Research map / Image generation / Fast diffusion sampling
Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Search a flexible GGDM sampler using DDSS | pre-trained DDPM · CIFAR10, 5 steps | FID: 32.66 -> 13.77 | Helped |
| Search a flexible GGDM sampler using DDSS | pre-trained DDPM · ImageNet 64x64, 10 steps | FID: 40.70 -> 37.32 | Helped |
| Search a flexible GGDM sampler using DDSS | pre-trained DDPM · LSUN Church 128x128, 10 steps | FID: 54.39 -> 11.59 | Helped |
| Add learned timestep parameters (ablation) | GGDM+PRED sampler · CIFAR10, 10 steps | FID: 8.617 -> 8.227 | Helped |
| Add learned PRED coefficients (ablation) | GGDM+TIME sampler · CIFAR10, 10 steps | FID: 7.858 -> 8.227 | Hurt |
| Add learned timestep parameters and PRED coefficients (ablation) | GGDM sampler · CIFAR10, 10 steps | FID: 8.154 -> 8.227 | No clear effect |
| Replace the linear KID kernel with the cubic kernel (ablation) | GGDM+PRED+TIME · CIFAR10, 10 steps | FID: 8.227 -> 8.2257 | No clear effect |
| Use pixel-space loss instead of Inception-feature KID (ablation) | DDSS sampler search · unconditional image generation | apparent sample quality | Hurt |
| Use the narrower DDIM sampler family instead of GGDM (ablation) | DDSS · CIFAR10, 10 steps | FID: 8.227 -> 16.29 | Hurt |
| Use VARS instead of GGDM (ablation) | DDSS · CIFAR10, 20 steps | FID: 4.722 -> 7.297 | Hurt |
| Optimize DDIM variance parameters (ablation) | DDSS (DDIM) · CIFAR10, 10 steps | FID: 13.62 -> 16.29 | Hurt |
Closest papers by meaning
- Directly Denoising Diffusion Models (ICML 2024 · similarity 0.90)
- End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler (ICLR 2025 · similarity 0.90)
- Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models (ICML 2025 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.90)
- Learning to Discretize Denoising Diffusion ODEs (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.