凤希AI - 人脸精修 · 轻量链 V2（全脸）

You rendered a wide shot - three people in a field, camera 20 metres back - and every face in it is a smear. A face occupying 70 pixels of a 1024px frame has no latent budget to spend, which is exactly why the detect-crop-upscale-repaint loop ("face detailer") exists.

The usual version of that loop has a side effect nobody warns beginners about: the repaint is a diffusion pass, so it invents a plausible face and you lose the person you actually rendered. FxAiFaceRestoreLiteV2 (凤希AI - 人脸精修 · 轻量链 V2（全脸）) does the same loop with the repaint swapped for a one-step generative superresolution pass, and then, if you give it reference photos, pins each face back to the identity it belongs to. "全脸" in the name means all the small faces, not just one designated protagonist.

The loop, mechanically

Detection is InsightFace's buffalo_l - the same backbone under InstantID, PuLID and every face-swap tool you've used. The pack adds ROI tracking: after the first frame it only re-detects a 2.2× crop around the previous box, which beats scanning a whole 1280×768 frame for a 43px face. It also pre-scans the clip once and caches boxes and embeddings.

Then the gate. 处理脸宽上限 defaults to 96: any detected face narrower than that gets processed, anything at or above it passes through untouched. That single number is the speed dial for the whole node - a clear face needs no repair, so don't pay for it. Set it to 0 and every face in every frame goes through.

Each face is cropped by its own width (1.5×, snapped up to 128/256/512), and the upscale factor is chosen so crop × factor lands exactly on 512 - the native training resolution of the VOSR2 superres engine. Then the reconstructed patch goes back through an ellipse mask (遮罩比例, 0.70 of face width) feathered at 羽化半径 (0.045 of the crop). Small numbers, but they're the difference between a repaired face and a visible pasted square.

Around the threshold there's a 15% hysteresis band keyed to the previous frame's box overlap: a face already being fixed has to grow to 1.15× before it stops, one already skipped has to shrink to 0.85× before it starts. Without it, a face hovering at 96px flickers - fixed, skipped, fixed - and reads as a twitch in the final video.

Identity: the swap isn't a face swap

This is the part that's easy to misread. 参考图 is optional: one image per character you want recognised. Every detected face is compared against all of them by cosine similarity, and the closest match above 身份相似度阈值 (default 0.10) is treated as that character - then swapped toward that person's own reference photo, not toward someone else. No protagonist slot; each character keeps their own face. Leave 参考图 unconnected and the node only superresolves, which is a perfectly good way to run it.

身份相似度阈值 is the one to tune from the report. Distant small faces realistically score 0.02–0.20, so 0.10 sits mid-range; drop to 0.06 or 0.04 when faces stop being written back, raise to 0.2 or 0.3 when the wrong person gets picked. 默认人物 (default 1) is a narrow fallback: one face in the frame and nothing clearing the threshold, write back reference #1 anyway.

换脸强度, 肤色回拉 and 纹理保留 clean up the swap afterwards. Skin pull-back matters more than it sounds: swap models inherit the target image's skin tone, and without pulling it back a face lit warm or dim comes out as a pale mask.

Outputs

处理后帧 is the whole batch, same order and resolution - wire it into a save node. 处理报告 is a multi-line string: frames processed, faces skipped as too clear, faces attributed per reference, plus every warning. Put it through a text preview on your first run - it's the only place the node admits it degraded.

Installing it

ComfyUI Manager, search fxai-toolkit (or 凤希), or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit
# restart ComfyUI

There is no requirements.txt; pyproject.toml lists no dependencies, and the pack's __init__.py pip-installs exactly soundfile and psutil on startup. Everything this node needs, you supply:

• pip install insightface onnxruntime-gpu into the same Python that runs ComfyUI (Windows portable: python_embeded\python.exe -m pip ...).
• buffalo_l into ComfyUI/models/insightface/models/buffalo_l/ - five .onnx files. Without it the node falls back to downloading, which is where runs mysteriously "hang while fetching a model".
• The VOSR2 bundle into ComfyUI/models/vosr2/VOSR2/: args.json, checkpoints/ema_model.safetensors, a Qwen-Image-vae-2d/ folder, dinov2_vitl14.safetensors.
• For the identity pass, inswapper_128.onnx or preferably hyperswap_1a_256.onnx in ComfyUI/models/insightface/models/ - the pack prefers the 256 model when both exist. Optional; the node runs without it.

Where people get burned

Missing VOSR2 keys don't error, they downgrade. If the bundle is absent or VRAM runs out, you don't get a crash - you get Lanczos interpolation, i.e. bigger blur. Flip 超分引擎 to 纯插值 (Lanczos) deliberately if you want that; if you didn't ask for it, read the report line before blaming the model.

It needs roughly 6.2 GiB free for VOSR2 + swap model + detector together, and only frees other models when memory is actually short - awkward if you launch with --disable-smart-memory, where ComfyUI won't unload your video sampler.

ONNX on CPU without saying so. If onnxruntime-gpu can't load cudnn64_9.dll, detection silently drops to CPU: the console prints 人脸检测 GPU 不可用，退回 CPU and per-frame time triples. Environment problem, not a settings problem.

Faces under 16px are ignored entirely, 换脸脸宽上限 is clamped to 处理脸宽上限 if you set it higher, and on Windows 显卡编号 above 0 won't take effect unless ComfyUI starts with --cuda-device all.

One last thing worth knowing before you put this in anything commercial: the InsightFace models are non-commercial, and HyperSwap ships under a research licence. Same trap as every other identity pipeline in the ecosystem - the code is permissive, the weights are not.