-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
437 lines (419 loc) · 16.5 KB
/
index.html
File metadata and controls
437 lines (419 loc) · 16.5 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta
name="description"
content="PRDP achieves stable black-box reward finetuning for diffusion models for the first time on large-scale prompt datasets with over 100K prompts, leading to superior generation quality on complex, unseen prompts."
/>
<meta
name="keywords"
content="PRDP, Proximal Reward Difference Prediction, diffusion models, Stable Diffusion, alignment, human preference, reward finetuning, large scale, reinforcement learning, RLHF, DDPO, DPOK, DPO, PPO"
/>
<title>PRDP</title>
<link
rel="stylesheet"
href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
/>
<link rel="stylesheet" href="./static/css/bulma.min.css" />
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css" />
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css" />
<link
rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css"
/>
<link rel="stylesheet" href="./static/css/index.css" />
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-3 publication-title">
PRDP: Proximal Reward Difference Prediction<br />
for Large-Scale Reward Finetuning of Diffusion Models
</h1>
<div class="is-size-4 publication-authors">
<span class="author-block">
<a
href="https://scholar.google.com/citations?hl=en&user=F-V72fUAAAAJ&view_op=list_works&sortby=pubdate"
>Fei Deng</a
><sup>1,2</sup>,
</span>
<span class="author-block">
<a href="https://research.google/people/qifei-wang/"
>Qifei Wang</a
><sup>1</sup>,
</span>
<span class="author-block">
<a href="http://www.weiwei.one/">Wei Wei</a><sup>1,3</sup>,
</span>
<span class="author-block">
<a href="https://research.google/people/matthias-grundmann/"
>Matthias Grundmann</a
><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://tbhou.github.io/">Tingbo Hou</a
><sup>1,4</sup>
</span>
</div>
<div class="is-size-4 publication-authors">
<span class="author-block"><sup>1</sup>Google,</span>
<span class="author-block"
><sup>2</sup>Rutgers University,
</span>
<span class="author-block"><sup>3</sup>Accenture,</span>
<span class="author-block"><sup>4</sup>Meta</span>
</div>
<div class="is-size-4 publication-authors">
<span class="author-block">CVPR 2024</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a
href="https://arxiv.org/abs/2402.08714"
class="external-link button is-normal is-rounded is-dark"
target="_blank"
rel="noreferrer"
>
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<span class="link-block">
<a
href="https://www.youtube.com/watch?v=nI47O9ccvaw"
class="external-link button is-normal is-rounded is-dark"
target="_blank"
rel="noreferrer"
>
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span>
<span class="link-block">
<a
href="./static/images/slides.pdf"
class="external-link button is-normal is-rounded is-dark"
target="_blank"
rel="noreferrer"
>
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Slides</span>
</a>
</span>
<span class="link-block">
<a
href="./static/images/poster.png"
class="external-link button is-normal is-rounded is-dark"
target="_blank"
rel="noreferrer"
>
<span class="icon">
<i class="fas fa-image"></i>
</span>
<span>Poster</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<div id="teaser-carousel" class="carousel results-carousel">
<div class="item">
<video
poster=""
id="teaser-1"
autoplay
muted
loop
playsinline
height="100%"
>
<source src="./static/videos/teaser_1.mp4" type="video/mp4" />
</video>
</div>
<div class="item">
<video
poster=""
id="teaser-2"
autoplay
muted
loop
playsinline
height="100%"
>
<source src="./static/videos/teaser_2.mp4" type="video/mp4" />
</video>
</div>
</div>
<div
class="content is-size-5 has-text-justified has-text-weight-semibold has-text-black"
>
<p>
<br />
PRDP achieves stable black-box reward finetuning for diffusion
models for the first time on large-scale prompt datasets with over
100K prompts, leading to superior generation quality on complex,
unseen prompts.
</p>
</div>
</div>
</div>
</section>
<hr />
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h2 class="title is-3">Abstract</h2>
<div class="content is-size-5 has-text-justified">
<p>
Reward finetuning has emerged as a promising approach to
aligning foundation models with downstream objectives.
Remarkable success has been achieved in the language domain by
using reinforcement learning (RL) to maximize rewards that
reflect human preference. However, in the vision domain,
existing RL-based reward finetuning methods are limited by their
instability in large-scale training, rendering them incapable of
generalizing to complex, unseen prompts. In this paper, we
propose Proximal Reward Difference Prediction (PRDP), enabling
stable black-box reward finetuning for diffusion models for the
first time on large-scale prompt datasets with over 100K
prompts. Our key innovation is the Reward Difference Prediction
(RDP) objective that has the same optimal solution as the RL
objective while enjoying better training stability.
Specifically, the RDP objective is a supervised regression
objective that tasks the diffusion model with predicting the
reward difference of generated image pairs from their denoising
trajectories. We theoretically prove that the diffusion model
that obtains perfect reward difference prediction is exactly the
maximizer of the RL objective. We further develop an online
algorithm with proximal updates to stably optimize the RDP
objective. In experiments, we demonstrate that PRDP can match
the reward maximization ability of well-established RL-based
methods in small-scale training. Furthermore, through
large-scale training on text prompts from the Human Preference
Dataset v2 and the Pick-a-Pic v1 dataset, PRDP achieves superior
generation quality on a diverse set of complex, unseen prompts
whereas RL-based methods completely fail.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<hr />
<section class="section">
<div class="container is-max-desktop">
<!-- Method. -->
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h2 class="title is-3">Method</h2>
<img src="./static/images/prdp.png" />
<div class="content is-size-5 has-text-justified">
<p>
<br />
PRDP mitigates the instability of policy gradient methods by
converting the RLHF objective to an equivalent supervised
regression objective. Specifically, given a text prompt, PRDP
samples two images, and tasks the diffusion model with
predicting the reward difference of these two images from their
denoising trajectories. The diffusion model is updated by
stochastic gradient descent on the MSE loss that measures the
prediction error. We prove that the MSE loss and the RLHF
objective have the same optimal solution.
</p>
</div>
</div>
</div>
<!--/ Method. -->
</div>
</section>
<hr />
<section class="section">
<div class="container is-max-desktop">
<!-- Results. -->
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h2 class="title is-3">Generation Samples on Unseen Prompts</h2>
</div>
</div>
<div class="hero-body">
<div
id="pick-a-pic-carousel"
class="autoplay-carousel results-carousel"
>
<div class="item">
<img src="./static/images/pick_a_pic/1.png" />
</div>
<div class="item">
<img src="./static/images/pick_a_pic/2.png" />
</div>
<div class="item">
<img src="./static/images/pick_a_pic/3.png" />
</div>
<div class="item">
<img src="./static/images/pick_a_pic/4.png" />
</div>
<div class="item">
<img src="./static/images/pick_a_pic/5.png" />
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h3 class="title is-4">Pick-a-Pic v1 Test</h3>
</div>
</div>
</div>
<div class="hero-body">
<div
id="animation-carousel"
class="autoplay-carousel results-carousel"
>
<div class="item">
<img src="./static/images/animation/1.png" />
</div>
<div class="item">
<img src="./static/images/animation/2.png" />
</div>
<div class="item">
<img src="./static/images/animation/3.png" />
</div>
<div class="item">
<img src="./static/images/animation/4.png" />
</div>
<div class="item">
<img src="./static/images/animation/5.png" />
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h3 class="title is-4">Human Preference Dataset v2 Animation</h3>
</div>
</div>
</div>
<div class="hero-body">
<div
id="concept-art-carousel"
class="autoplay-carousel results-carousel"
>
<div class="item">
<img src="./static/images/concept_art/1.png" />
</div>
<div class="item">
<img src="./static/images/concept_art/2.png" />
</div>
<div class="item">
<img src="./static/images/concept_art/3.png" />
</div>
<div class="item">
<img src="./static/images/concept_art/4.png" />
</div>
<div class="item">
<img src="./static/images/concept_art/5.png" />
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h3 class="title is-4">
Human Preference Dataset v2 Concept Art
</h3>
</div>
</div>
</div>
<div class="hero-body">
<div
id="painting-carousel"
class="autoplay-carousel results-carousel"
>
<div class="item">
<img src="./static/images/painting/1.png" />
</div>
<div class="item">
<img src="./static/images/painting/2.png" />
</div>
<div class="item">
<img src="./static/images/painting/3.png" />
</div>
<div class="item">
<img src="./static/images/painting/4.png" />
</div>
<div class="item">
<img src="./static/images/painting/5.png" />
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h3 class="title is-4">Human Preference Dataset v2 Painting</h3>
</div>
</div>
</div>
<div class="hero-body">
<div id="photo-carousel" class="autoplay-carousel results-carousel">
<div class="item">
<img src="./static/images/photo/1.png" />
</div>
<div class="item">
<img src="./static/images/photo/2.png" />
</div>
<div class="item">
<img src="./static/images/photo/3.png" />
</div>
<div class="item">
<img src="./static/images/photo/4.png" />
</div>
<div class="item">
<img src="./static/images/photo/5.png" />
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-max-desktop">
<h3 class="title is-4">Human Preference Dataset v2 Photo</h3>
</div>
</div>
</div>
<!--/ Results. -->
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<p>
This website is based on the
<a href="https://github.com/nerfies/nerfies.github.io">Nerfies</a>
template, licensed under a
<a
rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/"
>Creative Commons Attribution-ShareAlike 4.0 International
License</a
>.
</p>
</div>
</div>
</footer>
</body>
</html>