Research map / Image generation / Super-resolution
Arbitrary-steps Image Super-resolution via Diffusion Inversion
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 InvSR, a diffusion-inversion image super-resolution method that trains a noise predictor to produce a partial-noise starting state of a frozen SD-Turbo model, enabling arbitrary 1-5 step sampling.
Problem
Diffusion-based super-resolution either modifies diffusion intermediate features or fine-tunes large T2I models, and existing efficient methods fix the number of sampling steps; the paper targets high-fidelity SR with flexible, efficient sampling built on diffusion priors.
Approach
A Partial Noise Prediction strategy: a trained noise predictor (VQGAN-encoder-based) estimates the noise to add to the LR latent at a high-SNR starting timestep (<=250), so sampling from that intermediate state with any accelerated sampler (1-5 steps) yields the HR result, while the diffusion backbone stays frozen.
Key results
- On ImageNet-Test, InvSR-1 reaches 24.14 PSNR and 0.2517 LPIPS, beating one-step OSEDiff-1 on all seven metrics as reported by the authors.
- InvSR-1 achieves 4.0284 NIQE and 72.29 MUSIQ on RealSet80/RealSR-type non-reference metrics, second-best or better among methods, while using about 34M learnable parameters.
- Varying the starting timestep with fixed steps trades off fidelity (reference metrics) against realism (non-reference metrics); more sampling steps lower both reference and non-reference scores in the reported ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InvSR-1 instead of ResShift-4 | ImageNet-Test ×4 super-resolution · ImageNet-Test | PSNR: 27.33 -> 24.14 | Hurt |
| Use InvSR-1 instead of StableSR-50 | ImageNet-Test ×4 super-resolution · ImageNet-Test | CLIPIQA: 0.7067 -> 0.7093 | Helped |
| Use InvSR-1 instead of LDM-50 | RealSR ×4 super-resolution · RealSR | PSNR: 26.75 -> 24.50 | Hurt |
| Use InvSR-1 instead of OSEDiff-1 | RealSR ×4 super-resolution · RealSR | CLIPIQA: 0.7008 -> 0.6918 | Hurt |
| Use InvSR-1 instead of DiffBIR-50 | RealSet80 ×4 super-resolution · RealSet80 | NIQE: 3.8630 -> 4.0284 | Hurt |
| Increase sampling from one step to multiple steps (ablation) | InvSR on ImageNet-Test · ImageNet-Test | PSNR: 24.66 -> 22.70 | Hurt |
| Start one-step sampling from timestep 100 rather than timestep 250 (ablation) | InvSR on ImageNet-Test · ImageNet-Test | PSNR: 23.84 -> 24.66 | Helped |
| Predict noise for intermediate states rather than using random noise (ablation) | InvSR · Image SR | performance | No clear effect |
Surprising findings
- Non-reference quality metrics decline when more sampling steps are used, because test images often contain noise that multi-step sampling amplifies.
- Adding a learned noise predictor for intermediate timesteps gave no significant gain, so random noise is used there and only the initial noise is predicted.
Limitations
- Non-reference metrics degrade with more sampling steps, so multi-step sampling can amplify noise in noisy inputs.
- Evaluation on the synthetic ImageNet-Test uses the authors' own constructed benchmark, and the two real-world sets are small (100 and 80 images).
- Limited to starting timesteps <=250 to keep high SNR, which constrains the degree of generative refinement.
Experimental setup
Datasets: LSDIR FFHQ ImageNet validation set RealSR RealSet80
Benchmarks: ImageNet-Test RealSR RealSet80
Metrics: PSNR SSIM LPIPS NIQE PI CLIPIQA MUSIQ
Baselines: BSRGAN RealESRGAN LDM StableSR DiffBIR SeeSR ResShift SinSR OSEDiff
Builds on: DDPM DDIM SD-Turbo RealESRGAN VQGAN
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.93)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.93)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.92)
- Restabilizing Diffusion Models with Predictive Noise Fusion Strategy for Image Super-Resolution (AAAI 2025 · similarity 0.92)
- Efficient Conditional Diffusion Model with Probability Flow Sampling for Image Super-resolution (AAAI 2024 · similarity 0.91)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 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.