YOLOv8s Face

Faces are where generation goes to betray you - that slightly-off mouth, the eyes that don't match - and the standard fix is to mask the face and regenerate just that region. YOLOv8s Face is the mask-making half of that fix, running on Runware's cloud. It detects faces in an image and returns a mask of them as an IMAGE, with edge controls built in. The "s" means small: a step up from nano, a step down from medium - the sweet spot for detection quality without the cost of the bigger models.

The workflow it enables is one of the oldest and most reliable in the ecosystem: detect the face, mask it, inpaint it with the rest of the image locked down. Locally that means installing a face-detection model and wiring up the segmentation plumbing; here it's a single node with one image input. If you've already got a mask-based inpainting setup and just needed a reliable face detector that doesn't live on your GPU, this drops straight in.

How it works

An imageMasking task on Runware's cloud (runware:35@2): the node uploads your image via the SDK, sends the request over REST, and downloads the resulting mask into an IMAGE tensor. All the real controls shape that mask:

• settings.confidence (0.5) - only faces above this score get masked. Lower for smaller/angled faces, but you'll also catch false positives.
• settings.maskBlur (5) - edge smoothing, so the mask composites without a harsh cut line.
• settings.maskPadding (10) - extend the mask by pixels. For face inpainting you usually want padding - regenerate a little context around the face, not just the skin.
• settings.maxDetections (6) - cap on faces masked, highest confidence first.

Output is image (IMAGE) - feed it into your mask-based inpaint/restoration workflow.

The inputs that matter

Required input: image. For a typical face-fix pass, start at confidence 0.5, maskPadding +10, maskBlur 5. Group photos will hit the maxDetections cap - bump it if you're processing a crowd, but remember every detection adds compute (and cost).

Install and API key

Install once for the whole pack:

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

Restart ComfyUI (or install Runware from ComfyUI Manager). API key from runware.ai/api-keys, set in ComfyUI Settings → Runware API key, exported as RUNWARE_API_KEY, or via runware auth login.

Gotchas

Profile shots and heavily angled faces are where the small model wobbles - if a face isn't detected, drop confidence before you blame the image. The mask is per-detection composited, so overlapping faces can merge into one blob; maskPadding down helps there. And the usual metered-API reminder: each image is a paid call, and unlike running YOLO locally (where the model is free and your GPU does the work), this costs per mask. If you're batch-processing hundreds of faces, the local route eventually wins on price - this node is for the occasional fix or when you don't have the local stack. The output is a mask image, so confirm your inpaint node consumes IMAGE masks the way it expects.