Research map / Image generation / Super-resolution
4KAgent: Agentic Any Image to 4K Super-Resolution
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 4KAgent, a profile-driven multi-agent system that uses VLM/IQA perception, a Q-MoE restoration agent with rollback, and a face pipeline to upscale any image to 4K across natural, AIGC, and scientific domains without retraining.
Problem
Universal super-resolution of arbitrary images (extremely degraded, diverse domains) to 4K resolution, where specialist SR models generalize poorly and users need flexible workflows.
Approach
An agentic framework with a Perception Agent (IQA experts + VLM degradation reasoning and planning), a Restoration Agent that runs every tool in a per-task toolbox and picks the best output by a quality score (Q-MoE), a recursive execution-reflection-rollback loop, a face restoration sub-pipeline with identity-preservation scoring, and user-configurable profiles.
Key results
- On RealSR 4x, 4KAgent (ExpSR-s4-P) reports the best MANIQA (0.6602) and CLIPIQA (0.7078) among compared methods, with PSNR 24.60 versus AgenticIR 22.45.
- On the multiple-degradation MiO100 Group C subset, 4KAgent (GenMIR-P) reaches MUSIQ 55.56 versus MAIR 51.36 and AgenticIR 48.68.
- On WebPhoto-Test face restoration, 4KAgent (GenSRFR-s4-P) achieves the best NIQE (4.15), CLIPIQA (0.7077), MUSIQ (75.92) and MANIQA (0.6576), outperforming AgenticIR by a 19.67 MUSIQ margin.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Switch from the fidelity profile to the perception profile | 4KAgent · classical image super-resolution benchmarks | perceptual quality | Mixed |
| Use 4KAgent instead of AgenticIR | agentic super-resolution on Set14 and B100 · Set14 and B100 | multiple metrics | Helped |
| Use 4KAgent instead of AgenticIR | real-world image super-resolution · RealSR and DrealSR | perceptual quality | Helped |
| Use 4KAgent instead of other compared methods | multiple-degradation restoration · MiO100 Group A, Group B, and Group C | MUSIQ: 4.2 lead of MUSIQ on Group C | Helped |
| Use 4KAgent instead of AgenticIR | face restoration · WebPhoto-Test | MUSIQ: 19.67 lead on MUSIQ | Helped |
Surprising findings
- Under the fidelity-oriented profile, 4KAgent's PSNR is competitive with dedicated fidelity SR models, while the perception profile trades PSNR for perceptual scores.
- Agentic 4KAgent (GenSR-s4-P) has lower PSNR/SSIM than several specialist SR models on classical benchmarks, despite higher perceptual gains in some cases.
Limitations
- Relies on many external tools and VLMs, making the system complex and slow (multi-step, multiple models per step).
- Fidelity metrics (PSNR/SSIM) on several benchmarks lag behind specialist fidelity-based SR models in some settings.
Experimental setup
Datasets: Set5 Set14 B100 Urban100 Manga109 RealSR DrealSR MiO100 (Group A/B/C) WebPhoto-Test RealSRSet GenAIBench-4K DiffusionDB-4K
Benchmarks: Set5 Set14 B100 Urban100 Manga109 RealSR DrealSR MiO100 Group A/B/C WebPhoto-Test DIV4K-50 GenAIBench-4K DiffusionDB-4K
Metrics: PSNR SSIM LPIPS DISTS FID NIQE CLIPIQA MUSIQ MANIQA CLIB-FIQA
Baselines: SwinIR X-Restormer HAT-L HMA DRCT DiffBIR OSEDiff PiSA-SR AgenticIR ResShift StableSR SeeSR
Builds on: AgenticIR HPSv2 DepictQA OSEDiff
Closest papers by meaning
- Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.85)
- RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.85)
- Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2025 · similarity 0.84)
- InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD (NeurIPS 2024 · similarity 0.84)
- An Intelligent Agentic System for Complex Image Restoration Problems (ICLR 2025 · similarity 0.84)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.84)
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.