Research map / Image generation / Fast diffusion sampling
InstaRevive: One-Step Image Enhancement via Dynamic Score Matching
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.
Distills a pre-trained text-to-image diffusion model into a one-step image enhancer using score-based distillation with dynamic noise-level and loss control plus BLIP captions as text prompts.
Problem
Diffusion-based image enhancement yields high quality but needs slow iterative sampling; naive score distillation for restoration gives inaccurate gradients and over-smoothing.
Approach
Adapts variational/score distillation to image enhancement with a dynamic diffusion-noise scope (T_max set from the current HQ-generated distance, scaled by a control factor) and a dynamic weighting that blends regression loss and score-matching loss, plus captioner-derived text prompts as conditioning.
Key results
- On CelebA-Test BFR, InstaRevive reaches 22.3259 PSNR and 19.78 FID, and 38.73 FID on LFW-Test, versus 43.36 for the best prior LFW FID shown (DMDNet).
- On RealSR and RealSet65 BSR, InstaRevive attains MANIQA 0.4722 and 0.4571 respectively, highest among compared methods including SinSR.
- Single-step InstaRevive runs at 0.14 s on RealSet65 with a 611M DiT-based generator, versus 1.13 s for 15-step ResShift and 10-14 s for multi-step SUPIR/DiffBIR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic noise and loss control (ablation) | InstaRevive · Blind image enhancement | image quality | Helped |
| add textual prompt guidance instead of null textual input (ablation) | InstaRevive · Blind image enhancement, especially challenging scenarios | performance | Helped |
| add score matching to regression-only distillation (ablation) | InstaRevive · Blind image enhancement | visual quality | Helped |
Surprising findings
- Restorer SwinIR-GAN alone performs poorly on ImageNet-Test, suggesting the diffusion generator contributes most of the gain.
Limitations
- Suboptimal results under extreme degradation or complex, dense content (failure cases).
- Cannot surpass larger multi-step models such as SUPIR in quality.
- Score estimate for generated distribution is less accurate, requiring dynamic loss balancing.
Experimental setup
Datasets: FFHQ ImageNet CelebA-Test LFW-Test WIDER-Test RealSR RealSet65 ImageNet-Test LoL BDD100k RainTrainH Rain-100L
Benchmarks: CelebA-Test LFW-Test WIDER-Test RealSR RealSet65 ImageNet-Test
Metrics: PSNR SSIM LPIPS FID IDS MANIQA MUSIQ CLIPIQA NIQE FPS
Baselines: GPEN GCFSR GFPGAN VQFR RestoreFormer DMDNet CodeFormer DifFace ResShift RealESRGAN BSRGAN SwinIR-GAN
Builds on: Variational Score Distillation Score-SDE ResShift LoRA BLIP
Closest papers by meaning
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.88)
- Contrastive Denoising Score for Text-guided Latent Diffusion Image Editing (CVPR 2024 · similarity 0.88)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.87)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.87)
- CoDi: Conditional Diffusion Distillation for Higher-Fidelity and Faster Image Generation (CVPR 2024 · similarity 0.87)
- 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.