0.91,
"content": "k \\geq 4",
"type": "inline_equation"
},
{
"bbox": [
214,
688,
505,
701
],
"score": 1.0,
"content": "need much more intermediate points as shown by previous work [31, 32]",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
700,
505,
712
],
"spans": [
{
"bbox": [
105,
700,
329,
712
],
"score": 1.0,
"content": "for exponential integrators. Therefore, we only consider",
"type": "text"
},
{
"bbox": [
329,
700,
335,
710
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
336,
700,
505,
712
],
"score": 1.0,
"content": "from 1 to 3 in this work, while leaving the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
710,
259,
723
],
"spans": [
{
"bbox": [
106,
710,
185,
723
],
"score": 1.0,
"content": "solvers with higher",
"type": "text"
},
{
"bbox": [
185,
711,
192,
721
],
"score": 0.81,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
192,
710,
259,
723
],
"score": 1.0,
"content": "for future study.",
"type": "text"
}
],
"index": 29
}
],
"index": 28,
"bbox_fs": [
105,
688,
505,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
72,
210,
84
],
"lines": [
{
"bbox": [
105,
71,
211,
86
],
"spans": [
{
"bbox": [
105,
71,
211,
86
],
"score": 1.0,
"content": "3.3 Step Size Schedule",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
92,
505,
181
],
"lines": [
{
"bbox": [
102,
90,
508,
110
],
"spans": [
{
"bbox": [
102,
90,
357,
110
],
"score": 1.0,
"content": "The proposed solvers in Sec. 3.2 need to specify the time steps",
"type": "text"
},
{
"bbox": [
358,
92,
389,
105
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
389,
90,
508,
110
],
"score": 1.0,
"content": "in advance. We propose two",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
104,
505,
116
],
"spans": [
{
"bbox": [
106,
104,
505,
116
],
"score": 1.0,
"content": "choices of the time step schedule. One choice is handcrafted, which is to uniformly split the interval",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
107,
114,
506,
128
],
"spans": [
{
"bbox": [
107,
114,
142,
126
],
"score": 0.88,
"content": "[ \\lambda _ { T } , \\lambda _ { 0 } ]",
"type": "inline_equation"
},
{
"bbox": [
142,
114,
163,
128
],
"score": 1.0,
"content": ", i.e.",
"type": "text"
},
{
"bbox": [
163,
114,
271,
128
],
"score": 0.85,
"content": "\\begin{array} { r } { \\lambda _ { t _ { i } } = \\dot { \\lambda _ { T } } + \\frac { i } { M } ( \\lambda _ { 0 } - \\lambda _ { T } ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
271,
114,
275,
128
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
276,
115,
333,
126
],
"score": 0.83,
"content": "i = 0 , \\ldots , M",
"type": "inline_equation"
},
{
"bbox": [
333,
114,
506,
128
],
"score": 1.0,
"content": ". Note that this is different from previous",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
506,
139
],
"spans": [
{
"bbox": [
105,
124,
287,
139
],
"score": 1.0,
"content": "work [2, 3] which chooses uniform steps for",
"type": "text"
},
{
"bbox": [
288,
127,
295,
137
],
"score": 0.85,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
296,
124,
506,
139
],
"score": 1.0,
"content": ". Empirically, DPM-Solver with uniform time steps",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
136,
506,
150
],
"spans": [
{
"bbox": [
106,
137,
120,
148
],
"score": 0.89,
"content": "\\lambda _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
120,
136,
506,
150
],
"score": 1.0,
"content": "can already generate quite good samples in few steps, where results are listed in Appendix E. As",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
146,
506,
161
],
"spans": [
{
"bbox": [
105,
146,
506,
161
],
"score": 1.0,
"content": "the other choice, we propose an adaptive step size algorithm, which dynamically adjusts the step size",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
159,
505,
171
],
"spans": [
{
"bbox": [
106,
159,
505,
171
],
"score": 1.0,
"content": "by combining different orders of DPM-Solver. The adaptive algorithm is inspired by [20] and we",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
169,
298,
182
],
"spans": [
{
"bbox": [
106,
169,
298,
182
],
"score": 1.0,
"content": "defer its implementation details to Appendix C.",
"type": "text"
}
],
"index": 8
}
],
"index": 4.5
},
{
"type": "text",
"bbox": [
107,
185,
505,
241
],
"lines": [
{
"bbox": [
106,
186,
505,
198
],
"spans": [
{
"bbox": [
106,
186,
505,
198
],
"score": 1.0,
"content": "For few-step sampling, we need to use up all the number of function evaluations (NFE). When the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
196,
505,
210
],
"spans": [
{
"bbox": [
105,
196,
505,
210
],
"score": 1.0,
"content": "NFE is not divisible by 3, we firstly apply DPM-Solver-3 as much as possible, and then add a single",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
207,
505,
220
],
"spans": [
{
"bbox": [
105,
207,
391,
220
],
"score": 1.0,
"content": "step of DPM-Solver-1 or DPM-Solver-2 (dependent on the reminder of",
"type": "text"
},
{
"bbox": [
391,
208,
402,
217
],
"score": 0.83,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
402,
207,
505,
220
],
"score": 1.0,
"content": "divided by 3), as detailed",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
218,
505,
231
],
"spans": [
{
"bbox": [
105,
218,
505,
231
],
"score": 1.0,
"content": "in Appendix D. In the subsequent experiments, we use such combination of solvers with the uniform",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
229,
426,
243
],
"spans": [
{
"bbox": [
105,
229,
194,
243
],
"score": 1.0,
"content": "step size schedule for",
"type": "text"
},
{
"bbox": [
194,
230,
237,
240
],
"score": 0.77,
"content": "\\mathrm { N F E } \\leq 2 0",
"type": "inline_equation"
},
{
"bbox": [
237,
229,
426,
243
],
"score": 1.0,
"content": ", and otherwise the adaptive step size schedule.",
"type": "text"
}
],
"index": 13
}
],
"index": 11
},
{
"type": "title",
"bbox": [
108,
254,
286,
266
],
"lines": [
{
"bbox": [
105,
253,
287,
268
],
"spans": [
{
"bbox": [
105,
253,
287,
268
],
"score": 1.0,
"content": "3.4 Sampling from Discrete-Time DPMs",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "text",
"bbox": [
106,
273,
506,
365
],
"lines": [
{
"bbox": [
102,
268,
510,
293
],
"spans": [
{
"bbox": [
102,
268,
353,
293
],
"score": 1.0,
"content": "Discrete-time DPMs [2] train the noise prediction model at",
"type": "text"
},
{
"bbox": [
353,
275,
363,
284
],
"score": 0.83,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
363,
268,
432,
293
],
"score": 1.0,
"content": "fixed time steps",
"type": "text"
},
{
"bbox": [
433,
273,
468,
286
],
"score": 0.93,
"content": "\\{ t _ { n } \\} _ { n = 1 } ^ { N }",
"type": "inline_equation"
},
{
"bbox": [
468,
268,
510,
293
],
"score": 1.0,
"content": ", and the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
285,
506,
298
],
"spans": [
{
"bbox": [
105,
285,
289,
298
],
"score": 1.0,
"content": "noise prediction model is parameterized by",
"type": "text"
},
{
"bbox": [
289,
286,
330,
298
],
"score": 0.92,
"content": "\\tilde { \\epsilon } _ { \\theta } ( { \\boldsymbol x } _ { n } , n )",
"type": "inline_equation"
},
{
"bbox": [
330,
285,
347,
298
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
347,
286,
425,
297
],
"score": 0.92,
"content": "n = 0 , \\ldots , N - 1",
"type": "inline_equation"
},
{
"bbox": [
426,
285,
480,
298
],
"score": 1.0,
"content": ", where each",
"type": "text"
},
{
"bbox": [
480,
287,
493,
297
],
"score": 0.86,
"content": "{ \\pmb x } _ { n }",
"type": "inline_equation"
},
{
"bbox": [
494,
285,
506,
298
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
297,
505,
308
],
"spans": [
{
"bbox": [
105,
297,
240,
308
],
"score": 1.0,
"content": "corresponding to the value at time",
"type": "text"
},
{
"bbox": [
241,
297,
261,
308
],
"score": 0.9,
"content": "t _ { n + 1 }",
"type": "inline_equation"
},
{
"bbox": [
261,
297,
505,
308
],
"score": 1.0,
"content": ". We can transform the discrete-time noise prediction model to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
305,
506,
325
],
"spans": [
{
"bbox": [
105,
305,
238,
325
],
"score": 1.0,
"content": "the continuous version by letting",
"type": "text"
},
{
"bbox": [
238,
308,
343,
323
],
"score": 0.91,
"content": "\\begin{array} { r } { \\epsilon _ { \\theta } ( x , t ) : = \\tilde { \\epsilon } _ { \\theta } ( x , \\frac { ( N - 1 ) t } { T } ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
343,
305,
373,
325
],
"score": 1.0,
"content": ", for all",
"type": "text"
},
{
"bbox": [
374,
308,
447,
322
],
"score": 0.93,
"content": "\\pmb { x } \\in \\mathbb { R } ^ { d } , t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
448,
305,
506,
325
],
"score": 1.0,
"content": ". Note that the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
320,
506,
333
],
"spans": [
{
"bbox": [
105,
320,
160,
333
],
"score": 1.0,
"content": "input time of",
"type": "text"
},
{
"bbox": [
160,
321,
171,
332
],
"score": 0.87,
"content": "\\tilde { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
171,
320,
506,
333
],
"score": 1.0,
"content": "may not be integers, but we find that the noise prediction model can still work well,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
331,
506,
344
],
"spans": [
{
"bbox": [
105,
331,
506,
344
],
"score": 1.0,
"content": "and we hypothesize that it is because of the smooth time embeddings (e.g., position embeddings [2]).",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
343,
506,
355
],
"spans": [
{
"bbox": [
106,
343,
506,
355
],
"score": 1.0,
"content": "By such reparameterization, the noise prediction model can adopt the continuous-time steps as input,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
352,
333,
366
],
"spans": [
{
"bbox": [
105,
352,
333,
366
],
"score": 1.0,
"content": "and thus we can also use DPM-Solver for fast sampling.",
"type": "text"
}
],
"index": 22
}
],
"index": 18.5
},
{
"type": "title",
"bbox": [
107,
380,
384,
395
],
"lines": [
{
"bbox": [
104,
379,
385,
398
],
"spans": [
{
"bbox": [
104,
379,
385,
398
],
"score": 1.0,
"content": "4 Comparison with Existing Fast Sampling Methods",
"type": "text"
}
],
"index": 23
}
],
"index": 23
},
{
"type": "text",
"bbox": [
106,
405,
505,
439
],
"lines": [
{
"bbox": [
105,
404,
505,
419
],
"spans": [
{
"bbox": [
105,
404,
505,
419
],
"score": 1.0,
"content": "Here, we discuss the relationship and highlight the difference between DPM-Solver and existing",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
416,
505,
429
],
"spans": [
{
"bbox": [
106,
416,
505,
429
],
"score": 1.0,
"content": "ODE-based fast sampling methods for DPMs. We further briefly discuss the advantage of training-free",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
428,
271,
441
],
"spans": [
{
"bbox": [
106,
428,
271,
441
],
"score": 1.0,
"content": "samplers over those training-based ones.",
"type": "text"
}
],
"index": 26
}
],
"index": 25
},
{
"type": "title",
"bbox": [
107,
452,
232,
464
],
"lines": [
{
"bbox": [
106,
452,
233,
465
],
"spans": [
{
"bbox": [
106,
452,
233,
465
],
"score": 1.0,
"content": "4.1 DDIM as DPM-Solver-1",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
107,
472,
505,
506
],
"lines": [
{
"bbox": [
105,
471,
505,
486
],
"spans": [
{
"bbox": [
105,
471,
505,
486
],
"score": 1.0,
"content": "Denoising Diffusion Implicit Models (DDIM) [19] design a deterministic method for fast sampling",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
483,
505,
497
],
"spans": [
{
"bbox": [
105,
483,
273,
497
],
"score": 1.0,
"content": "from DPMs. For two adjacent time steps",
"type": "text"
},
{
"bbox": [
274,
485,
292,
496
],
"score": 0.89,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
292,
483,
310,
497
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
310,
485,
318,
495
],
"score": 0.86,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
319,
483,
450,
497
],
"score": 1.0,
"content": ", assume that we have a solution",
"type": "text"
},
{
"bbox": [
451,
484,
473,
497
],
"score": 0.92,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
473,
483,
505,
497
],
"score": 1.0,
"content": "at time",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
107,
495,
351,
507
],
"spans": [
{
"bbox": [
107,
496,
124,
506
],
"score": 0.86,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
125,
495,
282,
507
],
"score": 1.0,
"content": ", then a single step of DDIM from time",
"type": "text"
},
{
"bbox": [
282,
496,
300,
506
],
"score": 0.9,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
300,
495,
331,
507
],
"score": 1.0,
"content": "to time",
"type": "text"
},
{
"bbox": [
332,
496,
340,
506
],
"score": 0.86,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
340,
495,
351,
507
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 30
}
],
"index": 29
},
{
"type": "interline_equation",
"bbox": [
188,
512,
423,
540
],
"lines": [
{
"bbox": [
188,
512,
423,
540
],
"spans": [
{
"bbox": [
188,
512,
423,
540
],
"score": 0.93,
"content": "\\tilde { \\pmb { x } } _ { t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\left( \\frac { \\sigma _ { t _ { i - 1 } } } { \\alpha _ { t _ { i - 1 } } } - \\frac { \\sigma _ { t _ { i } } } { \\alpha _ { t _ { i } } } \\right) \\epsilon _ { \\theta } ( \\tilde { \\pmb { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) .",
"type": "interline_equation",
"image_path": "3ad93903d54c31c8256e52d597943286bbf9fba385901a6d278538f4eb730eed.jpg"
}
]
}
],
"index": 31.5,
"virtual_lines": [
{
"bbox": [
188,
512,
423,
526.0
],
"spans": [],
"index": 31
},
{
"bbox": [
188,
526.0,
423,
540.0
],
"spans": [],
"index": 32
}
]
},
{
"type": "text",
"bbox": [
107,
545,
505,
606
],
"lines": [
{
"bbox": [
105,
545,
505,
557
],
"spans": [
{
"bbox": [
105,
545,
505,
557
],
"score": 1.0,
"content": "Although motivated by entirely different perspectives, we show that the updates of DPM-Solver-1",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
556,
505,
568
],
"spans": [
{
"bbox": [
105,
556,
460,
568
],
"score": 1.0,
"content": "and Denoising Diffusion Implicit Models (DDIM) [19] are identical. By the definition of",
"type": "text"
},
{
"bbox": [
460,
557,
466,
566
],
"score": 0.78,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
467,
556,
505,
568
],
"score": 1.0,
"content": ", we have",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
108,
566,
506,
585
],
"spans": [
{
"bbox": [
108,
567,
174,
585
],
"score": 0.92,
"content": "\\frac { \\sigma _ { t _ { i - 1 } } } { \\alpha _ { t _ { i - 1 } } } = e ^ { - \\lambda _ { t _ { i - 1 } } ^ { - } }",
"type": "inline_equation"
},
{
"bbox": [
174,
566,
193,
581
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
193,
568,
243,
585
],
"score": 0.91,
"content": "\\begin{array} { r } { \\frac { \\sigma _ { t _ { i } } } { \\alpha _ { t _ { i } } } = e ^ { - \\lambda _ { t _ { i } } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
243,
567,
324,
583
],
"score": 1.0,
"content": ". Plugging these and",
"type": "text"
},
{
"bbox": [
324,
569,
390,
582
],
"score": 0.93,
"content": "h _ { i } = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
390,
567,
506,
583
],
"score": 1.0,
"content": "to Eq. (4.1) results in exactly",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
584,
506,
596
],
"spans": [
{
"bbox": [
105,
584,
506,
596
],
"score": 1.0,
"content": "a step of DPM-Solver-1 in Eq. (3.7). However, the semi-linear ODE formulation of DPM-Solver",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
595,
473,
607
],
"spans": [
{
"bbox": [
105,
595,
473,
607
],
"score": 1.0,
"content": "allows for principled generalization to higher-order solvers and convergence order analysis.",
"type": "text"
}
],
"index": 37
}
],
"index": 35
},
{
"type": "text",
"bbox": [
107,
610,
506,
666
],
"lines": [
{
"bbox": [
106,
611,
506,
623
],
"spans": [
{
"bbox": [
106,
611,
506,
623
],
"score": 1.0,
"content": "Recent work [13] also show that DDIM is a first-order discretization of diffusion ODEs by differenti-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
621,
505,
633
],
"spans": [
{
"bbox": [
106,
621,
505,
633
],
"score": 1.0,
"content": "ating both sides of Eq. (4.1). However, they cannot explain the difference between DDIM and the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
632,
506,
645
],
"spans": [
{
"bbox": [
106,
632,
506,
645
],
"score": 1.0,
"content": "first-order Euler discretization of diffusion ODEs. In contrast, by showing that DDIM is a special",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
644,
506,
657
],
"spans": [
{
"bbox": [
106,
644,
506,
657
],
"score": 1.0,
"content": "case of DPM-Solver, we reveal that DDIM makes full use of the semi-linearity of diffusion ODEs,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
654,
352,
667
],
"spans": [
{
"bbox": [
106,
654,
352,
667
],
"score": 1.0,
"content": "which explains its superiority over traditional Euler methods.",
"type": "text"
}
],
"index": 42
}
],
"index": 40
},
{
"type": "title",
"bbox": [
106,
679,
353,
691
],
"lines": [
{
"bbox": [
104,
677,
354,
695
],
"spans": [
{
"bbox": [
104,
677,
354,
695
],
"score": 1.0,
"content": "4.2 Comparison with Traditional Runge-Kutta Methods",
"type": "text"
}
],
"index": 43
}
],
"index": 43
},
{
"type": "text",
"bbox": [
106,
699,
503,
722
],
"lines": [
{
"bbox": [
106,
699,
505,
712
],
"spans": [
{
"bbox": [
106,
699,
505,
712
],
"score": 1.0,
"content": "One can obtain a high-order solver by directly applying traditional explicit Runge-Kutta (RK) methods",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
709,
505,
724
],
"spans": [
{
"bbox": [
105,
709,
505,
724
],
"score": 1.0,
"content": "to the diffusion ODE in Eq. (2.7). Specifically, RK methods write the solution of Eq. (2.7) in the",
"type": "text"
}
],
"index": 45
}
],
"index": 44.5
}
],
"page_idx": 6,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
309,
750
],
"lines": [
{
"bbox": [
302,
741,
309,
752
],
"spans": [
{
"bbox": [
302,
741,
309,
752
],
"score": 1.0,
"content": "7",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
72,
210,
84
],
"lines": [
{
"bbox": [
105,
71,
211,
86
],
"spans": [
{
"bbox": [
105,
71,
211,
86
],
"score": 1.0,
"content": "3.3 Step Size Schedule",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
92,
505,
181
],
"lines": [
{
"bbox": [
102,
90,
508,
110
],
"spans": [
{
"bbox": [
102,
90,
357,
110
],
"score": 1.0,
"content": "The proposed solvers in Sec. 3.2 need to specify the time steps",
"type": "text"
},
{
"bbox": [
358,
92,
389,
105
],
"score": 0.93,
"content": "\\{ t _ { i } \\} _ { i = 0 } ^ { M }",
"type": "inline_equation"
},
{
"bbox": [
389,
90,
508,
110
],
"score": 1.0,
"content": "in advance. We propose two",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
104,
505,
116
],
"spans": [
{
"bbox": [
106,
104,
505,
116
],
"score": 1.0,
"content": "choices of the time step schedule. One choice is handcrafted, which is to uniformly split the interval",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
107,
114,
506,
128
],
"spans": [
{
"bbox": [
107,
114,
142,
126
],
"score": 0.88,
"content": "[ \\lambda _ { T } , \\lambda _ { 0 } ]",
"type": "inline_equation"
},
{
"bbox": [
142,
114,
163,
128
],
"score": 1.0,
"content": ", i.e.",
"type": "text"
},
{
"bbox": [
163,
114,
271,
128
],
"score": 0.85,
"content": "\\begin{array} { r } { \\lambda _ { t _ { i } } = \\dot { \\lambda _ { T } } + \\frac { i } { M } ( \\lambda _ { 0 } - \\lambda _ { T } ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
271,
114,
275,
128
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
276,
115,
333,
126
],
"score": 0.83,
"content": "i = 0 , \\ldots , M",
"type": "inline_equation"
},
{
"bbox": [
333,
114,
506,
128
],
"score": 1.0,
"content": ". Note that this is different from previous",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
506,
139
],
"spans": [
{
"bbox": [
105,
124,
287,
139
],
"score": 1.0,
"content": "work [2, 3] which chooses uniform steps for",
"type": "text"
},
{
"bbox": [
288,
127,
295,
137
],
"score": 0.85,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
296,
124,
506,
139
],
"score": 1.0,
"content": ". Empirically, DPM-Solver with uniform time steps",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
136,
506,
150
],
"spans": [
{
"bbox": [
106,
137,
120,
148
],
"score": 0.89,
"content": "\\lambda _ { t _ { i } }",
"type": "inline_equation"
},
{
"bbox": [
120,
136,
506,
150
],
"score": 1.0,
"content": "can already generate quite good samples in few steps, where results are listed in Appendix E. As",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
146,
506,
161
],
"spans": [
{
"bbox": [
105,
146,
506,
161
],
"score": 1.0,
"content": "the other choice, we propose an adaptive step size algorithm, which dynamically adjusts the step size",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
159,
505,
171
],
"spans": [
{
"bbox": [
106,
159,
505,
171
],
"score": 1.0,
"content": "by combining different orders of DPM-Solver. The adaptive algorithm is inspired by [20] and we",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
169,
298,
182
],
"spans": [
{
"bbox": [
106,
169,
298,
182
],
"score": 1.0,
"content": "defer its implementation details to Appendix C.",
"type": "text"
}
],
"index": 8
}
],
"index": 4.5,
"bbox_fs": [
102,
90,
508,
182
]
},
{
"type": "text",
"bbox": [
107,
185,
505,
241
],
"lines": [
{
"bbox": [
106,
186,
505,
198
],
"spans": [
{
"bbox": [
106,
186,
505,
198
],
"score": 1.0,
"content": "For few-step sampling, we need to use up all the number of function evaluations (NFE). When the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
196,
505,
210
],
"spans": [
{
"bbox": [
105,
196,
505,
210
],
"score": 1.0,
"content": "NFE is not divisible by 3, we firstly apply DPM-Solver-3 as much as possible, and then add a single",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
207,
505,
220
],
"spans": [
{
"bbox": [
105,
207,
391,
220
],
"score": 1.0,
"content": "step of DPM-Solver-1 or DPM-Solver-2 (dependent on the reminder of",
"type": "text"
},
{
"bbox": [
391,
208,
402,
217
],
"score": 0.83,
"content": "K",
"type": "inline_equation"
},
{
"bbox": [
402,
207,
505,
220
],
"score": 1.0,
"content": "divided by 3), as detailed",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
218,
505,
231
],
"spans": [
{
"bbox": [
105,
218,
505,
231
],
"score": 1.0,
"content": "in Appendix D. In the subsequent experiments, we use such combination of solvers with the uniform",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
229,
426,
243
],
"spans": [
{
"bbox": [
105,
229,
194,
243
],
"score": 1.0,
"content": "step size schedule for",
"type": "text"
},
{
"bbox": [
194,
230,
237,
240
],
"score": 0.77,
"content": "\\mathrm { N F E } \\leq 2 0",
"type": "inline_equation"
},
{
"bbox": [
237,
229,
426,
243
],
"score": 1.0,
"content": ", and otherwise the adaptive step size schedule.",
"type": "text"
}
],
"index": 13
}
],
"index": 11,
"bbox_fs": [
105,
186,
505,
243
]
},
{
"type": "title",
"bbox": [
108,
254,
286,
266
],
"lines": [
{
"bbox": [
105,
253,
287,
268
],
"spans": [
{
"bbox": [
105,
253,
287,
268
],
"score": 1.0,
"content": "3.4 Sampling from Discrete-Time DPMs",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "text",
"bbox": [
106,
273,
506,
365
],
"lines": [
{
"bbox": [
102,
268,
510,
293
],
"spans": [
{
"bbox": [
102,
268,
353,
293
],
"score": 1.0,
"content": "Discrete-time DPMs [2] train the noise prediction model at",
"type": "text"
},
{
"bbox": [
353,
275,
363,
284
],
"score": 0.83,
"content": "N",
"type": "inline_equation"
},
{
"bbox": [
363,
268,
432,
293
],
"score": 1.0,
"content": "fixed time steps",
"type": "text"
},
{
"bbox": [
433,
273,
468,
286
],
"score": 0.93,
"content": "\\{ t _ { n } \\} _ { n = 1 } ^ { N }",
"type": "inline_equation"
},
{
"bbox": [
468,
268,
510,
293
],
"score": 1.0,
"content": ", and the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
285,
506,
298
],
"spans": [
{
"bbox": [
105,
285,
289,
298
],
"score": 1.0,
"content": "noise prediction model is parameterized by",
"type": "text"
},
{
"bbox": [
289,
286,
330,
298
],
"score": 0.92,
"content": "\\tilde { \\epsilon } _ { \\theta } ( { \\boldsymbol x } _ { n } , n )",
"type": "inline_equation"
},
{
"bbox": [
330,
285,
347,
298
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
347,
286,
425,
297
],
"score": 0.92,
"content": "n = 0 , \\ldots , N - 1",
"type": "inline_equation"
},
{
"bbox": [
426,
285,
480,
298
],
"score": 1.0,
"content": ", where each",
"type": "text"
},
{
"bbox": [
480,
287,
493,
297
],
"score": 0.86,
"content": "{ \\pmb x } _ { n }",
"type": "inline_equation"
},
{
"bbox": [
494,
285,
506,
298
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
297,
505,
308
],
"spans": [
{
"bbox": [
105,
297,
240,
308
],
"score": 1.0,
"content": "corresponding to the value at time",
"type": "text"
},
{
"bbox": [
241,
297,
261,
308
],
"score": 0.9,
"content": "t _ { n + 1 }",
"type": "inline_equation"
},
{
"bbox": [
261,
297,
505,
308
],
"score": 1.0,
"content": ". We can transform the discrete-time noise prediction model to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
305,
506,
325
],
"spans": [
{
"bbox": [
105,
305,
238,
325
],
"score": 1.0,
"content": "the continuous version by letting",
"type": "text"
},
{
"bbox": [
238,
308,
343,
323
],
"score": 0.91,
"content": "\\begin{array} { r } { \\epsilon _ { \\theta } ( x , t ) : = \\tilde { \\epsilon } _ { \\theta } ( x , \\frac { ( N - 1 ) t } { T } ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
343,
305,
373,
325
],
"score": 1.0,
"content": ", for all",
"type": "text"
},
{
"bbox": [
374,
308,
447,
322
],
"score": 0.93,
"content": "\\pmb { x } \\in \\mathbb { R } ^ { d } , t \\in [ 0 , T ]",
"type": "inline_equation"
},
{
"bbox": [
448,
305,
506,
325
],
"score": 1.0,
"content": ". Note that the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
320,
506,
333
],
"spans": [
{
"bbox": [
105,
320,
160,
333
],
"score": 1.0,
"content": "input time of",
"type": "text"
},
{
"bbox": [
160,
321,
171,
332
],
"score": 0.87,
"content": "\\tilde { \\epsilon } _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
171,
320,
506,
333
],
"score": 1.0,
"content": "may not be integers, but we find that the noise prediction model can still work well,",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
331,
506,
344
],
"spans": [
{
"bbox": [
105,
331,
506,
344
],
"score": 1.0,
"content": "and we hypothesize that it is because of the smooth time embeddings (e.g., position embeddings [2]).",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
343,
506,
355
],
"spans": [
{
"bbox": [
106,
343,
506,
355
],
"score": 1.0,
"content": "By such reparameterization, the noise prediction model can adopt the continuous-time steps as input,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
352,
333,
366
],
"spans": [
{
"bbox": [
105,
352,
333,
366
],
"score": 1.0,
"content": "and thus we can also use DPM-Solver for fast sampling.",
"type": "text"
}
],
"index": 22
}
],
"index": 18.5,
"bbox_fs": [
102,
268,
510,
366
]
},
{
"type": "title",
"bbox": [
107,
380,
384,
395
],
"lines": [
{
"bbox": [
104,
379,
385,
398
],
"spans": [
{
"bbox": [
104,
379,
385,
398
],
"score": 1.0,
"content": "4 Comparison with Existing Fast Sampling Methods",
"type": "text"
}
],
"index": 23
}
],
"index": 23
},
{
"type": "text",
"bbox": [
106,
405,
505,
439
],
"lines": [
{
"bbox": [
105,
404,
505,
419
],
"spans": [
{
"bbox": [
105,
404,
505,
419
],
"score": 1.0,
"content": "Here, we discuss the relationship and highlight the difference between DPM-Solver and existing",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
416,
505,
429
],
"spans": [
{
"bbox": [
106,
416,
505,
429
],
"score": 1.0,
"content": "ODE-based fast sampling methods for DPMs. We further briefly discuss the advantage of training-free",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
428,
271,
441
],
"spans": [
{
"bbox": [
106,
428,
271,
441
],
"score": 1.0,
"content": "samplers over those training-based ones.",
"type": "text"
}
],
"index": 26
}
],
"index": 25,
"bbox_fs": [
105,
404,
505,
441
]
},
{
"type": "title",
"bbox": [
107,
452,
232,
464
],
"lines": [
{
"bbox": [
106,
452,
233,
465
],
"spans": [
{
"bbox": [
106,
452,
233,
465
],
"score": 1.0,
"content": "4.1 DDIM as DPM-Solver-1",
"type": "text"
}
],
"index": 27
}
],
"index": 27
},
{
"type": "text",
"bbox": [
107,
472,
505,
506
],
"lines": [
{
"bbox": [
105,
471,
505,
486
],
"spans": [
{
"bbox": [
105,
471,
505,
486
],
"score": 1.0,
"content": "Denoising Diffusion Implicit Models (DDIM) [19] design a deterministic method for fast sampling",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
483,
505,
497
],
"spans": [
{
"bbox": [
105,
483,
273,
497
],
"score": 1.0,
"content": "from DPMs. For two adjacent time steps",
"type": "text"
},
{
"bbox": [
274,
485,
292,
496
],
"score": 0.89,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
292,
483,
310,
497
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
310,
485,
318,
495
],
"score": 0.86,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
319,
483,
450,
497
],
"score": 1.0,
"content": ", assume that we have a solution",
"type": "text"
},
{
"bbox": [
451,
484,
473,
497
],
"score": 0.92,
"content": "\\tilde { \\boldsymbol { x } } _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
473,
483,
505,
497
],
"score": 1.0,
"content": "at time",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
107,
495,
351,
507
],
"spans": [
{
"bbox": [
107,
496,
124,
506
],
"score": 0.86,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
125,
495,
282,
507
],
"score": 1.0,
"content": ", then a single step of DDIM from time",
"type": "text"
},
{
"bbox": [
282,
496,
300,
506
],
"score": 0.9,
"content": "t _ { i - 1 }",
"type": "inline_equation"
},
{
"bbox": [
300,
495,
331,
507
],
"score": 1.0,
"content": "to time",
"type": "text"
},
{
"bbox": [
332,
496,
340,
506
],
"score": 0.86,
"content": "t _ { i }",
"type": "inline_equation"
},
{
"bbox": [
340,
495,
351,
507
],
"score": 1.0,
"content": "is",
"type": "text"
}
],
"index": 30
}
],
"index": 29,
"bbox_fs": [
105,
471,
505,
507
]
},
{
"type": "interline_equation",
"bbox": [
188,
512,
423,
540
],
"lines": [
{
"bbox": [
188,
512,
423,
540
],
"spans": [
{
"bbox": [
188,
512,
423,
540
],
"score": 0.93,
"content": "\\tilde { \\pmb { x } } _ { t _ { i } } = \\frac { \\alpha _ { t _ { i } } } { \\alpha _ { t _ { i - 1 } } } \\tilde { \\pmb { x } } _ { t _ { i - 1 } } - \\alpha _ { t _ { i } } \\left( \\frac { \\sigma _ { t _ { i - 1 } } } { \\alpha _ { t _ { i - 1 } } } - \\frac { \\sigma _ { t _ { i } } } { \\alpha _ { t _ { i } } } \\right) \\epsilon _ { \\theta } ( \\tilde { \\pmb { x } } _ { t _ { i - 1 } } , t _ { i - 1 } ) .",
"type": "interline_equation",
"image_path": "3ad93903d54c31c8256e52d597943286bbf9fba385901a6d278538f4eb730eed.jpg"
}
]
}
],
"index": 31.5,
"virtual_lines": [
{
"bbox": [
188,
512,
423,
526.0
],
"spans": [],
"index": 31
},
{
"bbox": [
188,
526.0,
423,
540.0
],
"spans": [],
"index": 32
}
]
},
{
"type": "text",
"bbox": [
107,
545,
505,
606
],
"lines": [
{
"bbox": [
105,
545,
505,
557
],
"spans": [
{
"bbox": [
105,
545,
505,
557
],
"score": 1.0,
"content": "Although motivated by entirely different perspectives, we show that the updates of DPM-Solver-1",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
556,
505,
568
],
"spans": [
{
"bbox": [
105,
556,
460,
568
],
"score": 1.0,
"content": "and Denoising Diffusion Implicit Models (DDIM) [19] are identical. By the definition of",
"type": "text"
},
{
"bbox": [
460,
557,
466,
566
],
"score": 0.78,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
467,
556,
505,
568
],
"score": 1.0,
"content": ", we have",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
108,
566,
506,
585
],
"spans": [
{
"bbox": [
108,
567,
174,
585
],
"score": 0.92,
"content": "\\frac { \\sigma _ { t _ { i - 1 } } } { \\alpha _ { t _ { i - 1 } } } = e ^ { - \\lambda _ { t _ { i - 1 } } ^ { - } }",
"type": "inline_equation"
},
{
"bbox": [
174,
566,
193,
581
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
193,
568,
243,
585
],
"score": 0.91,
"content": "\\begin{array} { r } { \\frac { \\sigma _ { t _ { i } } } { \\alpha _ { t _ { i } } } = e ^ { - \\lambda _ { t _ { i } } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
243,
567,
324,
583
],
"score": 1.0,
"content": ". Plugging these and",
"type": "text"
},
{
"bbox": [
324,
569,
390,
582
],
"score": 0.93,
"content": "h _ { i } = \\lambda _ { t _ { i } } - \\lambda _ { t _ { i - 1 } }",
"type": "inline_equation"
},
{
"bbox": [
390,
567,
506,
583
],
"score": 1.0,
"content": "to Eq. (4.1) results in exactly",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
584,
506,
596
],
"spans": [
{
"bbox": [
105,
584,
506,
596
],
"score": 1.0,
"content": "a step of DPM-Solver-1 in Eq. (3.7). However, the semi-linear ODE formulation of DPM-Solver",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
595,
473,
607
],
"spans": [
{
"bbox": [
105,
595,
473,
607
],
"score": 1.0,
"content": "allows for principled generalization to higher-order solvers and convergence order analysis.",
"type": "text"
}
],
"index": 37
}
],
"index": 35,
"bbox_fs": [
105,
545,
506,
607
]
},
{
"type": "text",
"bbox": [
107,
610,
506,
666
],
"lines": [
{
"bbox": [
106,
611,
506,
623
],
"spans": [
{
"bbox": [
106,
611,
506,
623
],
"score": 1.0,
"content": "Recent work [13] also show that DDIM is a first-order discretization of diffusion ODEs by differenti-",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
621,
505,
633
],
"spans": [
{
"bbox": [
106,
621,
505,
633
],
"score": 1.0,
"content": "ating both sides of Eq. (4.1). However, they cannot explain the difference between DDIM and the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
632,
506,
645
],
"spans": [
{
"bbox": [
106,
632,
506,
645
],
"score": 1.0,
"content": "first-order Euler discretization of diffusion ODEs. In contrast, by showing that DDIM is a special",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
644,
506,
657
],
"spans": [
{
"bbox": [
106,
644,
506,
657
],
"score": 1.0,
"content": "case of DPM-Solver, we reveal that DDIM makes full use of the semi-linearity of diffusion ODEs,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
654,
352,
667
],
"spans": [
{
"bbox": [
106,
654,
352,
667
],
"score": 1.0,
"content": "which explains its superiority over traditional Euler methods.",
"type": "text"
}
],
"index": 42
}
],
"index": 40,
"bbox_fs": [
106,
611,
506,
667
]
},
{
"type": "title",
"bbox": [
106,
679,
353,
691
],
"lines": [
{
"bbox": [
104,
677,
354,
695
],
"spans": [
{
"bbox": [
104,
677,
354,
695
],
"score": 1.0,
"content": "4.2 Comparison with Traditional Runge-Kutta Methods",
"type": "text"
}
],
"index": 43
}
],
"index": 43
},
{
"type": "text",
"bbox": [
106,
699,
503,
722
],
"lines": [
{
"bbox": [
106,
699,
505,
712
],
"spans": [
{
"bbox": [
106,
699,
505,
712
],
"score": 1.0,
"content": "One can obtain a high-order solver by directly applying traditional explicit Runge-Kutta (RK) methods",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
709,
505,
724
],
"spans": [
{
"bbox": [
105,
709,
505,
724
],
"score": 1.0,
"content": "to the diffusion ODE in Eq. (2.7). Specifically, RK methods write the solution of Eq. (2.7) in the",
"type": "text"
}
],
"index": 45
}
],
"index": 44.5,
"bbox_fs": [
105,
699,
505,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
146,
109,
462,
196
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
78,
505,
108
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
90
],
"spans": [
{
"bbox": [
105,
77,
505,
90
],
"score": 1.0,
"content": "Table 1: FID ↓ on CIFAR-10 for different orders of Runge-Kutta (RK) methods and DPM-Solvers, varying the",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
88,
505,
99
],
"spans": [
{
"bbox": [
105,
88,
445,
99
],
"score": 1.0,
"content": "number of function evaluations (NFE). For RK methods, we evaluate diffusion ODEs w.r.t. both",
"type": "text"
},
{
"bbox": [
445,
89,
450,
97
],
"score": 0.51,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
450,
88,
505,
99
],
"score": 1.0,
"content": "(Eq. (2.7)) and",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
107,
98,
505,
110
],
"spans": [
{
"bbox": [
107,
99,
113,
107
],
"score": 0.5,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
114,
98,
256,
110
],
"score": 1.0,
"content": "(Eq. (E.1)). We use uniform step size in",
"type": "text"
},
{
"bbox": [
257,
99,
262,
107
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
262,
98,
389,
110
],
"score": 1.0,
"content": "for RK (t), and uniform step size in",
"type": "text"
},
{
"bbox": [
389,
99,
396,
107
],
"score": 0.58,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
397,
98,
424,
110
],
"score": 1.0,
"content": "for RK",
"type": "text"
},
{
"bbox": [
424,
99,
436,
108
],
"score": 0.43,
"content": "( \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
437,
98,
505,
110
],
"score": 1.0,
"content": "and DPM-Solvers.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
146,
109,
462,
196
],
"group_id": 0,
"lines": [
{
"bbox": [
146,
109,
462,
196
],
"spans": [
{
"bbox": [
146,
109,
462,
196
],
"score": 0.98,
"html": "| Sampling method\\NFE | 12 | 18 | 24 | 30 | 36 | 42 | 48 |
| RK2 (t) | 16.40 | 7.25 | 3.90 | 3.63 | 3.58 | 3.59 | 3.54 |
| RK2(入) | 107.81 | 42.04 | 17.71 | 7.65 | 4.62 | 3.58 | 3.17 |
| DPM-Solver-2 | 5.28 | 3.43 | 3.02 | 2.85 | 2.78 | 2.72 | 2.69 |
| RK3 (t) | 48.75 | 21.86 | 10.90 | 6.96 | 5.22 | 4.56 | 4.12 |
| RK3 (入) | 34.29 | 4.90 | 3.50 | 3.03 | 2.85 | 2.74 | 2.69 |
| DPM-Solver-3 | 6.03 | 2.90 | 2.75 | 2.70 | 2.67 | 2.65 | 2.65 |
",
"type": "table",
"image_path": "114b9edc4cc71bed7f9d053944dde80589018ddf790160907d7334f50b5c35fb.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
146,
109,
462,
138.0
],
"spans": [],
"index": 3
},
{
"bbox": [
146,
138.0,
462,
167.0
],
"spans": [],
"index": 4
},
{
"bbox": [
146,
167.0,
462,
196.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
218,
204,
230
],
"lines": [
{
"bbox": [
106,
217,
205,
232
],
"spans": [
{
"bbox": [
106,
217,
205,
232
],
"score": 1.0,
"content": "following integral form:",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "interline_equation",
"bbox": [
154,
236,
456,
265
],
"lines": [
{
"bbox": [
154,
236,
456,
265
],
"spans": [
{
"bbox": [
154,
236,
456,
265
],
"score": 0.94,
"content": "{ \\bf { x } } _ { t } = { \\bf { x } } _ { s } + \\int _ { s } ^ { t } h _ { \\theta } ( { \\bf { x } } _ { \\tau } , \\tau ) \\mathrm { { d } } \\tau = { \\bf { x } } _ { s } + \\int _ { s } ^ { t } \\left( f ( \\tau ) { \\bf { x } } _ { \\tau } + \\frac { g ^ { 2 } ( \\tau ) } { 2 \\sigma _ { \\tau } } \\epsilon _ { \\theta } ( { \\bf { x } } _ { \\tau } , \\tau ) \\right) \\mathrm { { d } } \\tau ,",
"type": "interline_equation",
"image_path": "dbc1e1d3bec4cc742d44654340bd9da788aefebc8fc4b4e2a9e38d739f018ceb.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
154,
236,
456,
245.66666666666666
],
"spans": [],
"index": 7
},
{
"bbox": [
154,
245.66666666666666,
456,
255.33333333333331
],
"spans": [],
"index": 8
},
{
"bbox": [
154,
255.33333333333331,
456,
265.0
],
"spans": [],
"index": 9
}
]
},
{
"type": "text",
"bbox": [
106,
271,
505,
370
],
"lines": [
{
"bbox": [
106,
271,
505,
284
],
"spans": [
{
"bbox": [
106,
271,
292,
284
],
"score": 1.0,
"content": "and use some intermediate time steps between",
"type": "text"
},
{
"bbox": [
293,
271,
312,
284
],
"score": 0.85,
"content": "[ t , s ]",
"type": "inline_equation"
},
{
"bbox": [
312,
271,
438,
284
],
"score": 1.0,
"content": "and combine the evaluations of",
"type": "text"
},
{
"bbox": [
439,
272,
451,
282
],
"score": 0.88,
"content": "h _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
451,
271,
505,
284
],
"score": 1.0,
"content": "at these time",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
282,
506,
295
],
"spans": [
{
"bbox": [
105,
282,
506,
295
],
"score": 1.0,
"content": "steps to approximate the whole integral. The approximation error of explicit RK methods depends on",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
294,
118,
304
],
"score": 0.85,
"content": "h _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
119,
294,
378,
306
],
"score": 1.0,
"content": ", which consists of the error corresponding to both the linear term",
"type": "text"
},
{
"bbox": [
378,
293,
410,
306
],
"score": 0.93,
"content": "f ( \\tau ) x _ { \\tau }",
"type": "inline_equation"
},
{
"bbox": [
410,
294,
505,
306
],
"score": 1.0,
"content": "and the nonlinear noise",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
304,
505,
317
],
"spans": [
{
"bbox": [
105,
304,
177,
317
],
"score": 1.0,
"content": "prediction model",
"type": "text"
},
{
"bbox": [
178,
305,
188,
315
],
"score": 0.82,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
189,
304,
505,
317
],
"score": 1.0,
"content": ". However, the error of the linear term may increase exponentially because the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
104,
314,
506,
329
],
"spans": [
{
"bbox": [
104,
314,
506,
329
],
"score": 1.0,
"content": "exact solution of the linear term has an exponential coefficient (as shown in Eq. (3.1)). There are many",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
325,
506,
339
],
"spans": [
{
"bbox": [
105,
325,
506,
339
],
"score": 1.0,
"content": "empirical evidence [25, 31] showing that directly using explicit RK methods for semi-linear ODEs",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
337,
505,
349
],
"spans": [
{
"bbox": [
106,
337,
505,
349
],
"score": 1.0,
"content": "may suffer from unstable numerical issues for large step size. We also demonstrate the empirical",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
347,
506,
360
],
"spans": [
{
"bbox": [
105,
347,
506,
360
],
"score": 1.0,
"content": "difference of the proposed DPM-Solver and the traditional explicit RK methods in Sec. 5.1, which",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
358,
506,
371
],
"spans": [
{
"bbox": [
106,
358,
506,
371
],
"score": 1.0,
"content": "shows that DPM-Solver have smaller discretization errors than the RK methods with the same order.",
"type": "text"
}
],
"index": 18
}
],
"index": 14
},
{
"type": "title",
"bbox": [
106,
384,
343,
397
],
"lines": [
{
"bbox": [
105,
384,
344,
399
],
"spans": [
{
"bbox": [
105,
384,
344,
399
],
"score": 1.0,
"content": "4.3 Training-based Fast Sampling Methods for DPMs",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
405,
506,
493
],
"lines": [
{
"bbox": [
105,
404,
506,
419
],
"spans": [
{
"bbox": [
105,
404,
506,
419
],
"score": 1.0,
"content": "Samplers that need extra training or optimization include knowledge distillation [13, 14], learning",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
416,
506,
429
],
"spans": [
{
"bbox": [
105,
416,
506,
429
],
"score": 1.0,
"content": "the noise level or variance [15, 16, 33], and learning the noise schedule or sample trajectory [17, 18].",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
427,
506,
440
],
"spans": [
{
"bbox": [
105,
427,
506,
440
],
"score": 1.0,
"content": "Although the progressive distillation method [13] can obtain a fast sampler within 4 steps, it needs",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
438,
506,
450
],
"spans": [
{
"bbox": [
105,
438,
506,
450
],
"score": 1.0,
"content": "further training costs and loses part of the information in the original DPM (e.g., after distillation, the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
449,
506,
462
],
"spans": [
{
"bbox": [
105,
449,
476,
462
],
"score": 1.0,
"content": "noise prediction model cannot predict the noise (score function) at every time step between",
"type": "text"
},
{
"bbox": [
477,
450,
502,
461
],
"score": 0.87,
"content": "[ 0 , T ] )",
"type": "inline_equation"
},
{
"bbox": [
503,
449,
506,
462
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
460,
506,
473
],
"spans": [
{
"bbox": [
105,
460,
506,
473
],
"score": 1.0,
"content": "In contrast, training-free samplers can keep all the information of the original model, and thereby can",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
471,
506,
483
],
"spans": [
{
"bbox": [
106,
471,
506,
483
],
"score": 1.0,
"content": "be directly extended to the conditional sampling by combining the original model and an external",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
482,
463,
494
],
"spans": [
{
"bbox": [
106,
482,
463,
494
],
"score": 1.0,
"content": "classifier [4] (e.g. see Appendix D for the conditional sampling with classifier guidance).",
"type": "text"
}
],
"index": 27
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
107,
498,
505,
553
],
"lines": [
{
"bbox": [
106,
498,
505,
510
],
"spans": [
{
"bbox": [
106,
498,
505,
510
],
"score": 1.0,
"content": "Beyond directly designing fast samplers for DPMs, several works also propose novel types of",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
509,
505,
521
],
"spans": [
{
"bbox": [
106,
509,
505,
521
],
"score": 1.0,
"content": "DPMs which supports faster sampling. For instance, defining a low-dimensional latent variable for",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
518,
506,
534
],
"spans": [
{
"bbox": [
104,
518,
506,
534
],
"score": 1.0,
"content": "DPMs [34]; designing special diffusion processes with bounded score functions [35]; combining",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
531,
506,
544
],
"spans": [
{
"bbox": [
105,
531,
506,
544
],
"score": 1.0,
"content": "GANs with the reverse process of DPMs [36]. The proposed DPM-Solver may also be suitable for",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
542,
417,
554
],
"spans": [
{
"bbox": [
106,
542,
417,
554
],
"score": 1.0,
"content": "accelerating the sampling of these DPMs, and we leave them for future work.",
"type": "text"
}
],
"index": 32
}
],
"index": 30
},
{
"type": "title",
"bbox": [
107,
570,
191,
584
],
"lines": [
{
"bbox": [
104,
568,
193,
587
],
"spans": [
{
"bbox": [
104,
568,
193,
587
],
"score": 1.0,
"content": "5 Experiments",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "text",
"bbox": [
107,
596,
505,
673
],
"lines": [
{
"bbox": [
105,
595,
505,
610
],
"spans": [
{
"bbox": [
105,
595,
505,
610
],
"score": 1.0,
"content": "In this section, we show that as a training-free sampler, DPM-Solver can greatly speedup the sampling",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
607,
505,
619
],
"spans": [
{
"bbox": [
105,
607,
505,
619
],
"score": 1.0,
"content": "of existing pre-trained DPMs, including both continuous-time and discrete-time ones, with both",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
618,
505,
630
],
"spans": [
{
"bbox": [
105,
618,
505,
630
],
"score": 1.0,
"content": "linear noise schedule [2, 19] and cosine noise schedule [16]. We vary different number of function",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
629,
505,
642
],
"spans": [
{
"bbox": [
105,
629,
412,
642
],
"score": 1.0,
"content": "evaluations (NFE) which is the number of calls to the noise prediction model",
"type": "text"
},
{
"bbox": [
412,
629,
448,
641
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
448,
629,
505,
642
],
"score": 1.0,
"content": ", and compare",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
640,
505,
652
],
"spans": [
{
"bbox": [
105,
640,
505,
652
],
"score": 1.0,
"content": "the sample quality between DPM-Solver and other methods. For each experiment, We draw 50K",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
650,
506,
664
],
"spans": [
{
"bbox": [
105,
650,
506,
664
],
"score": 1.0,
"content": "samples and use the widely adopted FID score [37] to evaluate the sample quality, where lower FID",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
662,
258,
676
],
"spans": [
{
"bbox": [
105,
662,
258,
676
],
"score": 1.0,
"content": "usually implies better sample quality.",
"type": "text"
}
],
"index": 40
}
],
"index": 37
},
{
"type": "text",
"bbox": [
108,
678,
505,
722
],
"lines": [
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "Unless explicitly mentioned, we always use the solver combination with the uniform step size",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
689,
505,
701
],
"spans": [
{
"bbox": [
106,
689,
505,
701
],
"score": 1.0,
"content": "schedule in Sec. 3.3 if the NFE budget is less than 20, and otherwise the DPM-Solver-3 with the",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
700,
505,
712
],
"spans": [
{
"bbox": [
106,
700,
505,
712
],
"score": 1.0,
"content": "adaptive step size schedule in Sec. 3.3. We refer to Appendix D for other implementation details of",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
710,
310,
724
],
"spans": [
{
"bbox": [
105,
710,
310,
724
],
"score": 1.0,
"content": "DPM-Solver and Appendix E for detailed settings.",
"type": "text"
}
],
"index": 44
}
],
"index": 42.5
}
],
"page_idx": 7,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
742,
308,
750
],
"lines": [
{
"bbox": [
301,
740,
310,
752
],
"spans": [
{
"bbox": [
301,
740,
310,
752
],
"score": 1.0,
"content": "",
"type": "text",
"height": 12,
"width": 9
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
146,
109,
462,
196
],
"blocks": [
{
"type": "table_caption",
"bbox": [
107,
78,
505,
108
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
77,
505,
90
],
"spans": [
{
"bbox": [
105,
77,
505,
90
],
"score": 1.0,
"content": "Table 1: FID ↓ on CIFAR-10 for different orders of Runge-Kutta (RK) methods and DPM-Solvers, varying the",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
88,
505,
99
],
"spans": [
{
"bbox": [
105,
88,
445,
99
],
"score": 1.0,
"content": "number of function evaluations (NFE). For RK methods, we evaluate diffusion ODEs w.r.t. both",
"type": "text"
},
{
"bbox": [
445,
89,
450,
97
],
"score": 0.51,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
450,
88,
505,
99
],
"score": 1.0,
"content": "(Eq. (2.7)) and",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
107,
98,
505,
110
],
"spans": [
{
"bbox": [
107,
99,
113,
107
],
"score": 0.5,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
114,
98,
256,
110
],
"score": 1.0,
"content": "(Eq. (E.1)). We use uniform step size in",
"type": "text"
},
{
"bbox": [
257,
99,
262,
107
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
262,
98,
389,
110
],
"score": 1.0,
"content": "for RK (t), and uniform step size in",
"type": "text"
},
{
"bbox": [
389,
99,
396,
107
],
"score": 0.58,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
397,
98,
424,
110
],
"score": 1.0,
"content": "for RK",
"type": "text"
},
{
"bbox": [
424,
99,
436,
108
],
"score": 0.43,
"content": "( \\lambda )",
"type": "inline_equation"
},
{
"bbox": [
437,
98,
505,
110
],
"score": 1.0,
"content": "and DPM-Solvers.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "table_body",
"bbox": [
146,
109,
462,
196
],
"group_id": 0,
"lines": [
{
"bbox": [
146,
109,
462,
196
],
"spans": [
{
"bbox": [
146,
109,
462,
196
],
"score": 0.98,
"html": "| Sampling method\\NFE | 12 | 18 | 24 | 30 | 36 | 42 | 48 |
| RK2 (t) | 16.40 | 7.25 | 3.90 | 3.63 | 3.58 | 3.59 | 3.54 |
| RK2(入) | 107.81 | 42.04 | 17.71 | 7.65 | 4.62 | 3.58 | 3.17 |
| DPM-Solver-2 | 5.28 | 3.43 | 3.02 | 2.85 | 2.78 | 2.72 | 2.69 |
| RK3 (t) | 48.75 | 21.86 | 10.90 | 6.96 | 5.22 | 4.56 | 4.12 |
| RK3 (入) | 34.29 | 4.90 | 3.50 | 3.03 | 2.85 | 2.74 | 2.69 |
| DPM-Solver-3 | 6.03 | 2.90 | 2.75 | 2.70 | 2.67 | 2.65 | 2.65 |
",
"type": "table",
"image_path": "114b9edc4cc71bed7f9d053944dde80589018ddf790160907d7334f50b5c35fb.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
146,
109,
462,
138.0
],
"spans": [],
"index": 3
},
{
"bbox": [
146,
138.0,
462,
167.0
],
"spans": [],
"index": 4
},
{
"bbox": [
146,
167.0,
462,
196.0
],
"spans": [],
"index": 5
}
]
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
218,
204,
230
],
"lines": [
{
"bbox": [
106,
217,
205,
232
],
"spans": [
{
"bbox": [
106,
217,
205,
232
],
"score": 1.0,
"content": "following integral form:",
"type": "text"
}
],
"index": 6
}
],
"index": 6,
"bbox_fs": [
106,
217,
205,
232
]
},
{
"type": "interline_equation",
"bbox": [
154,
236,
456,
265
],
"lines": [
{
"bbox": [
154,
236,
456,
265
],
"spans": [
{
"bbox": [
154,
236,
456,
265
],
"score": 0.94,
"content": "{ \\bf { x } } _ { t } = { \\bf { x } } _ { s } + \\int _ { s } ^ { t } h _ { \\theta } ( { \\bf { x } } _ { \\tau } , \\tau ) \\mathrm { { d } } \\tau = { \\bf { x } } _ { s } + \\int _ { s } ^ { t } \\left( f ( \\tau ) { \\bf { x } } _ { \\tau } + \\frac { g ^ { 2 } ( \\tau ) } { 2 \\sigma _ { \\tau } } \\epsilon _ { \\theta } ( { \\bf { x } } _ { \\tau } , \\tau ) \\right) \\mathrm { { d } } \\tau ,",
"type": "interline_equation",
"image_path": "dbc1e1d3bec4cc742d44654340bd9da788aefebc8fc4b4e2a9e38d739f018ceb.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
154,
236,
456,
245.66666666666666
],
"spans": [],
"index": 7
},
{
"bbox": [
154,
245.66666666666666,
456,
255.33333333333331
],
"spans": [],
"index": 8
},
{
"bbox": [
154,
255.33333333333331,
456,
265.0
],
"spans": [],
"index": 9
}
]
},
{
"type": "text",
"bbox": [
106,
271,
505,
370
],
"lines": [
{
"bbox": [
106,
271,
505,
284
],
"spans": [
{
"bbox": [
106,
271,
292,
284
],
"score": 1.0,
"content": "and use some intermediate time steps between",
"type": "text"
},
{
"bbox": [
293,
271,
312,
284
],
"score": 0.85,
"content": "[ t , s ]",
"type": "inline_equation"
},
{
"bbox": [
312,
271,
438,
284
],
"score": 1.0,
"content": "and combine the evaluations of",
"type": "text"
},
{
"bbox": [
439,
272,
451,
282
],
"score": 0.88,
"content": "h _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
451,
271,
505,
284
],
"score": 1.0,
"content": "at these time",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
282,
506,
295
],
"spans": [
{
"bbox": [
105,
282,
506,
295
],
"score": 1.0,
"content": "steps to approximate the whole integral. The approximation error of explicit RK methods depends on",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
293,
505,
306
],
"spans": [
{
"bbox": [
106,
294,
118,
304
],
"score": 0.85,
"content": "h _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
119,
294,
378,
306
],
"score": 1.0,
"content": ", which consists of the error corresponding to both the linear term",
"type": "text"
},
{
"bbox": [
378,
293,
410,
306
],
"score": 0.93,
"content": "f ( \\tau ) x _ { \\tau }",
"type": "inline_equation"
},
{
"bbox": [
410,
294,
505,
306
],
"score": 1.0,
"content": "and the nonlinear noise",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
304,
505,
317
],
"spans": [
{
"bbox": [
105,
304,
177,
317
],
"score": 1.0,
"content": "prediction model",
"type": "text"
},
{
"bbox": [
178,
305,
188,
315
],
"score": 0.82,
"content": "\\epsilon _ { \\theta }",
"type": "inline_equation"
},
{
"bbox": [
189,
304,
505,
317
],
"score": 1.0,
"content": ". However, the error of the linear term may increase exponentially because the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
104,
314,
506,
329
],
"spans": [
{
"bbox": [
104,
314,
506,
329
],
"score": 1.0,
"content": "exact solution of the linear term has an exponential coefficient (as shown in Eq. (3.1)). There are many",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
325,
506,
339
],
"spans": [
{
"bbox": [
105,
325,
506,
339
],
"score": 1.0,
"content": "empirical evidence [25, 31] showing that directly using explicit RK methods for semi-linear ODEs",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
337,
505,
349
],
"spans": [
{
"bbox": [
106,
337,
505,
349
],
"score": 1.0,
"content": "may suffer from unstable numerical issues for large step size. We also demonstrate the empirical",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
347,
506,
360
],
"spans": [
{
"bbox": [
105,
347,
506,
360
],
"score": 1.0,
"content": "difference of the proposed DPM-Solver and the traditional explicit RK methods in Sec. 5.1, which",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
358,
506,
371
],
"spans": [
{
"bbox": [
106,
358,
506,
371
],
"score": 1.0,
"content": "shows that DPM-Solver have smaller discretization errors than the RK methods with the same order.",
"type": "text"
}
],
"index": 18
}
],
"index": 14,
"bbox_fs": [
104,
271,
506,
371
]
},
{
"type": "title",
"bbox": [
106,
384,
343,
397
],
"lines": [
{
"bbox": [
105,
384,
344,
399
],
"spans": [
{
"bbox": [
105,
384,
344,
399
],
"score": 1.0,
"content": "4.3 Training-based Fast Sampling Methods for DPMs",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
405,
506,
493
],
"lines": [
{
"bbox": [
105,
404,
506,
419
],
"spans": [
{
"bbox": [
105,
404,
506,
419
],
"score": 1.0,
"content": "Samplers that need extra training or optimization include knowledge distillation [13, 14], learning",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
416,
506,
429
],
"spans": [
{
"bbox": [
105,
416,
506,
429
],
"score": 1.0,
"content": "the noise level or variance [15, 16, 33], and learning the noise schedule or sample trajectory [17, 18].",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
427,
506,
440
],
"spans": [
{
"bbox": [
105,
427,
506,
440
],
"score": 1.0,
"content": "Although the progressive distillation method [13] can obtain a fast sampler within 4 steps, it needs",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
438,
506,
450
],
"spans": [
{
"bbox": [
105,
438,
506,
450
],
"score": 1.0,
"content": "further training costs and loses part of the information in the original DPM (e.g., after distillation, the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
449,
506,
462
],
"spans": [
{
"bbox": [
105,
449,
476,
462
],
"score": 1.0,
"content": "noise prediction model cannot predict the noise (score function) at every time step between",
"type": "text"
},
{
"bbox": [
477,
450,
502,
461
],
"score": 0.87,
"content": "[ 0 , T ] )",
"type": "inline_equation"
},
{
"bbox": [
503,
449,
506,
462
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
460,
506,
473
],
"spans": [
{
"bbox": [
105,
460,
506,
473
],
"score": 1.0,
"content": "In contrast, training-free samplers can keep all the information of the original model, and thereby can",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
471,
506,
483
],
"spans": [
{
"bbox": [
106,
471,
506,
483
],
"score": 1.0,
"content": "be directly extended to the conditional sampling by combining the original model and an external",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
482,
463,
494
],
"spans": [
{
"bbox": [
106,
482,
463,
494
],
"score": 1.0,
"content": "classifier [4] (e.g. see Appendix D for the conditional sampling with classifier guidance).",
"type": "text"
}
],
"index": 27
}
],
"index": 23.5,
"bbox_fs": [
105,
404,
506,
494
]
},
{
"type": "text",
"bbox": [
107,
498,
505,
553
],
"lines": [
{
"bbox": [
106,
498,
505,
510
],
"spans": [
{
"bbox": [
106,
498,
505,
510
],
"score": 1.0,
"content": "Beyond directly designing fast samplers for DPMs, several works also propose novel types of",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
509,
505,
521
],
"spans": [
{
"bbox": [
106,
509,
505,
521
],
"score": 1.0,
"content": "DPMs which supports faster sampling. For instance, defining a low-dimensional latent variable for",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
518,
506,
534
],
"spans": [
{
"bbox": [
104,
518,
506,
534
],
"score": 1.0,
"content": "DPMs [34]; designing special diffusion processes with bounded score functions [35]; combining",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
531,
506,
544
],
"spans": [
{
"bbox": [
105,
531,
506,
544
],
"score": 1.0,
"content": "GANs with the reverse process of DPMs [36]. The proposed DPM-Solver may also be suitable for",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
542,
417,
554
],
"spans": [
{
"bbox": [
106,
542,
417,
554
],
"score": 1.0,
"content": "accelerating the sampling of these DPMs, and we leave them for future work.",
"type": "text"
}
],
"index": 32
}
],
"index": 30,
"bbox_fs": [
104,
498,
506,
554
]
},
{
"type": "title",
"bbox": [
107,
570,
191,
584
],
"lines": [
{
"bbox": [
104,
568,
193,
587
],
"spans": [
{
"bbox": [
104,
568,
193,
587
],
"score": 1.0,
"content": "5 Experiments",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "text",
"bbox": [
107,
596,
505,
673
],
"lines": [
{
"bbox": [
105,
595,
505,
610
],
"spans": [
{
"bbox": [
105,
595,
505,
610
],
"score": 1.0,
"content": "In this section, we show that as a training-free sampler, DPM-Solver can greatly speedup the sampling",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
607,
505,
619
],
"spans": [
{
"bbox": [
105,
607,
505,
619
],
"score": 1.0,
"content": "of existing pre-trained DPMs, including both continuous-time and discrete-time ones, with both",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
618,
505,
630
],
"spans": [
{
"bbox": [
105,
618,
505,
630
],
"score": 1.0,
"content": "linear noise schedule [2, 19] and cosine noise schedule [16]. We vary different number of function",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
629,
505,
642
],
"spans": [
{
"bbox": [
105,
629,
412,
642
],
"score": 1.0,
"content": "evaluations (NFE) which is the number of calls to the noise prediction model",
"type": "text"
},
{
"bbox": [
412,
629,
448,
641
],
"score": 0.93,
"content": "\\epsilon _ { \\theta } ( x _ { t } , t )",
"type": "inline_equation"
},
{
"bbox": [
448,
629,
505,
642
],
"score": 1.0,
"content": ", and compare",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
640,
505,
652
],
"spans": [
{
"bbox": [
105,
640,
505,
652
],
"score": 1.0,
"content": "the sample quality between DPM-Solver and other methods. For each experiment, We draw 50K",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
650,
506,
664
],
"spans": [
{
"bbox": [
105,
650,
506,
664
],
"score": 1.0,
"content": "samples and use the widely adopted FID score [37] to evaluate the sample quality, where lower FID",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
662,
258,
676
],
"spans": [
{
"bbox": [
105,
662,
258,
676
],
"score": 1.0,
"content": "usually implies better sample quality.",
"type": "text"
}
],
"index": 40
}
],
"index": 37,
"bbox_fs": [
105,
595,
506,
676
]
},
{
"type": "text",
"bbox": [
108,
678,
505,
722
],
"lines": [
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "Unless explicitly mentioned, we always use the solver combination with the uniform step size",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
689,
505,
701
],
"spans": [
{
"bbox": [
106,
689,
505,
701
],
"score": 1.0,
"content": "schedule in Sec. 3.3 if the NFE budget is less than 20, and otherwise the DPM-Solver-3 with the",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
700,
505,
712
],
"spans": [
{
"bbox": [
106,
700,
505,
712
],
"score": 1.0,
"content": "adaptive step size schedule in Sec. 3.3. We refer to Appendix D for other implementation details of",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
710,
310,
724
],
"spans": [
{
"bbox": [
105,
710,
310,
724
],
"score": 1.0,
"content": "DPM-Solver and Appendix E for detailed settings.",
"type": "text"
}
],
"index": 44
}
],
"index": 42.5,
"bbox_fs": [
105,
677,
506,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
110,
69,
500,
289
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
69,
500,
289
],
"group_id": 0,
"lines": [
{
"bbox": [
110,
69,
500,
289
],
"spans": [
{
"bbox": [
110,
69,
500,
289
],
"score": 0.976,
"type": "image",
"image_path": "4e1bcabdec913d95ec9ff692dca4450fca481c16e793a3c86ffe1cf469e22305.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
69,
500,
142.33333333333331
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
142.33333333333331,
500,
215.66666666666663
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
215.66666666666663,
500,
288.99999999999994
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
295,
505,
356
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
296,
505,
307
],
"spans": [
{
"bbox": [
106,
296,
264,
307
],
"score": 1.0,
"content": "Figure 2: Sample quality measured by FID",
"type": "text"
},
{
"bbox": [
264,
297,
271,
306
],
"score": 0.31,
"content": "\\downarrow",
"type": "inline_equation"
},
{
"bbox": [
272,
296,
505,
307
],
"score": 1.0,
"content": "of different sampling methods for DPMs on CIFAR-10 with both",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
306,
505,
317
],
"spans": [
{
"bbox": [
106,
306,
428,
317
],
"score": 1.0,
"content": "continuous-time and discrete-time models, CelebA 64x64, ImageNet 64x64, ImageNet",
"type": "text"
},
{
"bbox": [
429,
306,
462,
315
],
"score": 0.42,
"content": "1 2 8 \\mathrm { x } 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
462,
306,
505,
317
],
"score": 1.0,
"content": "and LSUN",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
316,
505,
327
],
"spans": [
{
"bbox": [
106,
316,
141,
327
],
"score": 1.0,
"content": "bedroom",
"type": "text"
},
{
"bbox": [
141,
316,
174,
325
],
"score": 0.63,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
174,
316,
505,
327
],
"score": 1.0,
"content": "with discrete-time models, varying the number of function evaluations (NFE). The method",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
107,
324,
506,
339
],
"spans": [
{
"bbox": [
107,
325,
140,
336
],
"score": 0.26,
"content": "^ { \\dag } { \\bf G } { \\bf G } { \\bf D } { \\bf M }",
"type": "inline_equation"
},
{
"bbox": [
140,
324,
506,
339
],
"score": 1.0,
"content": "[18] needs extra training to optimize the sample trajectory, while other methods are training-free. To",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
104,
335,
507,
347
],
"spans": [
{
"bbox": [
104,
335,
507,
347
],
"score": 1.0,
"content": "get the strongest baseline, we use the quadratic step size for DDIM on CelebA, which has a better FID than that",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
346,
288,
357
],
"spans": [
{
"bbox": [
106,
346,
288,
357
],
"score": 1.0,
"content": "of the uniform step size in the original paper [19].",
"type": "text"
}
],
"index": 8
}
],
"index": 5.5
}
],
"index": 3.25
},
{
"type": "title",
"bbox": [
107,
366,
364,
378
],
"lines": [
{
"bbox": [
105,
365,
365,
381
],
"spans": [
{
"bbox": [
105,
365,
365,
381
],
"score": 1.0,
"content": "5.1 Comparison with Continuous-Time Sampling Methods",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
107,
388,
505,
443
],
"lines": [
{
"bbox": [
106,
388,
505,
401
],
"spans": [
{
"bbox": [
106,
388,
505,
401
],
"score": 1.0,
"content": "We firstly compare DPM-Solver with other continuous-time sampling methods for DPMs. The",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
399,
505,
412
],
"spans": [
{
"bbox": [
105,
399,
505,
412
],
"score": 1.0,
"content": "compared methods include the Euler-Maruyama discretization for diffusion SDEs [3], the adaptive",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
409,
506,
424
],
"spans": [
{
"bbox": [
105,
409,
506,
424
],
"score": 1.0,
"content": "step size solver for diffusion SDEs [20] and the RK methods for diffusion ODEs [3, 28] in Eq. (2.7).",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
419,
506,
435
],
"spans": [
{
"bbox": [
105,
419,
506,
435
],
"score": 1.0,
"content": "We compare these methods for sampling from a pre-trained continuous-time “VP deep” model [3] on",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
432,
338,
444
],
"spans": [
{
"bbox": [
106,
432,
338,
444
],
"score": 1.0,
"content": "the CIFAR-10 dataset [29] with the linear noise schedule.",
"type": "text"
}
],
"index": 14
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
448,
505,
525
],
"lines": [
{
"bbox": [
105,
448,
505,
461
],
"spans": [
{
"bbox": [
105,
448,
505,
461
],
"score": 1.0,
"content": "Fig. 2a shows the efficiency of compared solvers. We use uniform time steps with 50, 200, 1000 NFE",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
459,
505,
471
],
"spans": [
{
"bbox": [
105,
459,
505,
471
],
"score": 1.0,
"content": "for the diffusion SDE with Euler discretization, and vary the tolerance hyperparameter [3, 20] for the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
470,
505,
483
],
"spans": [
{
"bbox": [
106,
470,
505,
483
],
"score": 1.0,
"content": "adaptive step size SDE solver [20] and RK45 ODE solver [28] to control the NFE. DPM-Solver can",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
481,
506,
494
],
"spans": [
{
"bbox": [
105,
481,
506,
494
],
"score": 1.0,
"content": "generate good sample quality within around 10 NFE, while other solvers have large discretization",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
491,
505,
504
],
"spans": [
{
"bbox": [
105,
491,
372,
504
],
"score": 1.0,
"content": "error even in 50 NFE, which shows that DPM-Solver can achieve",
"type": "text"
},
{
"bbox": [
372,
492,
387,
502
],
"score": 0.85,
"content": "{ \\sim } 5 ",
"type": "inline_equation"
},
{
"bbox": [
387,
491,
505,
504
],
"score": 1.0,
"content": "speedup of the previous best",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
503,
505,
515
],
"spans": [
{
"bbox": [
105,
503,
505,
515
],
"score": 1.0,
"content": "solver. In particular, we achieve 4.70 FID with 10 NFE, 3.75 FID with 12 NFE, 3.24 FID with 15",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
514,
413,
525
],
"spans": [
{
"bbox": [
106,
514,
413,
525
],
"score": 1.0,
"content": "NFE, and 2.87 FID with 20 NFE, which is the fastest sampler on CIFAR-10.",
"type": "text"
}
],
"index": 21
}
],
"index": 18
},
{
"type": "text",
"bbox": [
107,
530,
505,
628
],
"lines": [
{
"bbox": [
106,
530,
506,
542
],
"spans": [
{
"bbox": [
106,
530,
506,
542
],
"score": 1.0,
"content": "As an ablation study, we also compare the second-order and third-order DPM-Solver and RK methods,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
540,
506,
554
],
"spans": [
{
"bbox": [
105,
540,
434,
554
],
"score": 1.0,
"content": "as shown in Table 1. We compare RK methods for diffusion ODEs w.r.t. both time",
"type": "text"
},
{
"bbox": [
434,
542,
440,
551
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
440,
540,
506,
554
],
"score": 1.0,
"content": "in Eq. (2.7) and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
551,
506,
565
],
"spans": [
{
"bbox": [
105,
551,
164,
565
],
"score": 1.0,
"content": "half-log-SNR",
"type": "text"
},
{
"bbox": [
164,
552,
172,
562
],
"score": 0.7,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
172,
551,
506,
565
],
"score": 1.0,
"content": "by applying change-of-variable (see detailed formulations in Appendix E.1). The",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
562,
505,
576
],
"spans": [
{
"bbox": [
105,
562,
505,
576
],
"score": 1.0,
"content": "results show that given the same NFE, the sample quality of DPM-Solver is consistently better than",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
573,
506,
587
],
"spans": [
{
"bbox": [
104,
573,
506,
587
],
"score": 1.0,
"content": "RK methods with the same order. The superior efficiency of DPM-Solver is particularly evident in",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
584,
505,
597
],
"spans": [
{
"bbox": [
106,
584,
505,
597
],
"score": 1.0,
"content": "the few-step regime under 15 NFE, where RK methods have rather large discretization errors. This",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
594,
506,
609
],
"spans": [
{
"bbox": [
105,
594,
506,
609
],
"score": 1.0,
"content": "is mainly because DPM-Solver analytically computes the linear term, avoiding the corresponding",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
606,
506,
619
],
"spans": [
{
"bbox": [
106,
606,
506,
619
],
"score": 1.0,
"content": "discretization error. Besides, the higher order DPM-Solver-3 converges faster than DPM-Solver-2,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
618,
307,
630
],
"spans": [
{
"bbox": [
106,
618,
307,
630
],
"score": 1.0,
"content": "which matches the order analysis in Theorem 3.2.",
"type": "text"
}
],
"index": 30
}
],
"index": 26
},
{
"type": "title",
"bbox": [
106,
645,
349,
657
],
"lines": [
{
"bbox": [
104,
643,
351,
661
],
"spans": [
{
"bbox": [
104,
643,
351,
661
],
"score": 1.0,
"content": "5.2 Comparison with Discrete-Time Sampling Methods",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
667,
505,
722
],
"lines": [
{
"bbox": [
105,
666,
506,
680
],
"spans": [
{
"bbox": [
105,
666,
506,
680
],
"score": 1.0,
"content": "We use the method in Sec. 3.4 for using DPM-Solver in discrete-time DPMs, and then compare",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "DPM-Solver with other discrete-time training-free samplers, including DDPM [2], DDIM [19],",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "Analytic-DDPM [21], Analytic-DDIM [21], PNDM [22], FastDPM [38] and Itô-Taylor [24]. We",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
700,
506,
713
],
"spans": [
{
"bbox": [
105,
700,
506,
713
],
"score": 1.0,
"content": "also compare with GGDM [18], which uses the same pre-trained model but needs further training for",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
711,
466,
723
],
"spans": [
{
"bbox": [
105,
711,
466,
723
],
"score": 1.0,
"content": "the sampling trajectory. We compare the sample quality by varying NFE from 10 to 1000.",
"type": "text"
}
],
"index": 36
}
],
"index": 34
}
],
"page_idx": 8,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
741,
308,
750
],
"lines": [
{
"bbox": [
302,
741,
309,
752
],
"spans": [
{
"bbox": [
302,
741,
309,
752
],
"score": 1.0,
"content": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
110,
69,
500,
289
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
69,
500,
289
],
"group_id": 0,
"lines": [
{
"bbox": [
110,
69,
500,
289
],
"spans": [
{
"bbox": [
110,
69,
500,
289
],
"score": 0.976,
"type": "image",
"image_path": "4e1bcabdec913d95ec9ff692dca4450fca481c16e793a3c86ffe1cf469e22305.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
69,
500,
142.33333333333331
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
142.33333333333331,
500,
215.66666666666663
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
215.66666666666663,
500,
288.99999999999994
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
295,
505,
356
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
296,
505,
307
],
"spans": [
{
"bbox": [
106,
296,
264,
307
],
"score": 1.0,
"content": "Figure 2: Sample quality measured by FID",
"type": "text"
},
{
"bbox": [
264,
297,
271,
306
],
"score": 0.31,
"content": "\\downarrow",
"type": "inline_equation"
},
{
"bbox": [
272,
296,
505,
307
],
"score": 1.0,
"content": "of different sampling methods for DPMs on CIFAR-10 with both",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
306,
505,
317
],
"spans": [
{
"bbox": [
106,
306,
428,
317
],
"score": 1.0,
"content": "continuous-time and discrete-time models, CelebA 64x64, ImageNet 64x64, ImageNet",
"type": "text"
},
{
"bbox": [
429,
306,
462,
315
],
"score": 0.42,
"content": "1 2 8 \\mathrm { x } 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
462,
306,
505,
317
],
"score": 1.0,
"content": "and LSUN",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
316,
505,
327
],
"spans": [
{
"bbox": [
106,
316,
141,
327
],
"score": 1.0,
"content": "bedroom",
"type": "text"
},
{
"bbox": [
141,
316,
174,
325
],
"score": 0.63,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
174,
316,
505,
327
],
"score": 1.0,
"content": "with discrete-time models, varying the number of function evaluations (NFE). The method",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
107,
324,
506,
339
],
"spans": [
{
"bbox": [
107,
325,
140,
336
],
"score": 0.26,
"content": "^ { \\dag } { \\bf G } { \\bf G } { \\bf D } { \\bf M }",
"type": "inline_equation"
},
{
"bbox": [
140,
324,
506,
339
],
"score": 1.0,
"content": "[18] needs extra training to optimize the sample trajectory, while other methods are training-free. To",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
104,
335,
507,
347
],
"spans": [
{
"bbox": [
104,
335,
507,
347
],
"score": 1.0,
"content": "get the strongest baseline, we use the quadratic step size for DDIM on CelebA, which has a better FID than that",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
346,
288,
357
],
"spans": [
{
"bbox": [
106,
346,
288,
357
],
"score": 1.0,
"content": "of the uniform step size in the original paper [19].",
"type": "text"
}
],
"index": 8
}
],
"index": 5.5
}
],
"index": 3.25
},
{
"type": "title",
"bbox": [
107,
366,
364,
378
],
"lines": [
{
"bbox": [
105,
365,
365,
381
],
"spans": [
{
"bbox": [
105,
365,
365,
381
],
"score": 1.0,
"content": "5.1 Comparison with Continuous-Time Sampling Methods",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
107,
388,
505,
443
],
"lines": [
{
"bbox": [
106,
388,
505,
401
],
"spans": [
{
"bbox": [
106,
388,
505,
401
],
"score": 1.0,
"content": "We firstly compare DPM-Solver with other continuous-time sampling methods for DPMs. The",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
399,
505,
412
],
"spans": [
{
"bbox": [
105,
399,
505,
412
],
"score": 1.0,
"content": "compared methods include the Euler-Maruyama discretization for diffusion SDEs [3], the adaptive",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
409,
506,
424
],
"spans": [
{
"bbox": [
105,
409,
506,
424
],
"score": 1.0,
"content": "step size solver for diffusion SDEs [20] and the RK methods for diffusion ODEs [3, 28] in Eq. (2.7).",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
419,
506,
435
],
"spans": [
{
"bbox": [
105,
419,
506,
435
],
"score": 1.0,
"content": "We compare these methods for sampling from a pre-trained continuous-time “VP deep” model [3] on",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
432,
338,
444
],
"spans": [
{
"bbox": [
106,
432,
338,
444
],
"score": 1.0,
"content": "the CIFAR-10 dataset [29] with the linear noise schedule.",
"type": "text"
}
],
"index": 14
}
],
"index": 12,
"bbox_fs": [
105,
388,
506,
444
]
},
{
"type": "text",
"bbox": [
107,
448,
505,
525
],
"lines": [
{
"bbox": [
105,
448,
505,
461
],
"spans": [
{
"bbox": [
105,
448,
505,
461
],
"score": 1.0,
"content": "Fig. 2a shows the efficiency of compared solvers. We use uniform time steps with 50, 200, 1000 NFE",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
459,
505,
471
],
"spans": [
{
"bbox": [
105,
459,
505,
471
],
"score": 1.0,
"content": "for the diffusion SDE with Euler discretization, and vary the tolerance hyperparameter [3, 20] for the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
470,
505,
483
],
"spans": [
{
"bbox": [
106,
470,
505,
483
],
"score": 1.0,
"content": "adaptive step size SDE solver [20] and RK45 ODE solver [28] to control the NFE. DPM-Solver can",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
481,
506,
494
],
"spans": [
{
"bbox": [
105,
481,
506,
494
],
"score": 1.0,
"content": "generate good sample quality within around 10 NFE, while other solvers have large discretization",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
491,
505,
504
],
"spans": [
{
"bbox": [
105,
491,
372,
504
],
"score": 1.0,
"content": "error even in 50 NFE, which shows that DPM-Solver can achieve",
"type": "text"
},
{
"bbox": [
372,
492,
387,
502
],
"score": 0.85,
"content": "{ \\sim } 5 ",
"type": "inline_equation"
},
{
"bbox": [
387,
491,
505,
504
],
"score": 1.0,
"content": "speedup of the previous best",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
503,
505,
515
],
"spans": [
{
"bbox": [
105,
503,
505,
515
],
"score": 1.0,
"content": "solver. In particular, we achieve 4.70 FID with 10 NFE, 3.75 FID with 12 NFE, 3.24 FID with 15",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
514,
413,
525
],
"spans": [
{
"bbox": [
106,
514,
413,
525
],
"score": 1.0,
"content": "NFE, and 2.87 FID with 20 NFE, which is the fastest sampler on CIFAR-10.",
"type": "text"
}
],
"index": 21
}
],
"index": 18,
"bbox_fs": [
105,
448,
506,
525
]
},
{
"type": "text",
"bbox": [
107,
530,
505,
628
],
"lines": [
{
"bbox": [
106,
530,
506,
542
],
"spans": [
{
"bbox": [
106,
530,
506,
542
],
"score": 1.0,
"content": "As an ablation study, we also compare the second-order and third-order DPM-Solver and RK methods,",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
540,
506,
554
],
"spans": [
{
"bbox": [
105,
540,
434,
554
],
"score": 1.0,
"content": "as shown in Table 1. We compare RK methods for diffusion ODEs w.r.t. both time",
"type": "text"
},
{
"bbox": [
434,
542,
440,
551
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
440,
540,
506,
554
],
"score": 1.0,
"content": "in Eq. (2.7) and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
551,
506,
565
],
"spans": [
{
"bbox": [
105,
551,
164,
565
],
"score": 1.0,
"content": "half-log-SNR",
"type": "text"
},
{
"bbox": [
164,
552,
172,
562
],
"score": 0.7,
"content": "\\lambda",
"type": "inline_equation"
},
{
"bbox": [
172,
551,
506,
565
],
"score": 1.0,
"content": "by applying change-of-variable (see detailed formulations in Appendix E.1). The",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
562,
505,
576
],
"spans": [
{
"bbox": [
105,
562,
505,
576
],
"score": 1.0,
"content": "results show that given the same NFE, the sample quality of DPM-Solver is consistently better than",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
104,
573,
506,
587
],
"spans": [
{
"bbox": [
104,
573,
506,
587
],
"score": 1.0,
"content": "RK methods with the same order. The superior efficiency of DPM-Solver is particularly evident in",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
584,
505,
597
],
"spans": [
{
"bbox": [
106,
584,
505,
597
],
"score": 1.0,
"content": "the few-step regime under 15 NFE, where RK methods have rather large discretization errors. This",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
594,
506,
609
],
"spans": [
{
"bbox": [
105,
594,
506,
609
],
"score": 1.0,
"content": "is mainly because DPM-Solver analytically computes the linear term, avoiding the corresponding",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
606,
506,
619
],
"spans": [
{
"bbox": [
106,
606,
506,
619
],
"score": 1.0,
"content": "discretization error. Besides, the higher order DPM-Solver-3 converges faster than DPM-Solver-2,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
618,
307,
630
],
"spans": [
{
"bbox": [
106,
618,
307,
630
],
"score": 1.0,
"content": "which matches the order analysis in Theorem 3.2.",
"type": "text"
}
],
"index": 30
}
],
"index": 26,
"bbox_fs": [
104,
530,
506,
630
]
},
{
"type": "title",
"bbox": [
106,
645,
349,
657
],
"lines": [
{
"bbox": [
104,
643,
351,
661
],
"spans": [
{
"bbox": [
104,
643,
351,
661
],
"score": 1.0,
"content": "5.2 Comparison with Discrete-Time Sampling Methods",
"type": "text"
}
],
"index": 31
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
667,
505,
722
],
"lines": [
{
"bbox": [
105,
666,
506,
680
],
"spans": [
{
"bbox": [
105,
666,
506,
680
],
"score": 1.0,
"content": "We use the method in Sec. 3.4 for using DPM-Solver in discrete-time DPMs, and then compare",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
677,
506,
691
],
"spans": [
{
"bbox": [
105,
677,
506,
691
],
"score": 1.0,
"content": "DPM-Solver with other discrete-time training-free samplers, including DDPM [2], DDIM [19],",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
688,
506,
702
],
"spans": [
{
"bbox": [
105,
688,
506,
702
],
"score": 1.0,
"content": "Analytic-DDPM [21], Analytic-DDIM [21], PNDM [22], FastDPM [38] and Itô-Taylor [24]. We",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
700,
506,
713
],
"spans": [
{
"bbox": [
105,
700,
506,
713
],
"score": 1.0,
"content": "also compare with GGDM [18], which uses the same pre-trained model but needs further training for",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
711,
466,
723
],
"spans": [
{
"bbox": [
105,
711,
466,
723
],
"score": 1.0,
"content": "the sampling trajectory. We compare the sample quality by varying NFE from 10 to 1000.",
"type": "text"
}
],
"index": 36
}
],
"index": 34,
"bbox_fs": [
105,
666,
506,
723
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
106,
72,
505,
182
],
"lines": [
{
"bbox": [
105,
72,
505,
86
],
"spans": [
{
"bbox": [
105,
72,
329,
86
],
"score": 1.0,
"content": "Specifically, we use the discrete-time model trained by",
"type": "text"
},
{
"bbox": [
329,
73,
356,
84
],
"score": 0.9,
"content": "L _ { \\mathrm { s i m p l e } }",
"type": "inline_equation"
},
{
"bbox": [
357,
72,
505,
86
],
"score": 1.0,
"content": "in [2] on the CIFAR-10 dataset with",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
506,
96
],
"spans": [
{
"bbox": [
105,
83,
506,
96
],
"score": 1.0,
"content": "linear noise schedule; the discrete-time model in [19] on CelebA 64x64 [39] with linear noise",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
93,
506,
109
],
"spans": [
{
"bbox": [
105,
93,
292,
109
],
"score": 1.0,
"content": "schedule; the discrete-time model trained by",
"type": "text"
},
{
"bbox": [
292,
95,
319,
106
],
"score": 0.89,
"content": "L _ { \\mathrm { h y b r i d } }",
"type": "inline_equation"
},
{
"bbox": [
320,
93,
506,
109
],
"score": 1.0,
"content": "in [16] on ImageNet 64x64 [26] with cosine",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
105,
505,
118
],
"spans": [
{
"bbox": [
106,
105,
505,
118
],
"score": 1.0,
"content": "noise schedule; the discrete-time model with classifier guidance in [4] on ImageNet 128x128 [26]",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
117,
505,
128
],
"spans": [
{
"bbox": [
106,
117,
426,
128
],
"score": 1.0,
"content": "with linear noise schedule; the discrete-time model in [4] on LSUN bedroom",
"type": "text"
},
{
"bbox": [
426,
117,
464,
127
],
"score": 0.29,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
464,
117,
505,
128
],
"score": 1.0,
"content": "[40] with",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
127,
506,
139
],
"spans": [
{
"bbox": [
105,
127,
506,
139
],
"score": 1.0,
"content": "linear noise schedule. For the models trained on ImageNet, we only use their “mean” model and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
138,
506,
150
],
"spans": [
{
"bbox": [
106,
138,
506,
150
],
"score": 1.0,
"content": "omit the “variance” model. As shown in Fig. 2, on all datasets, DPM-Solver can obtain reasonable",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
149,
505,
161
],
"spans": [
{
"bbox": [
105,
149,
505,
161
],
"score": 1.0,
"content": "samples within 12 steps (FID 4.65 on CIFAR-10, FID 3.71 on CelebA 64x64 and FID 19.97 on",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
159,
506,
172
],
"spans": [
{
"bbox": [
106,
159,
268,
172
],
"score": 1.0,
"content": "ImageNet 64x64, FID 4.08 on ImageNet",
"type": "text"
},
{
"bbox": [
268,
160,
304,
171
],
"score": 0.33,
"content": "1 2 8 \\mathbf { x } 1 2 8 _ { \\rho }",
"type": "inline_equation"
},
{
"bbox": [
304,
159,
345,
172
],
"score": 1.0,
"content": "), which is",
"type": "text"
},
{
"bbox": [
345,
160,
383,
171
],
"score": 0.92,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
383,
159,
506,
172
],
"score": 1.0,
"content": "faster than the previous fastest",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
170,
496,
185
],
"spans": [
{
"bbox": [
105,
170,
496,
185
],
"score": 1.0,
"content": "training-free sampler. DPM-Solver even outperforms GGDM, which requires additional training.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5
},
{
"type": "title",
"bbox": [
107,
199,
187,
212
],
"lines": [
{
"bbox": [
104,
197,
190,
214
],
"spans": [
{
"bbox": [
104,
197,
190,
214
],
"score": 1.0,
"content": "6 Conclusions",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
224,
505,
345
],
"lines": [
{
"bbox": [
105,
223,
506,
237
],
"spans": [
{
"bbox": [
105,
223,
506,
237
],
"score": 1.0,
"content": "We tackle the problem of fast and training-free sampling from DPMs. We propose DPM-Solver,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
235,
506,
248
],
"spans": [
{
"bbox": [
105,
235,
506,
248
],
"score": 1.0,
"content": "a fast dedicated training-free solver of diffusion ODEs for fast sampling of DPMs in around 10",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
246,
505,
258
],
"spans": [
{
"bbox": [
105,
246,
505,
258
],
"score": 1.0,
"content": "steps of function evaluations. DPM-Solver leverages the semi-linearity of diffusion ODEs and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
257,
505,
269
],
"spans": [
{
"bbox": [
105,
257,
505,
269
],
"score": 1.0,
"content": "it directly approximates a simplified formulation of exact solutions of diffusion ODEs, which",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
268,
506,
280
],
"spans": [
{
"bbox": [
106,
268,
506,
280
],
"score": 1.0,
"content": "consists of an exponentially weighted integral of the noise prediction model. Inspired by numerical",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
279,
506,
291
],
"spans": [
{
"bbox": [
105,
279,
506,
291
],
"score": 1.0,
"content": "methods for exponential integrators, we propose first-order, second-order and third-order DPM-",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
290,
505,
302
],
"spans": [
{
"bbox": [
105,
290,
505,
302
],
"score": 1.0,
"content": "Solver to approximate the exponentially weighted integral of noise prediction models with theoretical",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
301,
505,
313
],
"spans": [
{
"bbox": [
105,
301,
505,
313
],
"score": 1.0,
"content": "convergence guarantee. We propose both handcrafted and adaptive step size schedule, and apply",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
311,
506,
324
],
"spans": [
{
"bbox": [
105,
311,
506,
324
],
"score": 1.0,
"content": "DPM-Solver for both continuous-time and discrete-time DPMs. Our experimental results show that",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
322,
506,
335
],
"spans": [
{
"bbox": [
105,
322,
506,
335
],
"score": 1.0,
"content": "DPM-Solver can generate high-quality samples in around 10 function evaluations on various datasets,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
333,
506,
346
],
"spans": [
{
"bbox": [
105,
333,
180,
346
],
"score": 1.0,
"content": "and it can achieve",
"type": "text"
},
{
"bbox": [
180,
334,
217,
344
],
"score": 0.9,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
218,
333,
506,
346
],
"score": 1.0,
"content": "speedup compared with previous state-of-the-art training-free samplers.",
"type": "text"
}
],
"index": 21
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
349,
505,
416
],
"lines": [
{
"bbox": [
105,
349,
506,
362
],
"spans": [
{
"bbox": [
105,
349,
506,
362
],
"score": 1.0,
"content": "Limitations and broader impact Despite the promising speedup performance, DPM-Solver is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
361,
505,
372
],
"spans": [
{
"bbox": [
106,
361,
505,
372
],
"score": 1.0,
"content": "designed for fast sampling, which may be not suitable for accelerating the likelihood evaluations of",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
371,
505,
384
],
"spans": [
{
"bbox": [
105,
371,
505,
384
],
"score": 1.0,
"content": "DPMs. Besides, compared to the commonly-used GANs, diffusion models with DPM-Solver are still",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
381,
506,
396
],
"spans": [
{
"bbox": [
105,
381,
506,
396
],
"score": 1.0,
"content": "not fast enough for real-time applications. In addition, like other deep generative models, DPMs may",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
393,
505,
406
],
"spans": [
{
"bbox": [
105,
393,
505,
406
],
"score": 1.0,
"content": "be used to generate adverse fake contents, and the proposed solver may further amplify the potential",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
404,
410,
417
],
"spans": [
{
"bbox": [
105,
404,
410,
417
],
"score": 1.0,
"content": "undesirable influence of deep generative models for malicious applications.",
"type": "text"
}
],
"index": 27
}
],
"index": 24.5
},
{
"type": "title",
"bbox": [
108,
432,
207,
445
],
"lines": [
{
"bbox": [
106,
431,
208,
448
],
"spans": [
{
"bbox": [
106,
431,
208,
448
],
"score": 1.0,
"content": "Acknowledgements",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
457,
506,
545
],
"lines": [
{
"bbox": [
105,
457,
506,
471
],
"spans": [
{
"bbox": [
105,
457,
506,
471
],
"score": 1.0,
"content": "This work was supported by National Key Research and Development Project of China (No.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
469,
506,
481
],
"spans": [
{
"bbox": [
105,
469,
506,
481
],
"score": 1.0,
"content": "2021ZD0110502); NSF of China Projects (Nos. 62061136001, 61620106010, 62076145, U19B2034,",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
479,
506,
493
],
"spans": [
{
"bbox": [
105,
479,
506,
493
],
"score": 1.0,
"content": "U1811461, U19A2081, 6197222, 62106120); Beijing NSF Project (No. JQ19016); Beijing Outstand-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
491,
505,
503
],
"spans": [
{
"bbox": [
106,
491,
505,
503
],
"score": 1.0,
"content": "ing Young Scientist Program NO. BJJWZYJH012019100020098; a grant from Tsinghua Institute",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "for Guo Qiang; the NVIDIA NVAIL Program with GPU/DGX Acceleration; the High Performance",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
512,
506,
525
],
"spans": [
{
"bbox": [
105,
512,
506,
525
],
"score": 1.0,
"content": "Computing Center, Tsinghua University; the Fundamental Research Funds for the Central Universi-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
522,
505,
537
],
"spans": [
{
"bbox": [
105,
522,
505,
537
],
"score": 1.0,
"content": "ties, and the Research Funds of Renmin University of China (22XNKJ13). J.Z is also supported by",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
534,
181,
546
],
"spans": [
{
"bbox": [
106,
534,
181,
546
],
"score": 1.0,
"content": "the XPlorer Prize.",
"type": "text"
}
],
"index": 36
}
],
"index": 32.5
},
{
"type": "title",
"bbox": [
107,
563,
163,
575
],
"lines": [
{
"bbox": [
106,
561,
165,
577
],
"spans": [
{
"bbox": [
106,
561,
165,
577
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "text",
"bbox": [
110,
582,
506,
722
],
"lines": [
{
"bbox": [
110,
581,
505,
598
],
"spans": [
{
"bbox": [
110,
581,
505,
598
],
"score": 1.0,
"content": "[1] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
592,
506,
609
],
"spans": [
{
"bbox": [
126,
592,
506,
609
],
"score": 1.0,
"content": "using nonequilibrium thermodynamics,” in International Conference on Machine Learning.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
126,
604,
250,
619
],
"spans": [
{
"bbox": [
126,
604,
250,
619
],
"score": 1.0,
"content": "PMLR, 2015, pp. 2256–2265.",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
111,
625,
506,
639
],
"spans": [
{
"bbox": [
111,
625,
506,
639
],
"score": 1.0,
"content": "[2] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
127,
636,
383,
650
],
"spans": [
{
"bbox": [
127,
636,
383,
650
],
"score": 1.0,
"content": "Information Processing Systems, vol. 33, 2020, pp. 6840–6851.",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
110,
657,
505,
671
],
"spans": [
{
"bbox": [
110,
657,
505,
671
],
"score": 1.0,
"content": "[3] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
127,
668,
505,
682
],
"spans": [
{
"bbox": [
127,
668,
505,
682
],
"score": 1.0,
"content": "generative modeling through stochastic differential equations,” in International Conference on",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
679,
260,
693
],
"spans": [
{
"bbox": [
127,
679,
260,
693
],
"score": 1.0,
"content": "Learning Representations, 2021.",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
111,
699,
505,
714
],
"spans": [
{
"bbox": [
111,
699,
505,
714
],
"score": 1.0,
"content": "[4] P. Dhariwal and A. Q. Nichol, “Diffusion models beat GANs on image synthesis,” in Advances",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
126,
711,
424,
724
],
"spans": [
{
"bbox": [
126,
711,
424,
724
],
"score": 1.0,
"content": "in Neural Information Processing Systems, vol. 34, 2021, pp. 8780–8794.",
"type": "text"
}
],
"index": 47
}
],
"index": 42.5
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
301,
741,
311,
750
],
"lines": [
{
"bbox": [
299,
740,
313,
755
],
"spans": [
{
"bbox": [
299,
740,
313,
755
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 14
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
106,
72,
505,
182
],
"lines": [
{
"bbox": [
105,
72,
505,
86
],
"spans": [
{
"bbox": [
105,
72,
329,
86
],
"score": 1.0,
"content": "Specifically, we use the discrete-time model trained by",
"type": "text"
},
{
"bbox": [
329,
73,
356,
84
],
"score": 0.9,
"content": "L _ { \\mathrm { s i m p l e } }",
"type": "inline_equation"
},
{
"bbox": [
357,
72,
505,
86
],
"score": 1.0,
"content": "in [2] on the CIFAR-10 dataset with",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
83,
506,
96
],
"spans": [
{
"bbox": [
105,
83,
506,
96
],
"score": 1.0,
"content": "linear noise schedule; the discrete-time model in [19] on CelebA 64x64 [39] with linear noise",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
93,
506,
109
],
"spans": [
{
"bbox": [
105,
93,
292,
109
],
"score": 1.0,
"content": "schedule; the discrete-time model trained by",
"type": "text"
},
{
"bbox": [
292,
95,
319,
106
],
"score": 0.89,
"content": "L _ { \\mathrm { h y b r i d } }",
"type": "inline_equation"
},
{
"bbox": [
320,
93,
506,
109
],
"score": 1.0,
"content": "in [16] on ImageNet 64x64 [26] with cosine",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
106,
105,
505,
118
],
"spans": [
{
"bbox": [
106,
105,
505,
118
],
"score": 1.0,
"content": "noise schedule; the discrete-time model with classifier guidance in [4] on ImageNet 128x128 [26]",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
117,
505,
128
],
"spans": [
{
"bbox": [
106,
117,
426,
128
],
"score": 1.0,
"content": "with linear noise schedule; the discrete-time model in [4] on LSUN bedroom",
"type": "text"
},
{
"bbox": [
426,
117,
464,
127
],
"score": 0.29,
"content": "2 5 6 \\times 2 5 6",
"type": "inline_equation"
},
{
"bbox": [
464,
117,
505,
128
],
"score": 1.0,
"content": "[40] with",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
127,
506,
139
],
"spans": [
{
"bbox": [
105,
127,
506,
139
],
"score": 1.0,
"content": "linear noise schedule. For the models trained on ImageNet, we only use their “mean” model and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
138,
506,
150
],
"spans": [
{
"bbox": [
106,
138,
506,
150
],
"score": 1.0,
"content": "omit the “variance” model. As shown in Fig. 2, on all datasets, DPM-Solver can obtain reasonable",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
149,
505,
161
],
"spans": [
{
"bbox": [
105,
149,
505,
161
],
"score": 1.0,
"content": "samples within 12 steps (FID 4.65 on CIFAR-10, FID 3.71 on CelebA 64x64 and FID 19.97 on",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
159,
506,
172
],
"spans": [
{
"bbox": [
106,
159,
268,
172
],
"score": 1.0,
"content": "ImageNet 64x64, FID 4.08 on ImageNet",
"type": "text"
},
{
"bbox": [
268,
160,
304,
171
],
"score": 0.33,
"content": "1 2 8 \\mathbf { x } 1 2 8 _ { \\rho }",
"type": "inline_equation"
},
{
"bbox": [
304,
159,
345,
172
],
"score": 1.0,
"content": "), which is",
"type": "text"
},
{
"bbox": [
345,
160,
383,
171
],
"score": 0.92,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
383,
159,
506,
172
],
"score": 1.0,
"content": "faster than the previous fastest",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
170,
496,
185
],
"spans": [
{
"bbox": [
105,
170,
496,
185
],
"score": 1.0,
"content": "training-free sampler. DPM-Solver even outperforms GGDM, which requires additional training.",
"type": "text"
}
],
"index": 9
}
],
"index": 4.5,
"bbox_fs": [
105,
72,
506,
185
]
},
{
"type": "title",
"bbox": [
107,
199,
187,
212
],
"lines": [
{
"bbox": [
104,
197,
190,
214
],
"spans": [
{
"bbox": [
104,
197,
190,
214
],
"score": 1.0,
"content": "6 Conclusions",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
107,
224,
505,
345
],
"lines": [
{
"bbox": [
105,
223,
506,
237
],
"spans": [
{
"bbox": [
105,
223,
506,
237
],
"score": 1.0,
"content": "We tackle the problem of fast and training-free sampling from DPMs. We propose DPM-Solver,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
235,
506,
248
],
"spans": [
{
"bbox": [
105,
235,
506,
248
],
"score": 1.0,
"content": "a fast dedicated training-free solver of diffusion ODEs for fast sampling of DPMs in around 10",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
246,
505,
258
],
"spans": [
{
"bbox": [
105,
246,
505,
258
],
"score": 1.0,
"content": "steps of function evaluations. DPM-Solver leverages the semi-linearity of diffusion ODEs and",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
257,
505,
269
],
"spans": [
{
"bbox": [
105,
257,
505,
269
],
"score": 1.0,
"content": "it directly approximates a simplified formulation of exact solutions of diffusion ODEs, which",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
268,
506,
280
],
"spans": [
{
"bbox": [
106,
268,
506,
280
],
"score": 1.0,
"content": "consists of an exponentially weighted integral of the noise prediction model. Inspired by numerical",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
279,
506,
291
],
"spans": [
{
"bbox": [
105,
279,
506,
291
],
"score": 1.0,
"content": "methods for exponential integrators, we propose first-order, second-order and third-order DPM-",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
290,
505,
302
],
"spans": [
{
"bbox": [
105,
290,
505,
302
],
"score": 1.0,
"content": "Solver to approximate the exponentially weighted integral of noise prediction models with theoretical",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
301,
505,
313
],
"spans": [
{
"bbox": [
105,
301,
505,
313
],
"score": 1.0,
"content": "convergence guarantee. We propose both handcrafted and adaptive step size schedule, and apply",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
311,
506,
324
],
"spans": [
{
"bbox": [
105,
311,
506,
324
],
"score": 1.0,
"content": "DPM-Solver for both continuous-time and discrete-time DPMs. Our experimental results show that",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
322,
506,
335
],
"spans": [
{
"bbox": [
105,
322,
506,
335
],
"score": 1.0,
"content": "DPM-Solver can generate high-quality samples in around 10 function evaluations on various datasets,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
333,
506,
346
],
"spans": [
{
"bbox": [
105,
333,
180,
346
],
"score": 1.0,
"content": "and it can achieve",
"type": "text"
},
{
"bbox": [
180,
334,
217,
344
],
"score": 0.9,
"content": "4 \\sim 1 6 \\times",
"type": "inline_equation"
},
{
"bbox": [
218,
333,
506,
346
],
"score": 1.0,
"content": "speedup compared with previous state-of-the-art training-free samplers.",
"type": "text"
}
],
"index": 21
}
],
"index": 16,
"bbox_fs": [
105,
223,
506,
346
]
},
{
"type": "text",
"bbox": [
107,
349,
505,
416
],
"lines": [
{
"bbox": [
105,
349,
506,
362
],
"spans": [
{
"bbox": [
105,
349,
506,
362
],
"score": 1.0,
"content": "Limitations and broader impact Despite the promising speedup performance, DPM-Solver is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
361,
505,
372
],
"spans": [
{
"bbox": [
106,
361,
505,
372
],
"score": 1.0,
"content": "designed for fast sampling, which may be not suitable for accelerating the likelihood evaluations of",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
371,
505,
384
],
"spans": [
{
"bbox": [
105,
371,
505,
384
],
"score": 1.0,
"content": "DPMs. Besides, compared to the commonly-used GANs, diffusion models with DPM-Solver are still",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
381,
506,
396
],
"spans": [
{
"bbox": [
105,
381,
506,
396
],
"score": 1.0,
"content": "not fast enough for real-time applications. In addition, like other deep generative models, DPMs may",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
393,
505,
406
],
"spans": [
{
"bbox": [
105,
393,
505,
406
],
"score": 1.0,
"content": "be used to generate adverse fake contents, and the proposed solver may further amplify the potential",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
404,
410,
417
],
"spans": [
{
"bbox": [
105,
404,
410,
417
],
"score": 1.0,
"content": "undesirable influence of deep generative models for malicious applications.",
"type": "text"
}
],
"index": 27
}
],
"index": 24.5,
"bbox_fs": [
105,
349,
506,
417
]
},
{
"type": "title",
"bbox": [
108,
432,
207,
445
],
"lines": [
{
"bbox": [
106,
431,
208,
448
],
"spans": [
{
"bbox": [
106,
431,
208,
448
],
"score": 1.0,
"content": "Acknowledgements",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
457,
506,
545
],
"lines": [
{
"bbox": [
105,
457,
506,
471
],
"spans": [
{
"bbox": [
105,
457,
506,
471
],
"score": 1.0,
"content": "This work was supported by National Key Research and Development Project of China (No.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
469,
506,
481
],
"spans": [
{
"bbox": [
105,
469,
506,
481
],
"score": 1.0,
"content": "2021ZD0110502); NSF of China Projects (Nos. 62061136001, 61620106010, 62076145, U19B2034,",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
479,
506,
493
],
"spans": [
{
"bbox": [
105,
479,
506,
493
],
"score": 1.0,
"content": "U1811461, U19A2081, 6197222, 62106120); Beijing NSF Project (No. JQ19016); Beijing Outstand-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
491,
505,
503
],
"spans": [
{
"bbox": [
106,
491,
505,
503
],
"score": 1.0,
"content": "ing Young Scientist Program NO. BJJWZYJH012019100020098; a grant from Tsinghua Institute",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "for Guo Qiang; the NVIDIA NVAIL Program with GPU/DGX Acceleration; the High Performance",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
512,
506,
525
],
"spans": [
{
"bbox": [
105,
512,
506,
525
],
"score": 1.0,
"content": "Computing Center, Tsinghua University; the Fundamental Research Funds for the Central Universi-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
522,
505,
537
],
"spans": [
{
"bbox": [
105,
522,
505,
537
],
"score": 1.0,
"content": "ties, and the Research Funds of Renmin University of China (22XNKJ13). J.Z is also supported by",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
534,
181,
546
],
"spans": [
{
"bbox": [
106,
534,
181,
546
],
"score": 1.0,
"content": "the XPlorer Prize.",
"type": "text"
}
],
"index": 36
}
],
"index": 32.5,
"bbox_fs": [
105,
457,
506,
546
]
},
{
"type": "title",
"bbox": [
107,
563,
163,
575
],
"lines": [
{
"bbox": [
106,
561,
165,
577
],
"spans": [
{
"bbox": [
106,
561,
165,
577
],
"score": 1.0,
"content": "References",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "list",
"bbox": [
110,
582,
506,
722
],
"lines": [
{
"bbox": [
110,
581,
505,
598
],
"spans": [
{
"bbox": [
110,
581,
505,
598
],
"score": 1.0,
"content": "[1] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning",
"type": "text"
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
126,
592,
506,
609
],
"spans": [
{
"bbox": [
126,
592,
506,
609
],
"score": 1.0,
"content": "using nonequilibrium thermodynamics,” in International Conference on Machine Learning.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
126,
604,
250,
619
],
"spans": [
{
"bbox": [
126,
604,
250,
619
],
"score": 1.0,
"content": "PMLR, 2015, pp. 2256–2265.",
"type": "text"
}
],
"index": 40,
"is_list_end_line": true
},
{
"bbox": [
111,
625,
506,
639
],
"spans": [
{
"bbox": [
111,
625,
506,
639
],
"score": 1.0,
"content": "[2] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural",
"type": "text"
}
],
"index": 41,
"is_list_start_line": true
},
{
"bbox": [
127,
636,
383,
650
],
"spans": [
{
"bbox": [
127,
636,
383,
650
],
"score": 1.0,
"content": "Information Processing Systems, vol. 33, 2020, pp. 6840–6851.",
"type": "text"
}
],
"index": 42,
"is_list_end_line": true
},
{
"bbox": [
110,
657,
505,
671
],
"spans": [
{
"bbox": [
110,
657,
505,
671
],
"score": 1.0,
"content": "[3] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based",
"type": "text"
}
],
"index": 43,
"is_list_start_line": true
},
{
"bbox": [
127,
668,
505,
682
],
"spans": [
{
"bbox": [
127,
668,
505,
682
],
"score": 1.0,
"content": "generative modeling through stochastic differential equations,” in International Conference on",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
127,
679,
260,
693
],
"spans": [
{
"bbox": [
127,
679,
260,
693
],
"score": 1.0,
"content": "Learning Representations, 2021.",
"type": "text"
}
],
"index": 45,
"is_list_end_line": true
},
{
"bbox": [
111,
699,
505,
714
],
"spans": [
{
"bbox": [
111,
699,
505,
714
],
"score": 1.0,
"content": "[4] P. Dhariwal and A. Q. Nichol, “Diffusion models beat GANs on image synthesis,” in Advances",
"type": "text"
}
],
"index": 46,
"is_list_start_line": true
},
{
"bbox": [
126,
711,
424,
724
],
"spans": [
{
"bbox": [
126,
711,
424,
724
],
"score": 1.0,
"content": "in Neural Information Processing Systems, vol. 34, 2021, pp. 8780–8794.",
"type": "text"
}
],
"index": 47,
"is_list_end_line": true
},
{
"bbox": [
109,
71,
507,
88
],
"spans": [
{
"bbox": [
109,
71,
507,
88
],
"score": 1.0,
"content": "[5] C. Meng, Y. Song, J. Song, J. Wu, J.-Y. Zhu, and S. Ermon, “SDEdit: Image synthesis and editing",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
124,
82,
507,
99
],
"spans": [
{
"bbox": [
124,
82,
507,
99
],
"score": 1.0,
"content": "with stochastic differential equations,” in International Conference on Learning Representations,",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
127,
93,
154,
108
],
"spans": [
{
"bbox": [
127,
93,
154,
108
],
"score": 1.0,
"content": "2022.",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_end_line": true
},
{
"bbox": [
110,
114,
505,
128
],
"spans": [
{
"bbox": [
110,
114,
505,
128
],
"score": 1.0,
"content": "[6] J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion",
"type": "text",
"cross_page": true
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
127,
124,
327,
138
],
"spans": [
{
"bbox": [
127,
124,
327,
138
],
"score": 1.0,
"content": "models,” arXiv preprint arXiv:2204.03458, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
109,
143,
505,
160
],
"spans": [
{
"bbox": [
109,
143,
505,
160
],
"score": 1.0,
"content": "[7] A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
125,
155,
414,
171
],
"spans": [
{
"bbox": [
125,
155,
414,
171
],
"score": 1.0,
"content": "generation with CLIP latents,” arXiv preprint arXiv:2204.06125, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 6,
"is_list_end_line": true
},
{
"bbox": [
109,
173,
506,
190
],
"spans": [
{
"bbox": [
109,
173,
506,
190
],
"score": 1.0,
"content": "[8] N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating",
"type": "text",
"cross_page": true
}
],
"index": 7,
"is_list_start_line": true
},
{
"bbox": [
125,
185,
507,
201
],
"spans": [
{
"bbox": [
125,
185,
507,
201
],
"score": 1.0,
"content": "gradients for waveform generation,” in International Conference on Learning Representations,",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
127,
196,
155,
211
],
"spans": [
{
"bbox": [
127,
196,
155,
211
],
"score": 1.0,
"content": "2021.",
"type": "text",
"cross_page": true
}
],
"index": 9,
"is_list_end_line": true
},
{
"bbox": [
110,
217,
505,
231
],
"spans": [
{
"bbox": [
110,
217,
505,
231
],
"score": 1.0,
"content": "[9] N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, N. Dehak, and W. Chan, “Wavegrad",
"type": "text",
"cross_page": true
}
],
"index": 10,
"is_list_start_line": true
},
{
"bbox": [
127,
228,
506,
242
],
"spans": [
{
"bbox": [
127,
228,
506,
242
],
"score": 1.0,
"content": "2: Iterative refinement for text-to-speech synthesis,” in International Speech Communication",
"type": "text",
"cross_page": true
}
],
"index": 11
},
{
"bbox": [
126,
239,
270,
253
],
"spans": [
{
"bbox": [
126,
239,
270,
253
],
"score": 1.0,
"content": "Association, 2021, pp. 3765–3769.",
"type": "text",
"cross_page": true
}
],
"index": 12,
"is_list_end_line": true
},
{
"bbox": [
105,
258,
506,
272
],
"spans": [
{
"bbox": [
105,
258,
506,
272
],
"score": 1.0,
"content": "[10] D. P. Kingma, T. Salimans, B. Poole, and J. Ho, “Variational diffusion models,” in Advances in",
"type": "text",
"cross_page": true
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
127,
270,
316,
282
],
"spans": [
{
"bbox": [
127,
270,
316,
282
],
"score": 1.0,
"content": "Neural Information Processing Systems, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 14,
"is_list_end_line": true
},
{
"bbox": [
104,
288,
507,
303
],
"spans": [
{
"bbox": [
104,
288,
507,
303
],
"score": 1.0,
"content": "[11] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville,",
"type": "text",
"cross_page": true
}
],
"index": 15,
"is_list_start_line": true
},
{
"bbox": [
126,
299,
505,
315
],
"spans": [
{
"bbox": [
126,
299,
505,
315
],
"score": 1.0,
"content": "and Y. Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing",
"type": "text",
"cross_page": true
}
],
"index": 16
},
{
"bbox": [
127,
311,
288,
325
],
"spans": [
{
"bbox": [
127,
311,
288,
325
],
"score": 1.0,
"content": "Systems, vol. 27, 2014, pp. 2672–2680.",
"type": "text",
"cross_page": true
}
],
"index": 17,
"is_list_end_line": true
},
{
"bbox": [
106,
331,
505,
343
],
"spans": [
{
"bbox": [
106,
331,
505,
343
],
"score": 1.0,
"content": "[12] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in International Conference",
"type": "text",
"cross_page": true
}
],
"index": 18,
"is_list_start_line": true
},
{
"bbox": [
126,
341,
274,
354
],
"spans": [
{
"bbox": [
126,
341,
274,
354
],
"score": 1.0,
"content": "on Learning Representations, 2014.",
"type": "text",
"cross_page": true
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
105,
361,
506,
375
],
"spans": [
{
"bbox": [
105,
361,
506,
375
],
"score": 1.0,
"content": "[13] T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in",
"type": "text",
"cross_page": true
}
],
"index": 20,
"is_list_start_line": true
},
{
"bbox": [
126,
372,
375,
386
],
"spans": [
{
"bbox": [
126,
372,
375,
386
],
"score": 1.0,
"content": "International Conference on Learning Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 21,
"is_list_end_line": true
},
{
"bbox": [
105,
391,
506,
405
],
"spans": [
{
"bbox": [
105,
391,
506,
405
],
"score": 1.0,
"content": "[14] E. Luhman and T. Luhman, “Knowledge distillation in iterative generative models for improved",
"type": "text",
"cross_page": true
}
],
"index": 22,
"is_list_start_line": true
},
{
"bbox": [
127,
402,
362,
416
],
"spans": [
{
"bbox": [
127,
402,
362,
416
],
"score": 1.0,
"content": "sampling speed,” arXiv preprint arXiv:2101.02388, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 23,
"is_list_end_line": true
},
{
"bbox": [
104,
420,
507,
437
],
"spans": [
{
"bbox": [
104,
420,
507,
437
],
"score": 1.0,
"content": "[15] R. San-Roman, E. Nachmani, and L. Wolf, “Noise estimation for generative diffusion models,”",
"type": "text",
"cross_page": true
}
],
"index": 24,
"is_list_start_line": true
},
{
"bbox": [
127,
434,
290,
446
],
"spans": [
{
"bbox": [
127,
434,
290,
446
],
"score": 1.0,
"content": "arXiv preprint arXiv:2104.02600, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 25,
"is_list_end_line": true
},
{
"bbox": [
106,
452,
507,
466
],
"spans": [
{
"bbox": [
106,
452,
507,
466
],
"score": 1.0,
"content": "[16] A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in Interna-",
"type": "text",
"cross_page": true
}
],
"index": 26,
"is_list_start_line": true
},
{
"bbox": [
126,
464,
424,
478
],
"spans": [
{
"bbox": [
126,
464,
424,
478
],
"score": 1.0,
"content": "tional Conference on Machine Learning. PMLR, 2021, pp. 8162–8171.",
"type": "text",
"cross_page": true
}
],
"index": 27,
"is_list_end_line": true
},
{
"bbox": [
106,
484,
505,
496
],
"spans": [
{
"bbox": [
106,
484,
505,
496
],
"score": 1.0,
"content": "[17] M. W. Lam, J. Wang, R. Huang, D. Su, and D. Yu, “Bilateral denoising diffusion models,” arXiv",
"type": "text",
"cross_page": true
}
],
"index": 28,
"is_list_start_line": true
},
{
"bbox": [
124,
494,
268,
509
],
"spans": [
{
"bbox": [
124,
494,
268,
509
],
"score": 1.0,
"content": "preprint arXiv:2108.11514, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
106,
514,
506,
528
],
"spans": [
{
"bbox": [
106,
514,
506,
528
],
"score": 1.0,
"content": "[18] D. Watson, W. Chan, J. Ho, and M. Norouzi, “Learning fast samplers for diffusion models by dif-",
"type": "text",
"cross_page": true
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
127,
525,
507,
539
],
"spans": [
{
"bbox": [
127,
525,
507,
539
],
"score": 1.0,
"content": "ferentiating through sample quality,” in International Conference on Learning Representations,",
"type": "text",
"cross_page": true
}
],
"index": 31
},
{
"bbox": [
126,
534,
155,
549
],
"spans": [
{
"bbox": [
126,
534,
155,
549
],
"score": 1.0,
"content": "2022.",
"type": "text",
"cross_page": true
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
106,
555,
505,
569
],
"spans": [
{
"bbox": [
106,
555,
505,
569
],
"score": 1.0,
"content": "[19] J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International",
"type": "text",
"cross_page": true
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
127,
567,
321,
580
],
"spans": [
{
"bbox": [
127,
567,
321,
580
],
"score": 1.0,
"content": "Conference on Learning Representations, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
106,
586,
506,
600
],
"spans": [
{
"bbox": [
106,
586,
506,
600
],
"score": 1.0,
"content": "[20] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, “Gotta go fast",
"type": "text",
"cross_page": true
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
126,
597,
486,
610
],
"spans": [
{
"bbox": [
126,
597,
486,
610
],
"score": 1.0,
"content": "when generating data with score-based models,” arXiv preprint arXiv:2105.14080, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 36,
"is_list_end_line": true
},
{
"bbox": [
106,
617,
506,
631
],
"spans": [
{
"bbox": [
106,
617,
506,
631
],
"score": 1.0,
"content": "[21] F. Bao, C. Li, J. Zhu, and B. Zhang, “Analytic-DPM: An analytic estimate of the optimal",
"type": "text",
"cross_page": true
}
],
"index": 37,
"is_list_start_line": true
},
{
"bbox": [
124,
626,
506,
643
],
"spans": [
{
"bbox": [
124,
626,
506,
643
],
"score": 1.0,
"content": "reverse variance in diffusion probabilistic models,” in International Conference on Learning",
"type": "text",
"cross_page": true
}
],
"index": 38
},
{
"bbox": [
126,
638,
223,
652
],
"spans": [
{
"bbox": [
126,
638,
223,
652
],
"score": 1.0,
"content": "Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
106,
657,
506,
671
],
"spans": [
{
"bbox": [
106,
657,
506,
671
],
"score": 1.0,
"content": "[22] L. Liu, Y. Ren, Z. Lin, and Z. Zhao, “Pseudo numerical methods for diffusion models on",
"type": "text",
"cross_page": true
}
],
"index": 40,
"is_list_start_line": true
},
{
"bbox": [
127,
669,
435,
683
],
"spans": [
{
"bbox": [
127,
669,
435,
683
],
"score": 1.0,
"content": "manifolds,” in International Conference on Learning Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 41,
"is_list_end_line": true
},
{
"bbox": [
105,
687,
507,
704
],
"spans": [
{
"bbox": [
105,
687,
507,
704
],
"score": 1.0,
"content": "[23] V. Popov, I. Vovk, V. Gogoryan, T. Sadekova, M. Kudinov, and J. Wei, “Diffusion-based voice",
"type": "text",
"cross_page": true
}
],
"index": 42,
"is_list_start_line": true
},
{
"bbox": [
125,
698,
507,
715
],
"spans": [
{
"bbox": [
125,
698,
507,
715
],
"score": 1.0,
"content": "conversion with fast maximum likelihood sampling scheme,” in International Conference on",
"type": "text",
"cross_page": true
}
],
"index": 43
},
{
"bbox": [
126,
711,
261,
725
],
"spans": [
{
"bbox": [
126,
711,
261,
725
],
"score": 1.0,
"content": "Learning Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 44,
"is_list_end_line": true
},
{
"bbox": [
105,
72,
506,
88
],
"spans": [
{
"bbox": [
105,
72,
506,
88
],
"score": 1.0,
"content": "[24] H. Tachibana, M. Go, M. Inahara, Y. Katayama, and Y. Watanabe, “Itô-Taylor sampling",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
126,
83,
506,
99
],
"spans": [
{
"bbox": [
126,
83,
506,
99
],
"score": 1.0,
"content": "scheme for denoising diffusion probabilistic models using ideal derivatives,” arXiv preprint",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
127,
94,
232,
108
],
"spans": [
{
"bbox": [
127,
94,
232,
108
],
"score": 1.0,
"content": "arXiv:2112.13339, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_end_line": true
},
{
"bbox": [
104,
112,
508,
132
],
"spans": [
{
"bbox": [
104,
112,
508,
132
],
"score": 1.0,
"content": "[25] M. Hochbruck and A. Ostermann, “Exponential integrators,” Acta Numerica, vol. 19, pp.",
"type": "text",
"cross_page": true
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
127,
126,
193,
140
],
"spans": [
{
"bbox": [
127,
126,
193,
140
],
"score": 1.0,
"content": "209–286, 2010.",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
106,
146,
506,
160
],
"spans": [
{
"bbox": [
106,
146,
506,
160
],
"score": 1.0,
"content": "[26] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
127,
157,
506,
171
],
"spans": [
{
"bbox": [
127,
157,
506,
171
],
"score": 1.0,
"content": "image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition. IEEE,",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
127,
168,
209,
182
],
"spans": [
{
"bbox": [
127,
168,
209,
182
],
"score": 1.0,
"content": "2009, pp. 248–255.",
"type": "text",
"cross_page": true
}
],
"index": 7,
"is_list_end_line": true
},
{
"bbox": [
105,
188,
507,
204
],
"spans": [
{
"bbox": [
105,
188,
507,
204
],
"score": 1.0,
"content": "[27] P. E. Kloeden and E. Platen, Numerical Solution of Stochastic Differential Equations. Springer,",
"type": "text",
"cross_page": true
}
],
"index": 8,
"is_list_start_line": true
},
{
"bbox": [
126,
199,
154,
214
],
"spans": [
{
"bbox": [
126,
199,
154,
214
],
"score": 1.0,
"content": "1992.",
"type": "text",
"cross_page": true
}
],
"index": 9,
"is_list_end_line": true
},
{
"bbox": [
106,
221,
507,
234
],
"spans": [
{
"bbox": [
106,
221,
507,
234
],
"score": 1.0,
"content": "[28] J. R. Dormand and P. J. Prince, “A family of embedded Runge-Kutta formulae,” Journal of",
"type": "text",
"cross_page": true
}
],
"index": 10,
"is_list_start_line": true
},
{
"bbox": [
125,
231,
416,
246
],
"spans": [
{
"bbox": [
125,
231,
416,
246
],
"score": 1.0,
"content": "computational and applied mathematics, vol. 6, no. 1, pp. 19–26, 1980.",
"type": "text",
"cross_page": true
}
],
"index": 11,
"is_list_end_line": true
},
{
"bbox": [
106,
252,
489,
266
],
"spans": [
{
"bbox": [
106,
252,
489,
266
],
"score": 1.0,
"content": "[29] A. Krizhevsky, “Learning multiple layers of features from tiny images,” Tech. Rep., 2009.",
"type": "text",
"cross_page": true
}
],
"index": 12,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
105,
272,
507,
286
],
"spans": [
{
"bbox": [
105,
272,
507,
286
],
"score": 1.0,
"content": "[30] K. Atkinson, W. Han, and D. E. Stewart, Numerical solution of ordinary differential equations.",
"type": "text",
"cross_page": true
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
127,
284,
273,
296
],
"spans": [
{
"bbox": [
127,
284,
273,
296
],
"score": 1.0,
"content": "John Wiley & Sons, 2011, vol. 108.",
"type": "text",
"cross_page": true
}
],
"index": 14,
"is_list_end_line": true
},
{
"bbox": [
104,
303,
506,
318
],
"spans": [
{
"bbox": [
104,
303,
506,
318
],
"score": 1.0,
"content": "[31] M. Hochbruck and A. Ostermann, “Explicit exponential Runge-Kutta methods for semilinear",
"type": "text",
"cross_page": true
}
],
"index": 15,
"is_list_start_line": true
},
{
"bbox": [
125,
314,
508,
330
],
"spans": [
{
"bbox": [
125,
314,
508,
330
],
"score": 1.0,
"content": "parabolic problems,” SIAM Journal on Numerical Analysis, vol. 43, no. 3, pp. 1069–1090, 2005.",
"type": "text",
"cross_page": true
}
],
"index": 16
},
{
"bbox": [
106,
335,
506,
349
],
"spans": [
{
"bbox": [
106,
335,
506,
349
],
"score": 1.0,
"content": "[32] V. T. Luan, “Efficient exponential Runge-Kutta methods of high order: Construction and",
"type": "text",
"cross_page": true
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
127,
347,
455,
361
],
"spans": [
{
"bbox": [
127,
347,
455,
361
],
"score": 1.0,
"content": "implementation,” BIT Numerical Mathematics, vol. 61, no. 2, pp. 535–560, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
106,
367,
506,
381
],
"spans": [
{
"bbox": [
106,
367,
506,
381
],
"score": 1.0,
"content": "[33] F. Bao, C. Li, J. Sun, J. Zhu, and B. Zhang, “Estimating the optimal covariance with imperfect",
"type": "text",
"cross_page": true
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
126,
378,
452,
392
],
"spans": [
{
"bbox": [
126,
378,
452,
392
],
"score": 1.0,
"content": "mean in diffusion probabilistic models,” arXiv preprint arXiv:2206.07309, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
105,
397,
506,
413
],
"spans": [
{
"bbox": [
105,
397,
506,
413
],
"score": 1.0,
"content": "[34] A. Vahdat, K. Kreis, and J. Kautz, “Score-based generative modeling in latent space,” in",
"type": "text",
"cross_page": true
}
],
"index": 21,
"is_list_start_line": true
},
{
"bbox": [
124,
408,
478,
424
],
"spans": [
{
"bbox": [
124,
408,
478,
424
],
"score": 1.0,
"content": "Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 11 287–11 302.",
"type": "text",
"cross_page": true
}
],
"index": 22,
"is_list_end_line": true
},
{
"bbox": [
106,
430,
506,
444
],
"spans": [
{
"bbox": [
106,
430,
506,
444
],
"score": 1.0,
"content": "[35] T. Dockhorn, A. Vahdat, and K. Kreis, “Score-based generative modeling with critically-damped",
"type": "text",
"cross_page": true
}
],
"index": 23,
"is_list_start_line": true
},
{
"bbox": [
127,
441,
470,
455
],
"spans": [
{
"bbox": [
127,
441,
470,
455
],
"score": 1.0,
"content": "Langevin diffusion,” in International Conference on Learning Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 24,
"is_list_end_line": true
},
{
"bbox": [
104,
460,
506,
477
],
"spans": [
{
"bbox": [
104,
460,
506,
477
],
"score": 1.0,
"content": "[36] Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learning trilemma with denoising",
"type": "text",
"cross_page": true
}
],
"index": 25,
"is_list_start_line": true
},
{
"bbox": [
127,
474,
458,
485
],
"spans": [
{
"bbox": [
127,
474,
458,
485
],
"score": 1.0,
"content": "diffusion GANs,” in International Conference on Learning Representations, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 26,
"is_list_end_line": true
},
{
"bbox": [
104,
492,
506,
507
],
"spans": [
{
"bbox": [
104,
492,
506,
507
],
"score": 1.0,
"content": "[37] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two",
"type": "text",
"cross_page": true
}
],
"index": 27,
"is_list_start_line": true
},
{
"bbox": [
127,
504,
506,
518
],
"spans": [
{
"bbox": [
127,
504,
506,
518
],
"score": 1.0,
"content": "time-scale update rule converge to a local Nash equilibrium,” in Advances in Neural Information",
"type": "text",
"cross_page": true
}
],
"index": 28
},
{
"bbox": [
126,
514,
508,
530
],
"spans": [
{
"bbox": [
126,
514,
508,
530
],
"score": 1.0,
"content": "Processing Systems, I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N.",
"type": "text",
"cross_page": true
}
],
"index": 29
},
{
"bbox": [
127,
525,
397,
540
],
"spans": [
{
"bbox": [
127,
525,
397,
540
],
"score": 1.0,
"content": "Vishwanathan, and R. Garnett, Eds., vol. 30, 2017, pp. 6626–6637.",
"type": "text",
"cross_page": true
}
],
"index": 30,
"is_list_end_line": true
},
{
"bbox": [
106,
546,
506,
560
],
"spans": [
{
"bbox": [
106,
546,
506,
560
],
"score": 1.0,
"content": "[38] Z. Kong and W. Ping, “On fast sampling of diffusion probabilistic models,” arXiv preprint",
"type": "text",
"cross_page": true
}
],
"index": 31,
"is_list_start_line": true
},
{
"bbox": [
127,
559,
231,
570
],
"spans": [
{
"bbox": [
127,
559,
231,
570
],
"score": 1.0,
"content": "arXiv:2106.00132, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
104,
576,
506,
593
],
"spans": [
{
"bbox": [
104,
576,
506,
593
],
"score": 1.0,
"content": "[39] Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in Proceedings",
"type": "text",
"cross_page": true
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
125,
588,
455,
603
],
"spans": [
{
"bbox": [
125,
588,
455,
603
],
"score": 1.0,
"content": "of the IEEE International Conference on Computer Vision, 2015, pp. 3730–3738.",
"type": "text",
"cross_page": true
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
105,
608,
506,
622
],
"spans": [
{
"bbox": [
105,
608,
506,
622
],
"score": 1.0,
"content": "[40] F. Yu, A. Seff, Y. Zhang, S. Song, T. Funkhouser, and J. Xiao, “LSUN: Construction of",
"type": "text",
"cross_page": true
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
125,
619,
506,
636
],
"spans": [
{
"bbox": [
125,
619,
506,
636
],
"score": 1.0,
"content": "a large-scale image dataset using deep learning with humans in the loop,” arXiv preprint",
"type": "text",
"cross_page": true
}
],
"index": 36
},
{
"bbox": [
127,
631,
231,
645
],
"spans": [
{
"bbox": [
127,
631,
231,
645
],
"score": 1.0,
"content": "arXiv:1506.03365, 2015.",
"type": "text",
"cross_page": true
}
],
"index": 37,
"is_list_end_line": true
},
{
"bbox": [
105,
652,
506,
665
],
"spans": [
{
"bbox": [
105,
652,
506,
665
],
"score": 1.0,
"content": "[41] Y. Song, C. Durkan, I. Murray, and S. Ermon, “Maximum likelihood training of score-based",
"type": "text",
"cross_page": true
}
],
"index": 38,
"is_list_start_line": true
},
{
"bbox": [
125,
660,
507,
679
],
"spans": [
{
"bbox": [
125,
660,
507,
679
],
"score": 1.0,
"content": "diffusion models,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp.",
"type": "text",
"cross_page": true
}
],
"index": 39
},
{
"bbox": [
127,
675,
178,
686
],
"spans": [
{
"bbox": [
127,
675,
178,
686
],
"score": 1.0,
"content": "1415–1428.",
"type": "text",
"cross_page": true
}
],
"index": 40,
"is_list_end_line": true
},
{
"bbox": [
104,
692,
508,
710
],
"spans": [
{
"bbox": [
104,
692,
508,
710
],
"score": 1.0,
"content": "[42] K. Yang, J. Yau, L. Fei-Fei, J. Deng, and O. Russakovsky, “A study of face obfuscation in",
"type": "text",
"cross_page": true
}
],
"index": 41,
"is_list_start_line": true
},
{
"bbox": [
126,
704,
340,
721
],
"spans": [
{
"bbox": [
126,
704,
340,
721
],
"score": 1.0,
"content": "ImageNet,” arXiv preprint arXiv:2103.06191, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 42,
"is_list_end_line": true
}
],
"index": 42.5,
"bbox_fs": [
110,
581,
506,
724
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
52,
507,
726
],
"lines": [
{
"bbox": [
109,
71,
507,
88
],
"spans": [
{
"bbox": [
109,
71,
507,
88
],
"score": 1.0,
"content": "[5] C. Meng, Y. Song, J. Song, J. Wu, J.-Y. Zhu, and S. Ermon, “SDEdit: Image synthesis and editing",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
124,
82,
507,
99
],
"spans": [
{
"bbox": [
124,
82,
507,
99
],
"score": 1.0,
"content": "with stochastic differential equations,” in International Conference on Learning Representations,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
93,
154,
108
],
"spans": [
{
"bbox": [
127,
93,
154,
108
],
"score": 1.0,
"content": "2022.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
110,
114,
505,
128
],
"spans": [
{
"bbox": [
110,
114,
505,
128
],
"score": 1.0,
"content": "[6] J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
124,
327,
138
],
"spans": [
{
"bbox": [
127,
124,
327,
138
],
"score": 1.0,
"content": "models,” arXiv preprint arXiv:2204.03458, 2022.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
109,
143,
505,
160
],
"spans": [
{
"bbox": [
109,
143,
505,
160
],
"score": 1.0,
"content": "[7] A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
125,
155,
414,
171
],
"spans": [
{
"bbox": [
125,
155,
414,
171
],
"score": 1.0,
"content": "generation with CLIP latents,” arXiv preprint arXiv:2204.06125, 2022.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
109,
173,
506,
190
],
"spans": [
{
"bbox": [
109,
173,
506,
190
],
"score": 1.0,
"content": "[8] N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
125,
185,
507,
201
],
"spans": [
{
"bbox": [
125,
185,
507,
201
],
"score": 1.0,
"content": "gradients for waveform generation,” in International Conference on Learning Representations,",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
127,
196,
155,
211
],
"spans": [
{
"bbox": [
127,
196,
155,
211
],
"score": 1.0,
"content": "2021.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
110,
217,
505,
231
],
"spans": [
{
"bbox": [
110,
217,
505,
231
],
"score": 1.0,
"content": "[9] N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, N. Dehak, and W. Chan, “Wavegrad",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
127,
228,
506,
242
],
"spans": [
{
"bbox": [
127,
228,
506,
242
],
"score": 1.0,
"content": "2: Iterative refinement for text-to-speech synthesis,” in International Speech Communication",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
126,
239,
270,
253
],
"spans": [
{
"bbox": [
126,
239,
270,
253
],
"score": 1.0,
"content": "Association, 2021, pp. 3765–3769.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
258,
506,
272
],
"spans": [
{
"bbox": [
105,
258,
506,
272
],
"score": 1.0,
"content": "[10] D. P. Kingma, T. Salimans, B. Poole, and J. Ho, “Variational diffusion models,” in Advances in",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
270,
316,
282
],
"spans": [
{
"bbox": [
127,
270,
316,
282
],
"score": 1.0,
"content": "Neural Information Processing Systems, 2021.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
104,
288,
507,
303
],
"spans": [
{
"bbox": [
104,
288,
507,
303
],
"score": 1.0,
"content": "[11] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville,",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
126,
299,
505,
315
],
"spans": [
{
"bbox": [
126,
299,
505,
315
],
"score": 1.0,
"content": "and Y. Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
127,
311,
288,
325
],
"spans": [
{
"bbox": [
127,
311,
288,
325
],
"score": 1.0,
"content": "Systems, vol. 27, 2014, pp. 2672–2680.",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
331,
505,
343
],
"spans": [
{
"bbox": [
106,
331,
505,
343
],
"score": 1.0,
"content": "[12] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in International Conference",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
126,
341,
274,
354
],
"spans": [
{
"bbox": [
126,
341,
274,
354
],
"score": 1.0,
"content": "on Learning Representations, 2014.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
361,
506,
375
],
"spans": [
{
"bbox": [
105,
361,
506,
375
],
"score": 1.0,
"content": "[13] T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
126,
372,
375,
386
],
"spans": [
{
"bbox": [
126,
372,
375,
386
],
"score": 1.0,
"content": "International Conference on Learning Representations, 2022.",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
391,
506,
405
],
"spans": [
{
"bbox": [
105,
391,
506,
405
],
"score": 1.0,
"content": "[14] E. Luhman and T. Luhman, “Knowledge distillation in iterative generative models for improved",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
127,
402,
362,
416
],
"spans": [
{
"bbox": [
127,
402,
362,
416
],
"score": 1.0,
"content": "sampling speed,” arXiv preprint arXiv:2101.02388, 2021.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
104,
420,
507,
437
],
"spans": [
{
"bbox": [
104,
420,
507,
437
],
"score": 1.0,
"content": "[15] R. San-Roman, E. Nachmani, and L. Wolf, “Noise estimation for generative diffusion models,”",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
127,
434,
290,
446
],
"spans": [
{
"bbox": [
127,
434,
290,
446
],
"score": 1.0,
"content": "arXiv preprint arXiv:2104.02600, 2021.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
452,
507,
466
],
"spans": [
{
"bbox": [
106,
452,
507,
466
],
"score": 1.0,
"content": "[16] A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in Interna-",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
126,
464,
424,
478
],
"spans": [
{
"bbox": [
126,
464,
424,
478
],
"score": 1.0,
"content": "tional Conference on Machine Learning. PMLR, 2021, pp. 8162–8171.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
484,
505,
496
],
"spans": [
{
"bbox": [
106,
484,
505,
496
],
"score": 1.0,
"content": "[17] M. W. Lam, J. Wang, R. Huang, D. Su, and D. Yu, “Bilateral denoising diffusion models,” arXiv",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
124,
494,
268,
509
],
"spans": [
{
"bbox": [
124,
494,
268,
509
],
"score": 1.0,
"content": "preprint arXiv:2108.11514, 2021.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
514,
506,
528
],
"spans": [
{
"bbox": [
106,
514,
506,
528
],
"score": 1.0,
"content": "[18] D. Watson, W. Chan, J. Ho, and M. Norouzi, “Learning fast samplers for diffusion models by dif-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
127,
525,
507,
539
],
"spans": [
{
"bbox": [
127,
525,
507,
539
],
"score": 1.0,
"content": "ferentiating through sample quality,” in International Conference on Learning Representations,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
126,
534,
155,
549
],
"spans": [
{
"bbox": [
126,
534,
155,
549
],
"score": 1.0,
"content": "2022.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
555,
505,
569
],
"spans": [
{
"bbox": [
106,
555,
505,
569
],
"score": 1.0,
"content": "[19] J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
127,
567,
321,
580
],
"spans": [
{
"bbox": [
127,
567,
321,
580
],
"score": 1.0,
"content": "Conference on Learning Representations, 2021.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
586,
506,
600
],
"spans": [
{
"bbox": [
106,
586,
506,
600
],
"score": 1.0,
"content": "[20] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, “Gotta go fast",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
126,
597,
486,
610
],
"spans": [
{
"bbox": [
126,
597,
486,
610
],
"score": 1.0,
"content": "when generating data with score-based models,” arXiv preprint arXiv:2105.14080, 2021.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
617,
506,
631
],
"spans": [
{
"bbox": [
106,
617,
506,
631
],
"score": 1.0,
"content": "[21] F. Bao, C. Li, J. Zhu, and B. Zhang, “Analytic-DPM: An analytic estimate of the optimal",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
124,
626,
506,
643
],
"spans": [
{
"bbox": [
124,
626,
506,
643
],
"score": 1.0,
"content": "reverse variance in diffusion probabilistic models,” in International Conference on Learning",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
126,
638,
223,
652
],
"spans": [
{
"bbox": [
126,
638,
223,
652
],
"score": 1.0,
"content": "Representations, 2022.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
657,
506,
671
],
"spans": [
{
"bbox": [
106,
657,
506,
671
],
"score": 1.0,
"content": "[22] L. Liu, Y. Ren, Z. Lin, and Z. Zhao, “Pseudo numerical methods for diffusion models on",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
127,
669,
435,
683
],
"spans": [
{
"bbox": [
127,
669,
435,
683
],
"score": 1.0,
"content": "manifolds,” in International Conference on Learning Representations, 2022.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
687,
507,
704
],
"spans": [
{
"bbox": [
105,
687,
507,
704
],
"score": 1.0,
"content": "[23] V. Popov, I. Vovk, V. Gogoryan, T. Sadekova, M. Kudinov, and J. Wei, “Diffusion-based voice",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
125,
698,
507,
715
],
"spans": [
{
"bbox": [
125,
698,
507,
715
],
"score": 1.0,
"content": "conversion with fast maximum likelihood sampling scheme,” in International Conference on",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
126,
711,
261,
725
],
"spans": [
{
"bbox": [
126,
711,
261,
725
],
"score": 1.0,
"content": "Learning Representations, 2022.",
"type": "text"
}
],
"index": 44
}
],
"index": 22
}
],
"page_idx": 10,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
741,
311,
751
],
"lines": [
{
"bbox": [
299,
740,
312,
754
],
"spans": [
{
"bbox": [
299,
740,
312,
754
],
"score": 1.0,
"content": "11",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
105,
52,
507,
726
],
"lines": [],
"index": 22,
"bbox_fs": [
104,
71,
507,
725
],
"lines_deleted": true
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
51,
507,
719
],
"lines": [
{
"bbox": [
105,
72,
506,
88
],
"spans": [
{
"bbox": [
105,
72,
506,
88
],
"score": 1.0,
"content": "[24] H. Tachibana, M. Go, M. Inahara, Y. Katayama, and Y. Watanabe, “Itô-Taylor sampling",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
126,
83,
506,
99
],
"spans": [
{
"bbox": [
126,
83,
506,
99
],
"score": 1.0,
"content": "scheme for denoising diffusion probabilistic models using ideal derivatives,” arXiv preprint",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
127,
94,
232,
108
],
"spans": [
{
"bbox": [
127,
94,
232,
108
],
"score": 1.0,
"content": "arXiv:2112.13339, 2021.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
104,
112,
508,
132
],
"spans": [
{
"bbox": [
104,
112,
508,
132
],
"score": 1.0,
"content": "[25] M. Hochbruck and A. Ostermann, “Exponential integrators,” Acta Numerica, vol. 19, pp.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
127,
126,
193,
140
],
"spans": [
{
"bbox": [
127,
126,
193,
140
],
"score": 1.0,
"content": "209–286, 2010.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
146,
506,
160
],
"spans": [
{
"bbox": [
106,
146,
506,
160
],
"score": 1.0,
"content": "[26] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
127,
157,
506,
171
],
"spans": [
{
"bbox": [
127,
157,
506,
171
],
"score": 1.0,
"content": "image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition. IEEE,",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
127,
168,
209,
182
],
"spans": [
{
"bbox": [
127,
168,
209,
182
],
"score": 1.0,
"content": "2009, pp. 248–255.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
188,
507,
204
],
"spans": [
{
"bbox": [
105,
188,
507,
204
],
"score": 1.0,
"content": "[27] P. E. Kloeden and E. Platen, Numerical Solution of Stochastic Differential Equations. Springer,",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
126,
199,
154,
214
],
"spans": [
{
"bbox": [
126,
199,
154,
214
],
"score": 1.0,
"content": "1992.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
221,
507,
234
],
"spans": [
{
"bbox": [
106,
221,
507,
234
],
"score": 1.0,
"content": "[28] J. R. Dormand and P. J. Prince, “A family of embedded Runge-Kutta formulae,” Journal of",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
125,
231,
416,
246
],
"spans": [
{
"bbox": [
125,
231,
416,
246
],
"score": 1.0,
"content": "computational and applied mathematics, vol. 6, no. 1, pp. 19–26, 1980.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
252,
489,
266
],
"spans": [
{
"bbox": [
106,
252,
489,
266
],
"score": 1.0,
"content": "[29] A. Krizhevsky, “Learning multiple layers of features from tiny images,” Tech. Rep., 2009.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
272,
507,
286
],
"spans": [
{
"bbox": [
105,
272,
507,
286
],
"score": 1.0,
"content": "[30] K. Atkinson, W. Han, and D. E. Stewart, Numerical solution of ordinary differential equations.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
127,
284,
273,
296
],
"spans": [
{
"bbox": [
127,
284,
273,
296
],
"score": 1.0,
"content": "John Wiley & Sons, 2011, vol. 108.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
104,
303,
506,
318
],
"spans": [
{
"bbox": [
104,
303,
506,
318
],
"score": 1.0,
"content": "[31] M. Hochbruck and A. Ostermann, “Explicit exponential Runge-Kutta methods for semilinear",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
125,
314,
508,
330
],
"spans": [
{
"bbox": [
125,
314,
508,
330
],
"score": 1.0,
"content": "parabolic problems,” SIAM Journal on Numerical Analysis, vol. 43, no. 3, pp. 1069–1090, 2005.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
335,
506,
349
],
"spans": [
{
"bbox": [
106,
335,
506,
349
],
"score": 1.0,
"content": "[32] V. T. Luan, “Efficient exponential Runge-Kutta methods of high order: Construction and",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
127,
347,
455,
361
],
"spans": [
{
"bbox": [
127,
347,
455,
361
],
"score": 1.0,
"content": "implementation,” BIT Numerical Mathematics, vol. 61, no. 2, pp. 535–560, 2021.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
367,
506,
381
],
"spans": [
{
"bbox": [
106,
367,
506,
381
],
"score": 1.0,
"content": "[33] F. Bao, C. Li, J. Sun, J. Zhu, and B. Zhang, “Estimating the optimal covariance with imperfect",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
126,
378,
452,
392
],
"spans": [
{
"bbox": [
126,
378,
452,
392
],
"score": 1.0,
"content": "mean in diffusion probabilistic models,” arXiv preprint arXiv:2206.07309, 2022.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
397,
506,
413
],
"spans": [
{
"bbox": [
105,
397,
506,
413
],
"score": 1.0,
"content": "[34] A. Vahdat, K. Kreis, and J. Kautz, “Score-based generative modeling in latent space,” in",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
124,
408,
478,
424
],
"spans": [
{
"bbox": [
124,
408,
478,
424
],
"score": 1.0,
"content": "Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 11 287–11 302.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
430,
506,
444
],
"spans": [
{
"bbox": [
106,
430,
506,
444
],
"score": 1.0,
"content": "[35] T. Dockhorn, A. Vahdat, and K. Kreis, “Score-based generative modeling with critically-damped",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
127,
441,
470,
455
],
"spans": [
{
"bbox": [
127,
441,
470,
455
],
"score": 1.0,
"content": "Langevin diffusion,” in International Conference on Learning Representations, 2022.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
104,
460,
506,
477
],
"spans": [
{
"bbox": [
104,
460,
506,
477
],
"score": 1.0,
"content": "[36] Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learning trilemma with denoising",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
127,
474,
458,
485
],
"spans": [
{
"bbox": [
127,
474,
458,
485
],
"score": 1.0,
"content": "diffusion GANs,” in International Conference on Learning Representations, 2022.",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
104,
492,
506,
507
],
"spans": [
{
"bbox": [
104,
492,
506,
507
],
"score": 1.0,
"content": "[37] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
127,
504,
506,
518
],
"spans": [
{
"bbox": [
127,
504,
506,
518
],
"score": 1.0,
"content": "time-scale update rule converge to a local Nash equilibrium,” in Advances in Neural Information",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
126,
514,
508,
530
],
"spans": [
{
"bbox": [
126,
514,
508,
530
],
"score": 1.0,
"content": "Processing Systems, I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
127,
525,
397,
540
],
"spans": [
{
"bbox": [
127,
525,
397,
540
],
"score": 1.0,
"content": "Vishwanathan, and R. Garnett, Eds., vol. 30, 2017, pp. 6626–6637.",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
546,
506,
560
],
"spans": [
{
"bbox": [
106,
546,
506,
560
],
"score": 1.0,
"content": "[38] Z. Kong and W. Ping, “On fast sampling of diffusion probabilistic models,” arXiv preprint",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
127,
559,
231,
570
],
"spans": [
{
"bbox": [
127,
559,
231,
570
],
"score": 1.0,
"content": "arXiv:2106.00132, 2021.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
104,
576,
506,
593
],
"spans": [
{
"bbox": [
104,
576,
506,
593
],
"score": 1.0,
"content": "[39] Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in Proceedings",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
125,
588,
455,
603
],
"spans": [
{
"bbox": [
125,
588,
455,
603
],
"score": 1.0,
"content": "of the IEEE International Conference on Computer Vision, 2015, pp. 3730–3738.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
608,
506,
622
],
"spans": [
{
"bbox": [
105,
608,
506,
622
],
"score": 1.0,
"content": "[40] F. Yu, A. Seff, Y. Zhang, S. Song, T. Funkhouser, and J. Xiao, “LSUN: Construction of",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
125,
619,
506,
636
],
"spans": [
{
"bbox": [
125,
619,
506,
636
],
"score": 1.0,
"content": "a large-scale image dataset using deep learning with humans in the loop,” arXiv preprint",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
127,
631,
231,
645
],
"spans": [
{
"bbox": [
127,
631,
231,
645
],
"score": 1.0,
"content": "arXiv:1506.03365, 2015.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
652,
506,
665
],
"spans": [
{
"bbox": [
105,
652,
506,
665
],
"score": 1.0,
"content": "[41] Y. Song, C. Durkan, I. Murray, and S. Ermon, “Maximum likelihood training of score-based",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
125,
660,
507,
679
],
"spans": [
{
"bbox": [
125,
660,
507,
679
],
"score": 1.0,
"content": "diffusion models,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
127,
675,
178,
686
],
"spans": [
{
"bbox": [
127,
675,
178,
686
],
"score": 1.0,
"content": "1415–1428.",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
104,
692,
508,
710
],
"spans": [
{
"bbox": [
104,
692,
508,
710
],
"score": 1.0,
"content": "[42] K. Yang, J. Yau, L. Fei-Fei, J. Deng, and O. Russakovsky, “A study of face obfuscation in",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
126,
704,
340,
721
],
"spans": [
{
"bbox": [
126,
704,
340,
721
],
"score": 1.0,
"content": "ImageNet,” arXiv preprint arXiv:2103.06191, 2021.",
"type": "text"
}
],
"index": 42
}
],
"index": 21
}
],
"page_idx": 11,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
741,
311,
751
],
"lines": [
{
"bbox": [
299,
740,
313,
754
],
"spans": [
{
"bbox": [
299,
740,
313,
754
],
"score": 1.0,
"content": "12",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
105,
51,
507,
719
],
"lines": [],
"index": 21,
"bbox_fs": [
104,
72,
508,
721
],
"lines_deleted": true
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
71,
156,
84
],
"lines": [
{
"bbox": [
106,
70,
158,
85
],
"spans": [
{
"bbox": [
106,
70,
158,
85
],
"score": 1.0,
"content": "Checklist",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
131,
92,
208,
103
],
"lines": [
{
"bbox": [
129,
91,
210,
105
],
"spans": [
{
"bbox": [
129,
91,
210,
105
],
"score": 1.0,
"content": "1. For all authors...",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "text",
"bbox": [
146,
107,
505,
190
],
"lines": [
{
"bbox": [
145,
106,
505,
119
],
"spans": [
{
"bbox": [
145,
106,
505,
119
],
"score": 1.0,
"content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
162,
117,
288,
130
],
"spans": [
{
"bbox": [
162,
117,
288,
130
],
"score": 1.0,
"content": "contributions and scope? [Yes]",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
144,
130,
434,
144
],
"spans": [
{
"bbox": [
144,
130,
434,
144
],
"score": 1.0,
"content": "(b) Did you describe the limitations of your work? [Yes] See section 6.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
144,
142,
506,
158
],
"spans": [
{
"bbox": [
144,
142,
506,
158
],
"score": 1.0,
"content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
160,
155,
203,
166
],
"spans": [
{
"bbox": [
160,
155,
203,
166
],
"score": 1.0,
"content": "section 6.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
144,
167,
506,
181
],
"spans": [
{
"bbox": [
144,
167,
506,
181
],
"score": 1.0,
"content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
161,
178,
214,
191
],
"spans": [
{
"bbox": [
161,
178,
214,
191
],
"score": 1.0,
"content": "them? [Yes]",
"type": "text"
}
],
"index": 8
}
],
"index": 5
},
{
"type": "text",
"bbox": [
131,
193,
302,
205
],
"lines": [
{
"bbox": [
129,
192,
304,
207
],
"spans": [
{
"bbox": [
129,
192,
304,
207
],
"score": 1.0,
"content": "2. If you are including theoretical results...",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
146,
208,
505,
244
],
"lines": [
{
"bbox": [
145,
207,
506,
223
],
"spans": [
{
"bbox": [
145,
207,
506,
223
],
"score": 1.0,
"content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] See Ap-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
160,
219,
204,
231
],
"spans": [
{
"bbox": [
160,
219,
204,
231
],
"score": 1.0,
"content": "pendix B.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
144,
231,
491,
246
],
"spans": [
{
"bbox": [
144,
231,
491,
246
],
"score": 1.0,
"content": "(b) Did you include complete proofs of all theoretical results? [Yes] See Appendix B.",
"type": "text"
}
],
"index": 12
}
],
"index": 11
},
{
"type": "text",
"bbox": [
131,
247,
241,
258
],
"lines": [
{
"bbox": [
128,
245,
243,
261
],
"spans": [
{
"bbox": [
128,
245,
243,
261
],
"score": 1.0,
"content": "3. If you ran experiments...",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
146,
263,
505,
421
],
"lines": [
{
"bbox": [
146,
262,
506,
275
],
"spans": [
{
"bbox": [
146,
262,
506,
275
],
"score": 1.0,
"content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
162,
273,
505,
284
],
"spans": [
{
"bbox": [
162,
273,
505,
284
],
"score": 1.0,
"content": "mental results (either in the supplemental material or as a URL)? [Yes] Code is attached",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
161,
284,
360,
297
],
"spans": [
{
"bbox": [
161,
284,
360,
297
],
"score": 1.0,
"content": "in the supplemental materials, with the appendix.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
145,
296,
505,
310
],
"spans": [
{
"bbox": [
145,
296,
505,
310
],
"score": 1.0,
"content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
161,
307,
505,
320
],
"spans": [
{
"bbox": [
161,
307,
505,
320
],
"score": 1.0,
"content": "chosen)? [Yes] Our method is training-free. But we also report the hyperparameters",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
161,
318,
338,
331
],
"spans": [
{
"bbox": [
161,
318,
338,
331
],
"score": 1.0,
"content": "for evaluations used in our proposed solver.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
145,
331,
507,
345
],
"spans": [
{
"bbox": [
145,
331,
507,
345
],
"score": 1.0,
"content": "(c) Did you report error bars (e.g., with respect to the random seed after running ex-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
161,
343,
505,
354
],
"spans": [
{
"bbox": [
161,
343,
505,
354
],
"score": 1.0,
"content": "periments multiple times)? [No] We observe that the standard deviation of the FID",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
162,
354,
505,
365
],
"spans": [
{
"bbox": [
162,
354,
505,
365
],
"score": 1.0,
"content": "evaluations of DPM-Solver are rather small (mainly less than 0.01) because the FID is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
162,
364,
505,
376
],
"spans": [
{
"bbox": [
162,
364,
505,
376
],
"score": 1.0,
"content": "already averaged over 50K samples, following existing work [18, 20, 21]. The small",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
161,
375,
367,
387
],
"spans": [
{
"bbox": [
161,
375,
367,
387
],
"score": 1.0,
"content": "standard deviation does not change the conclusion.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
146,
388,
505,
401
],
"spans": [
{
"bbox": [
146,
388,
505,
401
],
"score": 1.0,
"content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
162,
399,
506,
412
],
"spans": [
{
"bbox": [
162,
399,
506,
412
],
"score": 1.0,
"content": "of GPUs, internal cluster, or cloud provider)? [Yes] The GPU type and amount is",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
162,
410,
259,
422
],
"spans": [
{
"bbox": [
162,
410,
259,
422
],
"score": 1.0,
"content": "detailed in Appendix E.",
"type": "text"
}
],
"index": 27
}
],
"index": 20.5
},
{
"type": "text",
"bbox": [
127,
425,
504,
437
],
"lines": [
{
"bbox": [
129,
423,
506,
439
],
"spans": [
{
"bbox": [
129,
423,
506,
439
],
"score": 1.0,
"content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
"type": "text"
}
],
"index": 28
}
],
"index": 28
},
{
"type": "text",
"bbox": [
146,
440,
505,
546
],
"lines": [
{
"bbox": [
146,
440,
424,
452
],
"spans": [
{
"bbox": [
146,
440,
424,
452
],
"score": 1.0,
"content": "(a) If your work uses existing assets, did you cite the creators? [Yes]",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
145,
453,
424,
465
],
"spans": [
{
"bbox": [
145,
453,
424,
465
],
"score": 1.0,
"content": "(b) Did you mention the license of the assets? [Yes] See Appendix E",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
145,
465,
505,
479
],
"spans": [
{
"bbox": [
145,
465,
505,
479
],
"score": 1.0,
"content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
161,
475,
368,
489
],
"spans": [
{
"bbox": [
161,
475,
368,
489
],
"score": 1.0,
"content": "We include our code in the supplemental materials.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
145,
489,
505,
502
],
"spans": [
{
"bbox": [
145,
489,
505,
502
],
"score": 1.0,
"content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
162,
501,
506,
512
],
"spans": [
{
"bbox": [
162,
501,
506,
512
],
"score": 1.0,
"content": "using/curating? [No] All of the datasets used in the experiments are publicly available.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
145,
513,
506,
526
],
"spans": [
{
"bbox": [
145,
513,
506,
526
],
"score": 1.0,
"content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
162,
523,
505,
536
],
"spans": [
{
"bbox": [
162,
523,
505,
536
],
"score": 1.0,
"content": "information or offensive content? [Yes] We mentioned the human privacy issues of the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
162,
534,
297,
547
],
"spans": [
{
"bbox": [
162,
534,
297,
547
],
"score": 1.0,
"content": "ImageNet dataset in Appendix E.",
"type": "text"
}
],
"index": 37
}
],
"index": 33
},
{
"type": "text",
"bbox": [
131,
550,
432,
561
],
"lines": [
{
"bbox": [
128,
548,
433,
564
],
"spans": [
{
"bbox": [
128,
548,
433,
564
],
"score": 1.0,
"content": "5. If you used crowdsourcing or conducted research with human subjects...",
"type": "text"
}
],
"index": 38
}
],
"index": 38
},
{
"type": "text",
"bbox": [
146,
564,
505,
635
],
"lines": [
{
"bbox": [
145,
564,
506,
577
],
"spans": [
{
"bbox": [
145,
564,
506,
577
],
"score": 1.0,
"content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
162,
576,
237,
588
],
"spans": [
{
"bbox": [
162,
576,
237,
588
],
"score": 1.0,
"content": "applicable? [N/A]",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
145,
588,
505,
601
],
"spans": [
{
"bbox": [
145,
588,
505,
601
],
"score": 1.0,
"content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
161,
599,
342,
612
],
"spans": [
{
"bbox": [
161,
599,
342,
612
],
"score": 1.0,
"content": "Board (IRB) approvals, if applicable? [N/A]",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
147,
612,
505,
625
],
"spans": [
{
"bbox": [
147,
612,
505,
625
],
"score": 1.0,
"content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
162,
624,
333,
636
],
"spans": [
{
"bbox": [
162,
624,
333,
636
],
"score": 1.0,
"content": "spent on participant compensation? [N/A]",
"type": "text"
}
],
"index": 44
}
],
"index": 41.5
}
],
"page_idx": 12,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
741,
311,
750
],
"lines": [
{
"bbox": [
299,
740,
313,
754
],
"spans": [
{
"bbox": [
299,
740,
313,
754
],
"score": 1.0,
"content": "13",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
71,
156,
84
],
"lines": [
{
"bbox": [
106,
70,
158,
85
],
"spans": [
{
"bbox": [
106,
70,
158,
85
],
"score": 1.0,
"content": "Checklist",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
131,
92,
208,
103
],
"lines": [
{
"bbox": [
129,
91,
210,
105
],
"spans": [
{
"bbox": [
129,
91,
210,
105
],
"score": 1.0,
"content": "1. For all authors...",
"type": "text"
}
],
"index": 1
}
],
"index": 1,
"bbox_fs": [
129,
91,
210,
105
]
},
{
"type": "list",
"bbox": [
146,
107,
505,
190
],
"lines": [
{
"bbox": [
145,
106,
505,
119
],
"spans": [
{
"bbox": [
145,
106,
505,
119
],
"score": 1.0,
"content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
"type": "text"
}
],
"index": 2,
"is_list_start_line": true
},
{
"bbox": [
162,
117,
288,
130
],
"spans": [
{
"bbox": [
162,
117,
288,
130
],
"score": 1.0,
"content": "contributions and scope? [Yes]",
"type": "text"
}
],
"index": 3,
"is_list_end_line": true
},
{
"bbox": [
144,
130,
434,
144
],
"spans": [
{
"bbox": [
144,
130,
434,
144
],
"score": 1.0,
"content": "(b) Did you describe the limitations of your work? [Yes] See section 6.",
"type": "text"
}
],
"index": 4,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
144,
142,
506,
158
],
"spans": [
{
"bbox": [
144,
142,
506,
158
],
"score": 1.0,
"content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See",
"type": "text"
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
160,
155,
203,
166
],
"spans": [
{
"bbox": [
160,
155,
203,
166
],
"score": 1.0,
"content": "section 6.",
"type": "text"
}
],
"index": 6,
"is_list_end_line": true
},
{
"bbox": [
144,
167,
506,
181
],
"spans": [
{
"bbox": [
144,
167,
506,
181
],
"score": 1.0,
"content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
"type": "text"
}
],
"index": 7,
"is_list_start_line": true
},
{
"bbox": [
161,
178,
214,
191
],
"spans": [
{
"bbox": [
161,
178,
214,
191
],
"score": 1.0,
"content": "them? [Yes]",
"type": "text"
}
],
"index": 8,
"is_list_end_line": true
}
],
"index": 5,
"bbox_fs": [
144,
106,
506,
191
]
},
{
"type": "text",
"bbox": [
131,
193,
302,
205
],
"lines": [
{
"bbox": [
129,
192,
304,
207
],
"spans": [
{
"bbox": [
129,
192,
304,
207
],
"score": 1.0,
"content": "2. If you are including theoretical results...",
"type": "text"
}
],
"index": 9
}
],
"index": 9,
"bbox_fs": [
129,
192,
304,
207
]
},
{
"type": "text",
"bbox": [
146,
208,
505,
244
],
"lines": [
{
"bbox": [
145,
207,
506,
223
],
"spans": [
{
"bbox": [
145,
207,
506,
223
],
"score": 1.0,
"content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] See Ap-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
160,
219,
204,
231
],
"spans": [
{
"bbox": [
160,
219,
204,
231
],
"score": 1.0,
"content": "pendix B.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
144,
231,
491,
246
],
"spans": [
{
"bbox": [
144,
231,
491,
246
],
"score": 1.0,
"content": "(b) Did you include complete proofs of all theoretical results? [Yes] See Appendix B.",
"type": "text"
}
],
"index": 12
}
],
"index": 11,
"bbox_fs": [
144,
207,
506,
246
]
},
{
"type": "text",
"bbox": [
131,
247,
241,
258
],
"lines": [
{
"bbox": [
128,
245,
243,
261
],
"spans": [
{
"bbox": [
128,
245,
243,
261
],
"score": 1.0,
"content": "3. If you ran experiments...",
"type": "text"
}
],
"index": 13
}
],
"index": 13,
"bbox_fs": [
128,
245,
243,
261
]
},
{
"type": "list",
"bbox": [
146,
263,
505,
421
],
"lines": [
{
"bbox": [
146,
262,
506,
275
],
"spans": [
{
"bbox": [
146,
262,
506,
275
],
"score": 1.0,
"content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
"type": "text"
}
],
"index": 14,
"is_list_start_line": true
},
{
"bbox": [
162,
273,
505,
284
],
"spans": [
{
"bbox": [
162,
273,
505,
284
],
"score": 1.0,
"content": "mental results (either in the supplemental material or as a URL)? [Yes] Code is attached",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
161,
284,
360,
297
],
"spans": [
{
"bbox": [
161,
284,
360,
297
],
"score": 1.0,
"content": "in the supplemental materials, with the appendix.",
"type": "text"
}
],
"index": 16,
"is_list_end_line": true
},
{
"bbox": [
145,
296,
505,
310
],
"spans": [
{
"bbox": [
145,
296,
505,
310
],
"score": 1.0,
"content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were",
"type": "text"
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
161,
307,
505,
320
],
"spans": [
{
"bbox": [
161,
307,
505,
320
],
"score": 1.0,
"content": "chosen)? [Yes] Our method is training-free. But we also report the hyperparameters",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
161,
318,
338,
331
],
"spans": [
{
"bbox": [
161,
318,
338,
331
],
"score": 1.0,
"content": "for evaluations used in our proposed solver.",
"type": "text"
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
145,
331,
507,
345
],
"spans": [
{
"bbox": [
145,
331,
507,
345
],
"score": 1.0,
"content": "(c) Did you report error bars (e.g., with respect to the random seed after running ex-",
"type": "text"
}
],
"index": 20,
"is_list_start_line": true
},
{
"bbox": [
161,
343,
505,
354
],
"spans": [
{
"bbox": [
161,
343,
505,
354
],
"score": 1.0,
"content": "periments multiple times)? [No] We observe that the standard deviation of the FID",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
162,
354,
505,
365
],
"spans": [
{
"bbox": [
162,
354,
505,
365
],
"score": 1.0,
"content": "evaluations of DPM-Solver are rather small (mainly less than 0.01) because the FID is",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
162,
364,
505,
376
],
"spans": [
{
"bbox": [
162,
364,
505,
376
],
"score": 1.0,
"content": "already averaged over 50K samples, following existing work [18, 20, 21]. The small",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
161,
375,
367,
387
],
"spans": [
{
"bbox": [
161,
375,
367,
387
],
"score": 1.0,
"content": "standard deviation does not change the conclusion.",
"type": "text"
}
],
"index": 24,
"is_list_end_line": true
},
{
"bbox": [
146,
388,
505,
401
],
"spans": [
{
"bbox": [
146,
388,
505,
401
],
"score": 1.0,
"content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
"type": "text"
}
],
"index": 25,
"is_list_start_line": true
},
{
"bbox": [
162,
399,
506,
412
],
"spans": [
{
"bbox": [
162,
399,
506,
412
],
"score": 1.0,
"content": "of GPUs, internal cluster, or cloud provider)? [Yes] The GPU type and amount is",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
162,
410,
259,
422
],
"spans": [
{
"bbox": [
162,
410,
259,
422
],
"score": 1.0,
"content": "detailed in Appendix E.",
"type": "text"
}
],
"index": 27,
"is_list_end_line": true
}
],
"index": 20.5,
"bbox_fs": [
145,
262,
507,
422
]
},
{
"type": "text",
"bbox": [
127,
425,
504,
437
],
"lines": [
{
"bbox": [
129,
423,
506,
439
],
"spans": [
{
"bbox": [
129,
423,
506,
439
],
"score": 1.0,
"content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
"type": "text"
}
],
"index": 28
}
],
"index": 28,
"bbox_fs": [
129,
423,
506,
439
]
},
{
"type": "list",
"bbox": [
146,
440,
505,
546
],
"lines": [
{
"bbox": [
146,
440,
424,
452
],
"spans": [
{
"bbox": [
146,
440,
424,
452
],
"score": 1.0,
"content": "(a) If your work uses existing assets, did you cite the creators? [Yes]",
"type": "text"
}
],
"index": 29,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
145,
453,
424,
465
],
"spans": [
{
"bbox": [
145,
453,
424,
465
],
"score": 1.0,
"content": "(b) Did you mention the license of the assets? [Yes] See Appendix E",
"type": "text"
}
],
"index": 30,
"is_list_start_line": true,
"is_list_end_line": true
},
{
"bbox": [
145,
465,
505,
479
],
"spans": [
{
"bbox": [
145,
465,
505,
479
],
"score": 1.0,
"content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]",
"type": "text"
}
],
"index": 31,
"is_list_start_line": true
},
{
"bbox": [
161,
475,
368,
489
],
"spans": [
{
"bbox": [
161,
475,
368,
489
],
"score": 1.0,
"content": "We include our code in the supplemental materials.",
"type": "text"
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
145,
489,
505,
502
],
"spans": [
{
"bbox": [
145,
489,
505,
502
],
"score": 1.0,
"content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
"type": "text"
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
162,
501,
506,
512
],
"spans": [
{
"bbox": [
162,
501,
506,
512
],
"score": 1.0,
"content": "using/curating? [No] All of the datasets used in the experiments are publicly available.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
145,
513,
506,
526
],
"spans": [
{
"bbox": [
145,
513,
506,
526
],
"score": 1.0,
"content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
"type": "text"
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
162,
523,
505,
536
],
"spans": [
{
"bbox": [
162,
523,
505,
536
],
"score": 1.0,
"content": "information or offensive content? [Yes] We mentioned the human privacy issues of the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
162,
534,
297,
547
],
"spans": [
{
"bbox": [
162,
534,
297,
547
],
"score": 1.0,
"content": "ImageNet dataset in Appendix E.",
"type": "text"
}
],
"index": 37,
"is_list_end_line": true
}
],
"index": 33,
"bbox_fs": [
145,
440,
506,
547
]
},
{
"type": "text",
"bbox": [
131,
550,
432,
561
],
"lines": [
{
"bbox": [
128,
548,
433,
564
],
"spans": [
{
"bbox": [
128,
548,
433,
564
],
"score": 1.0,
"content": "5. If you used crowdsourcing or conducted research with human subjects...",
"type": "text"
}
],
"index": 38
}
],
"index": 38,
"bbox_fs": [
128,
548,
433,
564
]
},
{
"type": "list",
"bbox": [
146,
564,
505,
635
],
"lines": [
{
"bbox": [
145,
564,
506,
577
],
"spans": [
{
"bbox": [
145,
564,
506,
577
],
"score": 1.0,
"content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
"type": "text"
}
],
"index": 39,
"is_list_start_line": true
},
{
"bbox": [
162,
576,
237,
588
],
"spans": [
{
"bbox": [
162,
576,
237,
588
],
"score": 1.0,
"content": "applicable? [N/A]",
"type": "text"
}
],
"index": 40,
"is_list_end_line": true
},
{
"bbox": [
145,
588,
505,
601
],
"spans": [
{
"bbox": [
145,
588,
505,
601
],
"score": 1.0,
"content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
"type": "text"
}
],
"index": 41,
"is_list_start_line": true
},
{
"bbox": [
161,
599,
342,
612
],
"spans": [
{
"bbox": [
161,
599,
342,
612
],
"score": 1.0,
"content": "Board (IRB) approvals, if applicable? [N/A]",
"type": "text"
}
],
"index": 42,
"is_list_end_line": true
},
{
"bbox": [
147,
612,
505,
625
],
"spans": [
{
"bbox": [
147,
612,
505,
625
],
"score": 1.0,
"content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
"type": "text"
}
],
"index": 43,
"is_list_start_line": true
},
{
"bbox": [
162,
624,
333,
636
],
"spans": [
{
"bbox": [
162,
624,
333,
636
],
"score": 1.0,
"content": "spent on participant compensation? [N/A]",
"type": "text"
}
],
"index": 44,
"is_list_end_line": true
}
],
"index": 41.5,
"bbox_fs": [
145,
564,
506,
636
]
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.2.2"
}