Extensions/ComfyUI-VideoQuality-Metrics
ComfyUI Extension

ComfyUI-VideoQuality-Metrics

A comprehensive suite of video quality assessment metrics for ComfyUI including fidelity metrics, temporal metrics, distributional metrics, and deep learning-based quality evaluation.

By jajos12·Created 7 months ago·Updated 7 months ago· 1
jajos12/ComfyUI-VideoQuality-Metrics
Nodes13
On cloudLocal install
CategoryVideo Quality/CLIP, Video Quality/DOVER
Stars1
Updated7 months ago

Nodes (13)

VQ CLIP Aesthetic Score

The name is a vibe check, because that's what this node does

Video Quality/CLIP
VQ DOVER Quality (Aesthetic + Technical)

One number can't tell you why a video is bad. DOVER splits the answer

Video Quality/DOVER
VQ FAST-VQA Score (No-Reference)

High-res, short, full of noise

Video Quality/FAST-VQA
VQ Fréchet Inception Distance (FID)

FID is the number every image-model paper still leads with — now it runs in your graph

VideoQuality/Distributional
VQ Full-Reference Metrics (PSNR/SSIM)

The old-school trio that still catches things ML metrics miss

VideoQuality
VQ Fréchet Video Distance (FVD)

The number video-model papers end their abstracts with, inside your workflow

VideoQuality/Distributional
VQ Workflow Comparison

The README says statistical significance testing. The code says otherwise

VideoQuality/Reporting
VQ Metrics Logger (JSON)

The glue node that turns your scores into JSON you can save, diff, and plot

VideoQuality/Reporting
VQ Motion Smoothness

Catches the robotic, jittery motion that still frames can't reveal

VideoQuality/Temporal
VQ Radar Chart

One number can't describe a video, so draw a fingerprint instead

VideoQuality/Reporting
VQ Temporal Consistency (Warping/Flickering)

The node that catches shimmer, melting objects, and strobe lighting in your video

VideoQuality/Temporal
VQ Text-Video Alignment

Tells you if your video actually followed your prompt — or slowly forgot it

Video Quality/CLIP
VQ Video Frame-by-Frame FID

The name says FID, but check what it actually computes before you quote it

VideoQuality/Distributional
Readme

ComfyUI Video Quality Metrics

A comprehensive suite of video quality assessment metrics for ComfyUI.

Features

Fidelity Metrics

  • PSNR - Peak Signal-to-Noise Ratio
  • SSIM - Structural Similarity Index
  • CIEDE2000 - Perceptually uniform color accuracy

Temporal Metrics

  • Warping Error - Optical flow-based temporal consistency
  • Flickering Detection - Brightness variance analysis
  • Motion Smoothness - Jerk-based motion quality

Distributional Metrics

  • FVD - Fréchet Video Distance
  • FID - Fréchet Inception Distance

No-Reference Metrics (Deep Learning)

  • CLIP-IQA - Aesthetic Scoring & Text Alignment
  • DOVER - Disentangled Aesthetic & Technical Quality
  • FAST-VQA - Efficient High-Res Quality Assessment

Visualization & Reporting

  • Radar Charts - Multi-metric comparison
  • JSON Export - Metrics logging
  • Statistical Tests - T-test, Wilcoxon for workflow comparison

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/yourusername/ComfyUI-VideoQuality-Metrics.git
pip install -r ComfyUI-VideoQuality-Metrics/requirements.txt

Restart ComfyUI. Nodes appear under VideoQuality/ category.

Video & Batch Handling

All nodes accept Video Tensors [T, H, W, 3] and Image Batches [N, H, W, 3].

  • Fidelity Metrics (PSNR, SSIM, ΔE2000): Return the average quality across all frames in the batch.
  • Temporal Metrics: Require at least 2 frames (T ≥ 2) to compute motion consistency.
  • Distributional Metrics: Compare the entire distribution of the generated batch against the reference batch.

Node Reference

| Node | Inputs | Outputs | |------|--------|---------| | VQ Full-Reference Metrics | images, reference | psnr, ssim, ciede2000, summary | | VQ Temporal Consistency | video_frames | warping_error, flickering_score, summary | | VQ Motion Smoothness | video_frames | smoothness_score, mean_jerk, summary | | VQ FVD | video_generated, video_reference | fvd, summary | | VQ FID | images_generated, images_reference | fid, summary | | VQ CLIP Aesthetic | images | aesthetic_score, summary | | VQ Text Alignment | video, prompt | alignment_score, summary | | VQ DOVER Quality | video | aesthetic, technical, overall, summary | | VQ FAST-VQA | video | quality_score, summary | | VQ Radar Chart | psnr, ssim, ... | radar_chart, metrics_json | | VQ Metrics Logger | metrics | json_output |

Interpretation Guidelines

| Metric | Excellent | Good | Fair | Poor | |--------|-----------|------|------|------| | PSNR | > 35 dB | 30-35 dB | 25-30 dB | < 25 dB | | SSIM | > 0.95 | 0.90-0.95 | 0.85-0.90 | < 0.85 | | ΔE2000 | < 1.0 | 1.0-2.0 | 2.0-5.0 | > 5.0 | | FVD | < 50 | 50-150 | 150-300 | > 300 | | CLIP Aesthetic | > 0.7 | 0.6-0.7 | 0.4-0.6 | < 0.4 | | DOVER (Overall) | > 0.6 | 0.4-0.6 | 0.2-0.4 | < 0.2 | | FAST-VQA | > 0.65 | 0.5-0.65 | 0.3-0.5 | < 0.3 |

Technical Documentation

See doc.md for the complete theoretical framework.

License

MIT License