图像质量评估器 (FID/IS/CLIP/LPIPS)

Let's be honest about what this node is before we get to how it works. ImageQualityEvaluator (the UI shows it as 图像质量评估器) is not a "is my image good" filter - there's no such thing, and your eyeballs are still the best judge of a single render. It's a batch-evaluation terminal for the four classic academic metrics - FID, Inception Score, CLIP-Score, and LPIPS - computed on whole folders of images sitting on your disk. You reach for it when you need numbers: benchmarking a LoRA or checkpoint against a reference set, tuning a sampler, comparing two pipelines, or writing up results that someone might actually check.

What makes it rare is the packaging. Getting these four metrics normally means leaving ComfyUI and running torch-fidelity or clean-fid scripts, wrangling pip environments, and pasting output back in. This node does all four in one shot and prints a report as a string you can save. It's small, MIT-licensed, and one author's hobby project rather than a big ecosystem pack - so set expectations accordingly, but the math underneath is the real thing.

How it works

The node reads image folders straight from disk (not from the ComfyUI graph - the wires here are optional), resizes everything to 299×299, and runs it through a stock torchvision Inception V3. The 2048-dim feature vectors give you FID (via scipy's matrix square root, with a touch of covariance regularization to avoid the infamous "Imaginary component" crash); the 1000-class softmax probabilities give you IS. CLIP-Score comes from LAION's clip-score package, and LPIPS from the lpips package with its AlexNet backbone. So the report is only as meaningful as its inputs: this is comparing your generated batch against a reference set, not grading individual images.

The author deliberately skipped clean-fid and torch-fidelity and hand-rolled the Inception path - the changelog says it was to kill the "size mismatch" and imaginary-component errors those libraries are famous for. That's a real, earned fix, and the code shows it: every failure is caught per-metric and reported as a line in the output instead of killing your whole queue.

The inputs that matter

Only a handful of fields deserve your attention:

• generated_images_dir - the folder of images you're evaluating. The only one that's a hard requirement; if this path is bad, the node errors out.
• real_images_dir - your reference/real folder for FID. If it's invalid, FID is soft-skipped and the rest still runs. Don't leave it empty unless you only want IS.
• prompts - one prompt per generated image, in filename-sorted order, one per line. Only needed for CLIP-Score; without it you get "⚠️ 未提供 prompts" rather than a crash.
• compare_images_dir - the reference folder for LPIPS, paired positionally with your generated folder.
• batch_size - default 32, and the first knob to turn if you hit CUDA OOM (the README's exact advice).
• device - cuda or cpu. Bonus: if you pick cuda and have no GPU, it silently drops to CPU instead of erroring.

The four enable_* toggles let you run just FID+IS (the common minimal setup) and skip the ones whose inputs you don't have. Output is a single evaluation_result STRING - a formatted ASCII report with per-metric verdicts - and if you fill save_path, the same text is written to that file, no SaveText node required.

Installing it

Easiest via ComfyUI Manager - search ComfyUI-ImageEval. Or by hand:

cd ComfyUI/custom_nodes/
git clone https://github.com/zx725tt/ComfyUI-ImageQualityEvaluator.git
pip install scipy clip-score lpips

Then restart. Two gotchas the README is unusually loud about: the heavy deps (scipy, clip-score, lpips) are not auto-installed, and if you run a portable build like ComfyUI-aki you must install them with ComfyUI's own embedded Python (...\python\python.exe -m pip install ...), not your system pip - otherwise the node silently does nothing. First run also downloads Inception V3 weights and the LPIPS/CLIP weights, so expect a one-time network pull.

Troubleshooting

• Runs but no output / "缺少依赖" - deps landed in the wrong Python. Reinstall with ComfyUI's embedded Python and restart. There's a eval_debug.log in the node's folder that spells out exactly which import failed.
• "Prompt has no outputs" - this node is an output node, so update the pack and restart; it's usually a stale install.
• Wildly unstable numbers - you're under the recommended image counts (the node itself prints them: FID wants 500+, IS/CLIP 100+, LPIPS 50 pairs). FID from 20 images is a coin flip; don't report it.
• CUDA OOM - drop batch_size to 8 or 4.

One honest caveat to finish on: it's a young, lightly-maintained pack from a single author, so treat it as a convenience wrapper rather than a research-grade harness. But if you need defensible FID/IS/CLIP/LPIPS numbers and you already live in ComfyUI, it's the path of least resistance.