Research map / Image generation / Fast diffusion sampling
Scale Space Diffusion
Soumik Mukhopadhyay, Prateksha Udhayanan, Abhinav Shrivastava
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 |
|---|---|---|---|
| use Flexi-UNet instead of full UNet (2 levels) (ablation) | SSD · CelebA-64 | FID: 2.33 -> 2.26 | Improved |
| use Flexi-UNet instead of full UNet (4 levels) (ablation) | SSD · CelebA-64 | FID: 4.90 -> 4.87 | Improved |
| use Flexi-UNet instead of full UNet (2 levels) (ablation) | SSD · CelebA-64 | Inference time (s): 16.19 -> 15.38 | Improved |
| add SSD formulation with ConvexDecay0.5 to Flexi-UNet (2 levels) (ablation) | Flexi-UNet · CelebA-64 | FID: 2.44 -> 2.26 | Improved |
| add SSD formulation with ConvexDecay0.5 to Flexi-UNet (4 levels) (ablation) | Flexi-UNet · CelebA-64 | FID: 5.79 -> 4.87 | Improved |
| increase SSD levels from 2 to 4 (ablation) | SSD · CelebA-64 | FID: 2.14 -> 4.28 | Hurt |
| use SSD (2L) instead of DDPM-eps | DDPM · CelebA-64 | Training time (hours): 70.30 -> 62.63 | Improved |
| use SSD (6L) instead of DDPM-eps | DDPM · CelebA-256 | Training time (hours): 87.31 -> 42.88 | Improved |
| use SSD (6L) instead of DDPM-eps | DDPM · CelebA-256 | FID: 5.52 -> 13.50 | Hurt |
| reduce sampling steps from 1000 to 250 (SSD vs DDPM-x0) (ablation) | SSD · CelebA-64 | FID: DDPM-x0 2.98 -> 14.00 vs SSD 2.14 -> 2.87 | Improved |
| replace 2x downsampling with gradual 1-pixel downsizing (w/o Flexi-UNet) (ablation) | SSD · CelebA-32 | FID: 5.55 -> 4.10 | Improved |
| use Lanczos non-isotropic sampling instead of torch.randn (ablation) | SSD (2L) · CelebA-64 | Inference time (s): 15.35 -> 15.38 | No clear effect |
Closest papers by meaning
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.90)
- Diffusion Models Without Attention (CVPR 2024 · similarity 0.89)
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.89)
- Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution (NeurIPS 2026 · similarity 0.89)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.89)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.88)
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.
