Research map / Image generation / Fast diffusion sampling
Shortcut Diffusion Training with Cumulative Consistency Loss: An Optimal Control View
Paribesh Regmi, Sandesh Ghimire, Rui Li
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 cumulative consistency loss (CSL) over the shortcut loss | Shortcut model · CelebA-256 one-step | FID-50K: 20.46 -> 18.37 | Improved |
| add cumulative consistency loss (CSL) | Shortcut model · CIFAR10 one-step | FID-50K: 19.80 -> 17.76 | Improved |
| add cumulative consistency loss (CSL) | Shortcut model · CIFAR10 two-step | FID-50K: 11.79 -> 9.24 | Improved |
| use unrolled (non-cumulative) loss USL instead of SL (ablation) | Shortcut model · CIFAR10 four-step | FID-50K: 9.15 -> 9.35 | Hurt |
| use USL instead of SL (ablation) | Shortcut model · CelebA-256 one-step | FID-50K: 20.46 -> 19.41 | Improved |
| add CSL over USL (ablation) | ST-USL · CIFAR10 two-step | FID-50K: 11.65 -> 9.24 | Improved |
| add CSL at 128-step base model | Shortcut model · CelebA-256 128-step | FID-50K: 7.83 -> 7.88 | No clear effect |
| increase number of CSL terms R from 1 to 4 (ablation) | ST-CSL · CIFAR10 one-step | FID-50K: 16.43 -> 13.10 | Improved |
| increase R from 1 to 2 (ablation) | ST-CSL · CIFAR10 two-step | FID-50K: 10.54 -> 7.94 | Improved |
| increase proportion of bootstrap targets (B:K 4:1 to 1:1) with CSL (ablation) | ST-CSL · CIFAR10 one-step | FID-50K: 17.76 -> 13.96 | Improved |
| add CSL at 1:1 ratio | Shortcut model · CIFAR10 128-step | FID-50K: 6.56 -> 6.67 | Hurt |
| add CSL on ImageNet-256 | Shortcut model (4:1) · ImageNet-256 one-step | FID: 64.12 -> 45.84 | Improved |
Closest papers by meaning
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.91)
- Truncated Consistency Models (ICLR 2025 · similarity 0.90)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.89)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.89)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.89)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 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.
