A comprehensive Image Quality Assessment (IQA) custom node collection for ComfyUI. Combines deep learning-based metrics (PyIQA), classical computer vision metrics (OpenCV), and advanced analysis tools for detailed image inspection — all directly in your ComfyUI workflow.
| Layer | Technology |
|---|---|
| Runtime | Python 3.10+ |
| ComfyUI API | Custom node (standard API) |
| Deep learning IQA | PyIQA (HyperIQA, MUSIQ, NIMA, LPIPS, FID, SSIM, and more) |
| Classical metrics | OpenCV |
| Acceleration | CUDA (auto-detected) / CPU fallback |
| Model caching | In-memory smart cache (avoids reload on repeated use) |
Access to a large zoo of state-of-the-art IQA models via the PyIQA library.
- Smart caching — models cached in memory across workflow runs
- Batch support — processes image batches with configurable score aggregation
- Auto device detection — CUDA if available, otherwise CPU
- Available models — HyperIQA, MUSIQ, NIMA, LPIPS, FID, SSIM, BRISQUE, NIQE, and many more
Classical and advanced metrics computed entirely without deep learning:
| Metric type | Available metrics |
|---|---|
| Classical | Blur, Brightness, Contrast, Colorfulness, Noise |
| Advanced | Color Harmony, Color Temperature, Defocus (FFT), Clipping, Entropy |
| Visualizations | Histograms, Heatmaps, Color Wheels, Edge Maps |
- Filter / Rank — filter or sort images by quality score
- Ensemble — combine multiple metric scores into a single value
- Frontend integration — real-time score display directly on the node in the ComfyUI canvas
cd ComfyUI/custom_nodes
git clone https://github.com/tazztone/ComfyUI-Image-Quality-Assessment
cd ComfyUI-Image-Quality-Assessment
pip install -r requirements.txt
# Restart ComfyUIOr install via ComfyUI Manager by searching for ComfyUI-Image-Quality-Assessment.
Nodes appear in the ComfyUI menu under Image → Quality Assessment. Connect any IMAGE output to a node input. Scores are returned as FLOAT values and displayed live on the node.
Typical workflow patterns:
- Score a batch of generated images → filter to keep only top-N by quality
- Compare two images with LPIPS or SSIM for similarity
- Use NIQE or HyperIQA as a no-reference quality gate before saving outputs
- Visualize color/exposure issues with heatmap and histogram nodes
MIT