4KAgent

What does this agent do, and when should you use it?

4KAgent is a NeurIPS 2025 agentic super-resolution generalist that aims to restore any image to 4K, spanning classical and real-world degradations, extreme low-quality inputs, AI-generated images, and scientific imaging such as remote sensing, microscopy, pathology and medical scans. It pairs a Perception Agent that uses large vision-language models to analyze image content and distortion with a Restoration Agent that runs an execution–reflection–rollback loop over a toolbox of restoration and upscaling methods. A Quality-Driven Mixture-of-Expert (Q-MoE) policy picks the best candidate at each step, and a separate face restoration pipeline handles faces. Tasks are configured through profile files, so the system adapts without additional training, and inference is driven by the infer_4kagent.py CLI, which writes per-image folders of logs and intermediate outputs alongside evaluation scripts in eval/.

4KAgent reads images from an input directory, asks its Perception Agent — backed by a vision-language model such as llama_vision or an externally launched DepictQA service — to diagnose content and distortion, and emits a restoration plan. The Restoration Agent follows a profile (ExpSR_s4_F, ExpSR_s4_P, FastGen4K_P, OldP4K_P, GenMIR_P and others) to invoke deblurring, denoising, brightening, gamma correction and super-resolution tools, iterating through execution–reflection–rollback and selecting the best image via Q-MoE; with FaceRestore enabled it applies codeformer, gfpgan or keeps the original face. Outputs are per-image folders containing logs and stage-wise images; utils/image_export.py extracts final outputs, utils/toolchain_export.py prints the toolchain used per image, and utils/face_restoration_tool_export.py reports face tools. Scripts in eval/ compute PSNR/SSIM/LPIPS plus no-reference metrics such as NIQE, MUSIQ, MANIQA and CLIPIQA.

1. Researchers reproducing or benchmarking 4K super-resolution on Set5, Urban100, DRealSR and similar datasets
2. Remote-sensing teams upscaling and scoring satellite imagery such as AID, DIOR, DOTA or WorldStrat
3. Medical and biological imaging groups enhancing microscopy, pathology, X-ray or ultrasound images
4. Old-photo restoration work using the OldP4K_P profile on degraded historical photographs
5. AI-art creators pushing generated images to 4K while repairing faces through the face pipeline
6. Engineers who want to chain multiple restoration tools and audit which toolchain each image used

How do you install or deploy this agent?

Setup details live in the repository's Installation Guide (installation/Installation.md); the README does not list the full dependency set. An API key must be filled into config.yml before inference. To use DepictQA as the perception VLM, prepare its environment separately:

cd ./DepictQA
conda activate depictqa

To fetch the DIV4K-50 dataset:

python -m pip install "huggingface_hub[cli]"
huggingface-cli download --repo-type dataset YSZuo/DIV4K-50 --local-dir ./dataset/DIV4K-50
cd ./dataset/DIV4K-50
unzip DIV4K-50.zip

How do you use this agent?

Fill the VLM API key into config.yml first. With llama_vision as the perception VLM, run inference directly:

CUDA_VISIBLE_DEVICES=1 python infer_4kagent.py \
  --input_dir ./assets/profile_test_example/classicsr \
  --output_dir ./outputs/4KAgent_test/classicsr \
  --profile_name ExpSR_s4_F \
  --tool_run_gpu_id 2

With DepictQA, start its service in one terminal and run inference in another (the FastGen4K_P profile is recommended for speed and perceptual quality):

# terminal A
cd ./DepictQA
conda activate depictqa
CUDA_VISIBLE_DEVICES=0 python src/app_eval.py

# terminal B
CUDA_VISIBLE_DEVICES=1 python infer_4kagent.py \
  --input_dir ./assets/profile_test_example/4ksr \
  --output_dir ./outputs/4KAgent_test/4ksr \
  --profile_name FastGen4K_P \
  --tool_run_gpu_id 2

Other profiles include OldP4K_P for old photos and GenMIR_P for multiple-degradation restoration. Custom profiles can be based on examples in pipeline/profiles, documented in pipeline/profiles/profile_setup.md.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

FAQ