Research map / Image generation / Fast diffusion sampling
Contrastive Sampling Chains in Diffusion Models
Junyu Zhang, Daochang Liu, Shichao Zhang, Chang Xu
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 |
|---|---|---|---|
| fine-tune with contrastive sampling chain (C++) | EDM · CIFAR-10 unconditional, 35 NFE | FID: 2.04 -> 1.88 | Improved |
| fine-tune with contrastive sampling chain (C++) | EDM · CIFAR-10 unconditional | NFE to match FID: 35 -> 25 NFE at 2.04 FID | Improved |
| fine-tune with contrastive sampling chain (C++) | LSGM · CIFAR-10 unconditional, 138 NFE | FID: 2.10 -> 1.99 | Improved |
| fine-tune with contrastive sampling chain (C++) | EDM · CIFAR-10 unconditional, 35 NFE | NLL: 2.60 -> 2.55 | Improved |
| fine-tune with contrastive sampling chain (C++) | EDM · CIFAR-10 conditional, 35 NFE | FID: 1.82 -> 1.73 | Improved |
| fine-tune with contrastive sampling chain (C++) | EDM · FFHQ 64x64, 79 NFE | FID: 2.39 -> 2.07 | Improved |
| fine-tune with contrastive sampling chain (C++) | STDDPM · CelebA 64x64, 131 NFE | FID: 1.90 -> 1.73 | Improved |
| fine-tune chain defined by DEIS sampler with C++ | DEIS-tAB3 on SDE VP · CIFAR-10, 10 NFE | FID: 4.17 -> 4.02 | Improved |
| fine-tune chain defined by DPM-Solver-3 with C++ | DPM-Solver-3 on SDE VP · CIFAR-10, 10 NFE | FID: 54.56 -> 43.81 | Improved |
| fine-tune chain defined by DEIS sampler with C++ | DEIS-tAB3 on IDDPM conditional · ImageNet 64x64, 10 NFE | FID: 6.65 -> 6.59 | No clear effect |
| add BPTT to contrastive loss (linear schedule) (ablation) | EDM contrastive fine-tuning · CIFAR-10 | FID: 1.98 -> 1.88 | Improved |
| compute contrastive loss over all chain steps instead of one pair (ablation) | EDM contrastive fine-tuning · CIFAR-10, nonlinear schedule | FID: 1.99 -> 2.39 | Hurt |
Closest papers by meaning
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.90)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.90)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.89)
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.
