David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training

Weijian Luo, Colin Zhang, Debing Zhang, Zhengyang Geng
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:41520-41539, 2025.

Abstract

We propose Diff-Instruct(DI), a data-efficient post-training approach to one-step text-to-image generative models to improve its human preferences without requiring image data. Our method frames alignment as online reinforcement learning from human feedback (RLHF), which optimizes a human reward function while regularizing the generator to stay close to a reference diffusion process. Unlike traditional RLHF approaches, which rely on the KL divergence for regularization, we introduce a novel score-based divergence regularization that substantially improves performance. Although such a score-based RLHF objective seems intractable when optimizing, we derive a strictly equivalent tractable loss function in theory that can efficiently compute its gradient for optimizations. Building upon this framework, we train DI-SDXL-1step, a 1-step text-to-image model based on Stable Diffusion-XL (2.6B parameters), capable of generating 1024x1024 resolution images in a single step. The 2.6B DI-SDXL-1step model outperforms the 12B FLUX-dev model in ImageReward, PickScore, and CLIP score on the Parti prompts benchmark while using only 1.88% of the inference time. This result strongly supports the thought that with proper post-training, the small one-step model is capable of beating huge multi-step models. We will open-source our industry-ready model to the community.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-luo25s, title = {David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training}, author = {Luo, Weijian and Zhang, Colin and Zhang, Debing and Geng, Zhengyang}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {41520--41539}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/luo25s/luo25s.pdf}, url = {https://proceedings.mlr.press/v267/luo25s.html}, abstract = {We propose Diff-Instruct(DI), a data-efficient post-training approach to one-step text-to-image generative models to improve its human preferences without requiring image data. Our method frames alignment as online reinforcement learning from human feedback (RLHF), which optimizes a human reward function while regularizing the generator to stay close to a reference diffusion process. Unlike traditional RLHF approaches, which rely on the KL divergence for regularization, we introduce a novel score-based divergence regularization that substantially improves performance. Although such a score-based RLHF objective seems intractable when optimizing, we derive a strictly equivalent tractable loss function in theory that can efficiently compute its gradient for optimizations. Building upon this framework, we train DI-SDXL-1step, a 1-step text-to-image model based on Stable Diffusion-XL (2.6B parameters), capable of generating 1024x1024 resolution images in a single step. The 2.6B DI-SDXL-1step model outperforms the 12B FLUX-dev model in ImageReward, PickScore, and CLIP score on the Parti prompts benchmark while using only 1.88% of the inference time. This result strongly supports the thought that with proper post-training, the small one-step model is capable of beating huge multi-step models. We will open-source our industry-ready model to the community.} }
Endnote
%0 Conference Paper %T David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training %A Weijian Luo %A Colin Zhang %A Debing Zhang %A Zhengyang Geng %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-luo25s %I PMLR %P 41520--41539 %U https://proceedings.mlr.press/v267/luo25s.html %V 267 %X We propose Diff-Instruct(DI), a data-efficient post-training approach to one-step text-to-image generative models to improve its human preferences without requiring image data. Our method frames alignment as online reinforcement learning from human feedback (RLHF), which optimizes a human reward function while regularizing the generator to stay close to a reference diffusion process. Unlike traditional RLHF approaches, which rely on the KL divergence for regularization, we introduce a novel score-based divergence regularization that substantially improves performance. Although such a score-based RLHF objective seems intractable when optimizing, we derive a strictly equivalent tractable loss function in theory that can efficiently compute its gradient for optimizations. Building upon this framework, we train DI-SDXL-1step, a 1-step text-to-image model based on Stable Diffusion-XL (2.6B parameters), capable of generating 1024x1024 resolution images in a single step. The 2.6B DI-SDXL-1step model outperforms the 12B FLUX-dev model in ImageReward, PickScore, and CLIP score on the Parti prompts benchmark while using only 1.88% of the inference time. This result strongly supports the thought that with proper post-training, the small one-step model is capable of beating huge multi-step models. We will open-source our industry-ready model to the community.
APA
Luo, W., Zhang, C., Zhang, D. & Geng, Z.. (2025). David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:41520-41539 Available from https://proceedings.mlr.press/v267/luo25s.html.

Related Material