Research map / Image generation / Fast diffusion sampling
UDPM: Upsampling Diffusion Probabilistic Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UDPM, a diffusion model whose forward process downsamples latents (blur + stride-2 subsampling) plus noise, so generation takes 3 cheap upsampling denoising steps with FID 6.86 on CIFAR10.
Problem
Standard DDPMs need many diffusion steps (hundreds to thousands) for good samples, and their latent spaces are hard to interpret; existing fast samplers still need many network evaluations or are distillation-based.
Approach
Generalize the diffusion forward process so each step both downsamples the latent (blur filter followed by stride-gamma subsampling) and adds Gaussian noise; the reverse process then denoises and upsamples, with a network trained as a super-resolution-style predictor of H^{l-1} x0, yielding very few steps and much smaller latents.
Key results
- UDPM reaches FID 6.86 on CIFAR10 with 3 steps costing about 0.3 of one DDPM/EDM step (13.35 vs 40.62 GFLOPs at 64x64).
- On FFHQ and AFHQv2 (64x64), UDPM gets FID 7.41 and 7.10 with 3 steps, while EDM at 5 steps degrades to FID 344.76 and 266.02.
- UDPM samples 64x64 images at 765.21 FPS versus 255 FPS for a single denoising step of a similar network on an RTX A6000.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UDPM instead of CT for one-step-class CIFAR10 generation | CIFAR10 generation · CIFAR10 | FID: 8.70 -> 6.86 | Helped |
| Use UDPM instead of TDPM for one-step-class CIFAR10 generation | CIFAR10 generation · CIFAR10 | FID: 8.91 -> 6.86 | Helped |
| Use UDPM instead of DDGAN | CIFAR10 generation · CIFAR10 | FID: 4.08 -> 6.86 | Hurt |
| Use UDPM instead of 5-step EDM | CIFAR10 generation · CIFAR10 | FID: 35.54 -> 6.86 | Helped |
| Use UDPM instead of 5-step EDM | FFHQ generation · FFHQ | FID: 344.763 -> 7.41 | Helped |
| Use UDPM instead of 5-step EDM | AFHQv2 generation · AFHQv2 | FID: 266.024 -> 7.10 | Helped |
| Replace a single denoising step with three UDPM steps | 64×64 image sampling · image generation | GFLOPS: 40.62 -> 13.35 | Helped |
| Add perceptual and adversarial losses to the simple training loss | UDPM · image generation | image sharpness | Helped |
| Set λ_fid and λ_per to 0 for l=3 (ablation) | UDPM training · image generation | generation quality | Helped |
Surprising findings
- Reducing EDM to 5 steps raises FID from about 2 to over 260 on FFHQ and AFHQv2, whereas UDPM with 3 steps stays near 7.
Limitations
- Evaluated only on small datasets (CIFAR10, FFHQ, AFHQv2) at low resolution due to compute limits.
- Latent space is less interpretable than GANs, per the authors.
- Noise scheduler is hand-set (six values) rather than analytically derived or tuned.
Experimental setup
Datasets: CIFAR10 FFHQ AFHQv2
Benchmarks: CIFAR10 FFHQ AFHQv2
Metrics: FID GFLOPs FPS
Baselines: DDIM DPM-Solver EDM GENIE DEIS GGDM DDGAN TDPM CT
Builds on: DDPM EDM DDGAN NCSN
Closest papers by meaning
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.90)
- Denoising Diffusion Step-aware Models (ICLR 2024 · similarity 0.90)
- Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models (AAAI 2024 · similarity 0.90)
- An Edit Friendly DDPM Noise Space: Inversion and Manipulations (CVPR 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.