Research map / Image generation / Super-resolution
DREAM: Diffusion Rectification and Estimation-Adaptive 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 DREAM, a three-line training modification for diffusion models that rectifies training with the model's own estimates and adaptively blends in ground truth, improving image super-resolution quality and training/sampling efficiency.
Problem
Diffusion models for super-resolution suffer a training-sampling discrepancy: training denoises from ground-truth-based noisy images while sampling conditions on self-generated estimates, which hurts quality and slows training and sampling.
Approach
Diffusion rectification re-noises the model's own one-step HR estimate during training (step-unrolling style), and estimation adaptation blends that estimate with ground truth using a weight (sqrt(1-alpha_bar_t))^p so the rectification term fades at small t, balancing perception and distortion.
Key results
- On 8x CelebA-HQ, SR3+DREAM improves PSNR from 23.85 to 24.63 dB and reduces FID from 61.98 to 56.01.
- On 4x DIV2K, DREAM gives SR3 +1.08 dB PSNR and ResShift +3.14 dB PSNR, with ResShift LPIPS dropping by 0.087.
- DREAM reaches convergence at about 100k-150k iterations versus about 400k for SR3, and the paper reports 2-3x faster training convergence and 10-20x fewer sampling steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DREAM with adaptive estimation (p=1) | SR3 · CelebA-HQ 8× SR | PSNR: 23.85 -> 24.63 | Helped |
| Add DREAM with adaptive estimation (p=1) | SR3 · CelebA-HQ 8× SR | FID: 61.98 -> 56.01 | Helped |
| Add DREAM | SR3 · DIV2K 4× SR | PSNR: 27.02 -> 28.10 | Helped |
| Add DREAM | ResShift · DIV2K 4× SR | PSNR: 25.30 -> 28.44 | Helped |
| Add DREAM | ResShift · DIV2K 4× SR | LPIPS: 0.211 -> 0.124 | Helped |
| Add static rectification without estimation adaptation (p=0) (ablation) | SR3 · CelebA-HQ 8× SR | PSNR: 23.85 -> 25.04 | Helped |
| Add static rectification without estimation adaptation (p=0) (ablation) | SR3 · CelebA-HQ 8× SR | FID: 61.98 -> 77.51 | Hurt |
| Use adaptive estimation (p=1) instead of static rectification (p=0) (ablation) | SR3 · CelebA-HQ 8× SR | FID: 77.51 -> 56.01 | Helped |
| Increase adaptation exponent from p=1 to p=2 (ablation) | SR3 + DREAM · CelebA-HQ 8× SR | FID: 56.01 -> 61.72 | Hurt |
| Increase adaptation exponent from p=1 to p=3 (ablation) | SR3 + DREAM · DIV2K 4× SR | PSNR: 28.10 -> 27.88 | Hurt |
| Apply DREAM to IDM | IDM · CelebA-HQ 8× SR | PSNR: 24.01 -> 24.50 | Helped |
| Apply DREAM to IDM | IDM · CelebA-HQ 8× SR | consistency (MSE ×10−5): 2.14 -> 1.26 | Helped |
Surprising findings
- Pure diffusion rectification (p=0) lowers distortion but noticeably worsens perceptual quality (higher FID, lower LPIPS), so the naive self-alignment alone is insufficient.
Limitations
- Evaluated only on super-resolution tasks and three diffusion SR baselines.
- Perceptual quality still trails GAN/flow methods such as HCFlow++ on LPIPS; the authors note GAN loss and architecture changes are orthogonal and left for future work.
- Official SR3 code was unavailable, so a third-party implementation was used.
Experimental setup
Datasets: FFHQ DIV2K Flickr2K CelebA-HQ CAT LSUN
Benchmarks: CelebA-HQ DIV2K CAT LSUN
Metrics: PSNR SSIM LPIPS FID consistency
Baselines: SR3 IDM ResShift PULSE FSRGAN EDSR RRDB ESRGAN SRFlow HCFlow HCFlow++ RankSRGAN
Builds on: DDPM DDIM SR3 IDM ResShift
Closest papers by meaning
- SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution (AAAI 2024 · similarity 0.91)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.90)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.90)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.90)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 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.