Research map / Image generation / Fast diffusion sampling
Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent
Giannis Daras, Yuval Dagan, Alex Dimakis, Constantinos Costis Daskalakis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add consistency regularization to EDM-VP training | EDM-VP · CIFAR-10 conditional | FID (best): 1.82 -> 1.77 | Improved |
| add consistency regularization to EDM-VP training | EDM-VP · CIFAR-10 unconditional | FID (best): 2.01 -> 1.95 | Improved |
| add consistency regularization to EDM-VP training | EDM-VP · AFHQ 64x64 | FID (best): 2.26 -> 2.21 | Improved |
| add consistency regularization | EDM-VP · AFHQ 64x64 | FID at 100k iterations: 2.61 -> 2.30 | Improved |
| add consistency regularization | EDM-VP · AFHQ 64x64 | FID at 70k iterations: 2.41 -> 2.44 | No clear effect |
| add consistency regularization | EDM-VP · CIFAR-10 conditional | FID at 150k iterations: 1.85 -> 1.88 | Hurt |
| finetune pretrained EDM-VP for 5k steps with consistency regularization | EDM-VP · FFHQ 64x64 | FID: 2.68 -> 2.61 | Improved |
| reduce batch size from 512 to 256 to fit memory | EDM-VP · AFHQ 64x64 | FID: 1.96 -> 2.29 | Hurt |
| add consistency regularization with DSM at all t (ablation) | EDM · CIFAR-10 conditional (10k steps) | FID: 5.81 -> 5.45 | Improved |
| drop DSM loss for 20% of t on clean side, keep consistency everywhere (ablation) | CDM · CIFAR-10 conditional (10k steps) | FID: 5.45 -> 6.59 | Hurt |
| stop sampling early at t_threshold instead of sampling through untrained times (ablation) | CDM trained without DSM for some t · CIFAR-10 conditional (10k steps) | FID: 6.59 -> 14.52 | Hurt |
| use 6 sampling steps in consistency loss | CDM · training cost | training time: 1.5x | Hurt |
Closest papers by meaning
- Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion (ICLR 2024 · similarity 0.91)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.91)
- Sampling is as easy as keeping the consistency: convergence guarantee for Consistency Models (ICML 2024 · similarity 0.91)
- Convergence of Consistency Model with Multistep Sampling under General Data Assumptions (ICML 2025 · similarity 0.91)
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.90)
- Truncated Consistency Models (ICLR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
