Research map / Image generation / Super-resolution
Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment
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.
Chain-of-Zoom wraps a fixed 4x diffusion SR model in an autoregressive scale-state loop with VLM-generated multi-scale prompts, and fine-tunes the prompt VLM with GRPO using a critic VLM, reaching 256x magnification without retraining the SR backbone.
Problem
Fast trained single-image super-resolution models are tied to their training scale factors and collapse when asked to magnify far beyond them, while training models directly for extreme factors is computationally prohibitive.
Approach
Factorize SR into an AR-2 chain of intermediate scale-states, reuse a 4x SR backbone at each step, and condition each step on text prompts from a VLM that sees the two previous scales; the prompt extractor is aligned via GRPO with a critic VLM reward plus phrase-exclusion and repetition penalties.
Key results
- At 256x on DIV2K, CoZ (VLM) reaches NIQE 9.83 versus 16.17 for direct SR (no-reference metrics, lower is better).
- S3Diff + CoZ improves 64x NIQE on DIV2K from 16.14 to 8.78 and 256x from 16.78 to 10.77.
- A 25-participant MOS study reports that GRPO-tuned prompts are preferred over untuned VLM prompts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CoZ instead of one-step direct SR | OSEDiff · DIV2K 64× | MUSIQ: 22.54 -> 52.13 | Helped |
| apply CoZ instead of one-step direct SR | OSEDiff · DIV8K 256× | MUSIQ: 28.90 -> 47.25 | Helped |
| add VLM prompts rather than use no text prompts (ablation) | CoZ with OSEDiff · DIV2K 256× | MUSIQ: 46.28 -> 47.83 | Helped |
| add VLM prompts rather than use no text prompts (ablation) | CoZ with OSEDiff · DIV8K 256× | MUSIQ: 46.56 -> 47.25 | Helped |
| use VLM prompts instead of DAPE prompts (ablation) | CoZ with OSEDiff · DIV2K 256× | MUSIQ: 46.22 -> 47.83 | Helped |
| use VLM prompts instead of DAPE prompts (ablation) | CoZ with OSEDiff · DIV8K 256× | MUSIQ: 45.81 -> 47.25 | Helped |
| use VLM prompts instead of DAPE prompts (ablation) | CoZ with OSEDiff · DIV2K 4× | MUSIQ: 67.01 -> 67.10 | Mixed |
| use VLM prompts instead of DAPE prompts (ablation) | CoZ with OSEDiff · DIV8K 16× | MUSIQ: 57.79 -> 57.99 | Helped |
Surprising findings
- At 4x, CoZ with the Null or VLM prompt does not improve over direct SR, so the gains appear only at high magnification.
- Cosine similarity of overlapping patches is reported as high at 256x, yet the reported mean values for 16x and above exceed 1, which the table does not explain.
Limitations
- Error accumulates across repeated zoom steps.
- Tiled full-image inference requires multiple VLM passes, which is a computational bottleneck.
- Evaluation relies on no-reference IQA metrics because ground truth is unavailable at 256x.
Experimental setup
Datasets: LSDIR FFHQ DIV2K DIV8K
Benchmarks: DIV2K DIV8K
Metrics: NIQE MUSIQ MANIQA CLIPIQA
Baselines: NN Interpolation Direct SR LIIF SeeSR S3Diff DAPE
Builds on: OSEDiff GRPO DAPE Stable Diffusion 3
Closest papers by meaning
- DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution (NeurIPS 2025 · similarity 0.90)
- The Power of Context: How Multimodality Improves Image Super-Resolution (CVPR 2025 · similarity 0.90)
- LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution (NeurIPS 2026 · similarity 0.89)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.88)
- SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning (NeurIPS 2025 · similarity 0.87)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · 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.