Research map / Image generation / Super-resolution

SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution

AAAI 2024 · Image generation › Super-resolution · approach: Diffusion-based super-resolution · 14 citations (OpenAlex)

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 SkipDiff, a DDPM inference scheme that runs coarse skipped diffusion for fidelity, then a PPO-trained policy network adaptively chooses fine diffusion steps for perceptual quality, for face and natural image super-resolution.

Problem

Perception-distortion trade-off in image super-resolution, together with the slow sampling of diffusion models.

Approach

Keep DDPM training unchanged and modify only inference: a coarse-grained skip diffusion (fidelity-driven) gives a latent approximation of the full chain, then a fine-grained stage whose steps are chosen by a PPO-trained skip proposal network (reward = NIQE or LPIPS of the final output) refines it for perception.

Key results

• On CelebA-HQ 8x face SR, SkipDiff reaches PSNR 25.60, NIQE 6.47 and LPIPS 0.0967, the best NIQE and LPIPS among the compared methods (SR3: 24.79 / 7.30 / 0.0992).
• On ImageNet 4x SR, SkipDiff with 20 steps reaches NIQE 4.95 and LPIPS 0.1886, versus SR3 (100 steps) at 5.00 and 0.1909.
• SkipDiff takes 0.69 s per image versus 4.17 s for SR3 and 2.23 s for DDRM on ImageNet.

Measured effects reported in the paper

Surprising findings

• Only removing the coarse stage (FSR alone over the full chain) gives no gain over the baseline, since the decision space becomes too large.
• The perceptual metrics are not monotonic in fine step count: the maximum fine step (9) did not give the best NIQE/LPIPS, so more fine steps are not always better.

Limitations

• Results are reported on a single 8x face and single 4x natural SR setting, with 1k-image evaluation subsets.
• Reward relies on NIQE or LPIPS, and the authors note NIQE is not claimed to be the best reward choice.
• Per-run training details are deferred to supplementary material, and the SPN training cost is only partially described.

Experimental setup

Datasets: FFHQ ImageNet 1K CelebA-HQ

Benchmarks: CelebA-HQ ImageNet validation

Metrics: PSNR SSIM NIQE LPIPS runtime (s) FLOPs (G)

Baselines: Bicubic FSRNet PULSE GCFSR SR3 IDM DGP DDRM

Builds on: DDPM SR3 PPO DDIM

Closest papers by meaning

• DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.91)
• SkipSR: Faster Super-Resolution with Token Skipping (NeurIPS 2026 · similarity 0.91)
• Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.89)
• The Surprising Effectiveness of Skip-Tuning in Diffusion Sampling (ICML 2024 · similarity 0.89)
• ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.89)
• Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.88)