Research map / Image generation / Fast diffusion sampling
f-DM: A Multi-stage Diffusion Model via Progressive Signal Transformation
Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Miguel Ángel Bautista, Joshua M. Susskind
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 the interpolation formulation between stages (ablation) | f-DM-DS · FFHQ | FID: 26.5 -> 10.8 | Improved |
| Add noise rescaling rather than using no rescaling (ablation) | f-DM-DS · FFHQ | FID: 14.5 -> 10.8 | Improved |
| Use VP rather than SP noise rescaling (ablation) | f-DM-DS · FFHQ | FID: 12.1 -> 10.8 | Improved |
| Use cosine rather than linear stage scheduling (ablation) | f-DM-DS · FFHQ | FID: 13.5 -> 10.8 | Improved |
| Use cosine rather than linear stage scheduling (ablation) | f-DM-VQVAE · FFHQ | FID: 12.7 -> 13.8 | Hurt |
| Add noise rescaling rather than using no rescaling (ablation) | f-DM-VQVAE · FFHQ | FID: 24.0 -> 12.7 | Improved |
| Use progressive downsampling instead of standard DDPM | f-DM-DS · AFHQ 256×256 | FID: 9.3 -> 6.4 | Improved |
| Use progressive downsampling instead of standard DDPM | f-DM-DS · LSUN-Bed 256×256 | FID: 8.0 -> 6.9 | Improved |
| Use VQVAE latent transformation instead of standard DDPM | f-DM-VQVAE · ImageNet 256×256 | FID: 10.9 -> 6.8 | Improved |
| Use f-DM-DS rather than DDPM with half as many timesteps | f-DM-DS · FFHQ 256×256 | FID: 16.8 -> 10.8 | Improved |
| Use f-DM-DS rather than DDPM | f-DM-DS · FFHQ 256×256 | Speed: ×1.0 -> ×2.1 | Improved |
| Use f-DM-VQGAN rather than f-DM-VQVAE | f-DM · AFHQ 256×256 | FID: 8.9 -> 5.6 | Improved |
Closest papers by meaning
- Neural Diffusion Models (ICML 2024 · similarity 0.91)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.90)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.90)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024 · similarity 0.90)
- Denoising Diffusion Bridge Models (ICLR 2024 · 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.
