Skip to content

Commit e9bd01b

Browse files
authored
Merge pull request #9 from felix-yuxiang/master
Add one blog post by Felix
2 parents 1382125 + 8e41516 commit e9bd01b

File tree

8 files changed

+569
-0
lines changed

8 files changed

+569
-0
lines changed

_posts/2025-08-18-diff-distill.md

Lines changed: 386 additions & 0 deletions
Large diffs are not rendered by default.
Lines changed: 183 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,183 @@
1+
@misc{lipman_flow_2023,
2+
title = {Flow Matching for Generative Modeling},
3+
url = {http://arxiv.org/abs/2210.02747},
4+
doi = {10.48550/arXiv.2210.02747},
5+
abstract = {We introduce a new paradigm for generative modeling built on Continuous Normalizing Flows ({CNFs}), allowing us to train {CNFs} at unprecedented scale. Specifically, we present the notion of Flow Matching ({FM}), a simulation-free approach for training {CNFs} based on regressing vector fields of fixed conditional probability paths. Flow Matching is compatible with a general family of Gaussian probability paths for transforming between noise and data samples -- which subsumes existing diffusion paths as specific instances. Interestingly, we find that employing {FM} with diffusion paths results in a more robust and stable alternative for training diffusion models. Furthermore, Flow Matching opens the door to training {CNFs} with other, non-diffusion probability paths. An instance of particular interest is using Optimal Transport ({OT}) displacement interpolation to define the conditional probability paths. These paths are more efficient than diffusion paths, provide faster training and sampling, and result in better generalization. Training {CNFs} using Flow Matching on {ImageNet} leads to consistently better performance than alternative diffusion-based methods in terms of both likelihood and sample quality, and allows fast and reliable sample generation using off-the-shelf numerical {ODE} solvers.},
6+
number = {{arXiv}:2210.02747},
7+
publisher = {{arXiv}},
8+
author = {Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matt},
9+
urldate = {2024-07-05},
10+
date = {2023-02-08},
11+
eprinttype = {arxiv},
12+
eprint = {2210.02747 [cs, stat]}
13+
}
14+
15+
@article{albergo2023stochastic,
16+
title={Stochastic interpolants: A unifying framework for flows and diffusions},
17+
author={Albergo, Michael S and Boffi, Nicholas M and Vanden-Eijnden, Eric},
18+
journal={arXiv preprint arXiv:2303.08797},
19+
year={2023}
20+
}
21+
22+
@article{tong2023improving,
23+
title={Improving and generalizing flow-based generative models with minibatch optimal transport},
24+
author={Tong, Alexander and Fatras, Kilian and Malkin, Nikolay and Huguet, Guillaume and Zhang, Yanlei and Rector-Brooks, Jarrid and Wolf, Guy and Bengio, Yoshua},
25+
journal={arXiv preprint arXiv:2302.00482},
26+
year={2023}
27+
}
28+
29+
@article{liu2022flow,
30+
title={Flow straight and fast: Learning to generate and transfer data with rectified flow},
31+
author={Liu, Xingchao and Gong, Chengyue and Liu, Qiang},
32+
journal={arXiv preprint arXiv:2209.03003},
33+
year={2022}
34+
}
35+
36+
@article{hu2021lora,
37+
title={Lora: Low-rank adaptation of large language models. arXiv 2021},
38+
author={Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu},
39+
journal={arXiv preprint arXiv:2106.09685},
40+
volume={10},
41+
year={2021}
42+
}
43+
44+
@article{micikevicius2017mixed,
45+
title={Mixed precision training},
46+
author={Micikevicius, Paulius and Narang, Sharan and Alben, Jonah and Diamos, Gregory and Elsen, Erich and Garcia, David and Ginsburg, Boris and Houston, Michael and Kuchaiev, Oleksii and Venkatesh, Ganesh and others},
47+
journal={arXiv preprint arXiv:1710.03740},
48+
year={2017}
49+
}
50+
51+
@inproceedings{fu2025moflowonestep,
52+
author = {Fu, Yuxiang and Yan, Qi and Wang, Lele and Li, Ke and Liao, Renjie},
53+
title = {MoFlow: One-Step Flow Matching for Human Trajectory Forecasting via Implicit Maximum Likelihood Estimation based Distillation},
54+
journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
55+
year = {2025},
56+
}
57+
58+
@misc{lipman2024flowmatchingguidecode,
59+
title={Flow Matching Guide and Code},
60+
author={Yaron Lipman and Marton Havasi and Peter Holderrieth and Neta Shaul and Matt Le and Brian Karrer and Ricky T. Q. Chen and David Lopez-Paz and Heli Ben-Hamu and Itai Gat},
61+
year={2024},
62+
eprint={2412.06264},
63+
archivePrefix={arXiv},
64+
primaryClass={cs.LG},
65+
url={https://arxiv.org/abs/2412.06264},
66+
}
67+
68+
@article{boffi2025build,
69+
title={How to build a consistency model: Learning flow maps via self-distillation},
70+
author={Boffi, Nicholas M and Albergo, Michael S and Vanden-Eijnden, Eric},
71+
journal={arXiv preprint arXiv:2505.18825},
72+
year={2025}
73+
}
74+
75+
@article{geng2025mean,
76+
title={Mean flows for one-step generative modeling},
77+
author={Geng, Zhengyang and Deng, Mingyang and Bai, Xingjian and Kolter, J Zico and He, Kaiming},
78+
journal={arXiv preprint arXiv:2505.13447},
79+
year={2025}
80+
}
81+
82+
@article{peng2025flow,
83+
title={Flow-Anchored Consistency Models},
84+
author={Peng, Yansong and Zhu, Kai and Liu, Yu and Wu, Pingyu and Li, Hebei and Sun, Xiaoyan and Wu, Feng},
85+
journal={arXiv preprint arXiv:2507.03738},
86+
year={2025}
87+
}
88+
89+
@article{guo2025splitmeanflow,
90+
title={SplitMeanFlow: Interval Splitting Consistency in Few-Step Generative Modeling},
91+
author={Guo, Yi and Wang, Wei and Yuan, Zhihang and Cao, Rong and Chen, Kuan and Chen, Zhengyang and Huo, Yuanyuan and Zhang, Yang and Wang, Yuping and Liu, Shouda and others},
92+
journal={arXiv preprint arXiv:2507.16884},
93+
year={2025}
94+
}
95+
96+
@article{ho2020denoising,
97+
title={Denoising diffusion probabilistic models},
98+
author={Ho, Jonathan and Jain, Ajay and Abbeel, Pieter},
99+
journal={Advances in neural information processing systems},
100+
volume={33},
101+
pages={6840--6851},
102+
year={2020}
103+
}
104+
105+
@article{song2020score,
106+
title={Score-based generative modeling through stochastic differential equations},
107+
author={Song, Yang and Sohl-Dickstein, Jascha and Kingma, Diederik P and Kumar, Abhishek and Ermon, Stefano and Poole, Ben},
108+
journal={arXiv preprint arXiv:2011.13456},
109+
year={2020}
110+
}
111+
112+
@article{lu2024simplifying,
113+
title={Simplifying, stabilizing and scaling continuous-time consistency models},
114+
author={Lu, Cheng and Song, Yang},
115+
journal={arXiv preprint arXiv:2410.11081},
116+
year={2024}
117+
}
118+
119+
@article{kim2023consistency,
120+
title={Consistency trajectory models: Learning probability flow ode trajectory of diffusion},
121+
author={Kim, Dongjun and Lai, Chieh-Hsin and Liao, Wei-Hsiang and Murata, Naoki and Takida, Yuhta and Uesaka, Toshimitsu and He, Yutong and Mitsufuji, Yuki and Ermon, Stefano},
122+
journal={arXiv preprint arXiv:2310.02279},
123+
year={2023}
124+
}
125+
126+
@article{sabour2025align,
127+
title={Align Your Flow: Scaling Continuous-Time Flow Map Distillation},
128+
author={Sabour, Amirmojtaba and Fidler, Sanja and Kreis, Karsten},
129+
journal={arXiv preprint arXiv:2506.14603},
130+
year={2025}
131+
}
132+
133+
@article{frans2024one,
134+
title={One step diffusion via shortcut models},
135+
author={Frans, Kevin and Hafner, Danijar and Levine, Sergey and Abbeel, Pieter},
136+
journal={arXiv preprint arXiv:2410.12557},
137+
year={2024}
138+
}
139+
140+
@article{zhou2025inductive,
141+
title={Inductive moment matching},
142+
author={Zhou, Linqi and Ermon, Stefano and Song, Jiaming},
143+
journal={arXiv preprint arXiv:2503.07565},
144+
year={2025}
145+
}
146+
147+
@article{yin2024improved,
148+
title={Improved distribution matching distillation for fast image synthesis},
149+
author={Yin, Tianwei and Gharbi, Micha{\"e}l and Park, Taesung and Zhang, Richard and Shechtman, Eli and Durand, Fredo and Freeman, Bill},
150+
journal={Advances in neural information processing systems},
151+
volume={37},
152+
pages={47455--47487},
153+
year={2024}
154+
}
155+
156+
@article{song2020denoising,
157+
title={Denoising diffusion implicit models},
158+
author={Song, Jiaming and Meng, Chenlin and Ermon, Stefano},
159+
journal={arXiv preprint arXiv:2010.02502},
160+
year={2020}
161+
}
162+
163+
@article{wang2025uni,
164+
title={Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction},
165+
author={Wang, Yifei and Bai, Weimin and Zhang, Colin and Zhang, Debing and Luo, Weijian and Sun, He},
166+
journal={arXiv preprint arXiv:2505.20755},
167+
year={2025}
168+
}
169+
170+
@inproceedings{zhou2024score,
171+
title={Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation},
172+
author={Mingyuan Zhou and Huangjie Zheng and Zhendong Wang and Mingzhang Yin and Hai Huang},
173+
booktitle={International Conference on Machine Learning},
174+
url={https://arxiv.org/abs/2404.04057},
175+
year={2024}
176+
}
177+
178+
@article{xu2025one,
179+
title={One-step Diffusion Models with $ f $-Divergence Distribution Matching},
180+
author={Xu, Yilun and Nie, Weili and Vahdat, Arash},
181+
journal={arXiv preprint arXiv:2502.15681},
182+
year={2025}
183+
}
203 KB
Loading

blog/2025/diff-distill/IMM.png

454 KB
Loading
1.35 MB
Binary file not shown.
768 KB
Loading
1.04 MB
Loading
181 KB
Loading

0 commit comments

Comments
 (0)