Research map / Image generation / Fast diffusion sampling
Wavelet Diffusion Models Are Fast and Scalable Image Generators
Hao Phung, Quan Dao, Anh Tran
This paper tried a common change: Add residual connections
add frequency residual connections · CelebA-HQ 256×256 · FID: 6.25 -> 5.94 Improved
What 546 accepted papers report for this change
- 449 improved
- 35 no clear effect
- 11 mixed
- 51 hurt
- Resolution as a Direction: Vector-Panning Feature Alignment for Cross-Resolution Re-Identification ICML 2026 · add residual connection to VP · MLR-Market-1501 · Rank-1 (%): 91.2 -> 94.1Improved
- MAGIC: Multi-Granularity Language-Informed Image Clustering ICML 2026 · add residual connection to semantic adapters · CIFAR-20 · ACC: 66.1 -> 59.9Hurt
- Domain Adaptive Object Detection via Dynamic Causal Refinement ICML 2026 · add residual connection to SCC · Cityscapes to Foggy Cityscapes (0.02 split) · mAP@50: 64.0 -> 64.6Improved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 |
|---|---|---|---|
| apply wavelet diffusion instead of DDGAN | DDGAN · CIFAR-10 32×32 | Time (s): 0.21 -> 0.08 | Improved |
| apply wavelet diffusion instead of DDGAN | DDGAN · CIFAR-10 32×32 | FID: 3.75 -> 4.01 | Hurt |
| use the wavelet diffusion scheme with the wavelet-embedded generator | DDGAN · STL-10 64×64 | FID: 21.79 -> 12.93 | Improved |
| use the wavelet diffusion scheme with the wavelet-embedded generator | DDGAN · STL-10 64×64 | Time (s): 0.58 -> 0.38 | Improved |
| add the wavelet-embedded generator (ablation) | Wavelet Diffusion · CelebA-HQ 256×256 | FID: 6.55 -> 5.94 | Improved |
| add frequency residual connections (ablation) | Wavelet Diffusion + W-Generator · CelebA-HQ 256×256 | FID: 6.25 -> 5.94 | Improved |
| add frequency-aware upsampling and downsampling blocks (ablation) | Wavelet Diffusion + W-Generator · CelebA-HQ 256×256 | FID: 6.23 -> 5.94 | Improved |
| add frequency-aware upsampling and downsampling blocks (ablation) | Wavelet Diffusion + W-Generator · CelebA-HQ 256×256 | Time (s): 0.61 -> 0.79 | Hurt |
| add frequency bottleneck blocks (ablation) | Wavelet Diffusion + W-Generator · CelebA-HQ 256×256 | FID: 6.18 -> 5.94 | Improved |
| add the reconstruction term (ablation) | Wavelet Diffusion · CelebA-HQ 256×256 | FID: around 0.6 points to 5.94 | Improved |
| process both low- and high-frequency bottleneck subbands instead of passing high-frequency subbands through (ablation) | Wavelet Diffusion + W-Generator · STL-10 64×64 | FID: 12.96 vs. 12.93 | No clear effect |
| use progressive two-level upsampling (ablation) | Wavelet Diffusion · CelebA-HQ 256×256 | FID: 13.11 vs. 5.94 | Hurt |
Closest papers by meaning
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
- Bridging Data Gaps in Diffusion Models with Adversarial Noise-Based Transfer Learning (ICML 2024 · similarity 0.89)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.88)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.87)
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.
