DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917227906727936 |
|---|---|
| author | Chen, Wenlin Zhang, Mingtian He, Jiajun Ou, Zijing Hernández-Lobato, José Miguel Schölkopf, Bernhard Barber, David |
| author_facet | Chen, Wenlin Zhang, Mingtian He, Jiajun Ou, Zijing Hernández-Lobato, José Miguel Schölkopf, Bernhard Barber, David |
| contents | Score-based distillation methods (e.g., variational score distillation) train one-step diffusion models by first pre-training a teacher score model and then distilling it into a one-step student model. However, the gradient estimator in the distillation stage usually suffers from two sources of bias: (1) biased teacher supervision due to score estimation error incurred during pre-training, and (2) the student model's score estimation error during distillation. These biases can degrade the quality of the resulting one-step diffusion model. To address this, we propose DiffRatio, a new framework for training one-step diffusion models: instead of estimating the teacher and student scores independently and then taking their difference, we directly estimate the score difference as the gradient of a learned log density ratio between the student and data distributions across diffusion time steps. This approach greatly simplifies the training pipeline, significantly reduces gradient estimation bias, and improves one-step generation quality. Additionally, it also reduces auxiliary network size by using a lightweight density-ratio network instead of two full score networks, which improves computational and memory efficiency. DiffRatio achieves competitive one-step generation results on CIFAR-10 and ImageNet (64x64 and 512x512), outperforming most teacher-supervised distillation methods. Moreover, the learned density ratio naturally serves as a verifier, enabling a principled inference-time parallel scaling scheme that further improves the generation quality without external rewards or additional sequential computation. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_08005 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision Chen, Wenlin Zhang, Mingtian He, Jiajun Ou, Zijing Hernández-Lobato, José Miguel Schölkopf, Bernhard Barber, David Machine Learning Computer Vision and Pattern Recognition Score-based distillation methods (e.g., variational score distillation) train one-step diffusion models by first pre-training a teacher score model and then distilling it into a one-step student model. However, the gradient estimator in the distillation stage usually suffers from two sources of bias: (1) biased teacher supervision due to score estimation error incurred during pre-training, and (2) the student model's score estimation error during distillation. These biases can degrade the quality of the resulting one-step diffusion model. To address this, we propose DiffRatio, a new framework for training one-step diffusion models: instead of estimating the teacher and student scores independently and then taking their difference, we directly estimate the score difference as the gradient of a learned log density ratio between the student and data distributions across diffusion time steps. This approach greatly simplifies the training pipeline, significantly reduces gradient estimation bias, and improves one-step generation quality. Additionally, it also reduces auxiliary network size by using a lightweight density-ratio network instead of two full score networks, which improves computational and memory efficiency. DiffRatio achieves competitive one-step generation results on CIFAR-10 and ImageNet (64x64 and 512x512), outperforming most teacher-supervised distillation methods. Moreover, the learned density ratio naturally serves as a verifier, enabling a principled inference-time parallel scaling scheme that further improves the generation quality without external rewards or additional sequential computation. |
| title | DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision |
| topic | Machine Learning Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2502.08005 |