MediaPipe Eyes Mesh

The hardest part of fixing someone's eyes in a generated image isn't the fixing - it's telling the model where the eyes are. This node does that for you in one shot: feed it a face image and you get back a soft, ready-to-use mask of just the eye region, computed by MediaPipe on Runware's cloud. It's the classic "make a mask, then inpaint through it" workflow, minus the hours of lassoing.

What it actually does

This is an imageMasking node in the Runware pack, which means the heavy lifting happens server-side. Your image gets uploaded, MediaPipe's face-mesh landmarks find the eyes, and the node returns the mask as an image output (an IMAGE tensor, not the MASK type). It's a mask of the eye areas, with the face itself left untouched.

The output is the mask, not an edited photo. So the real workflow is: this node → mask → an inpainting model → done. In vanilla ComfyUI you'll usually run the mask through the core ImageToMask node (or feed it to a Runware inpaint node that takes IMAGE masks directly) and then use it to pin the inpaint to the eyes. The unmasked pixels stay bit-identical, which is exactly what masked inpainting still owns over whole-image edit models.

The settings that matter

Four of them, and they're all straightforward:

• settings.confidence (default 0.5) - how sure MediaPipe has to be before a detection counts. If it keeps missing eyes, drop it; if it's grabbing random blobs, raise it.
• settings.maskBlur (default 5) - feathering on the mask edges. 5 is a good start for eye work; go lower if you want hard edges for a relight, higher if you're getting seams.
• settings.maskPadding (default 10) - pixels to extend the mask beyond the detected eyes. Positive grows it (safer for inpaint), negative shrinks it.
• settings.maxDetections (default 6) - how many faces to process when there's a crowd.

The ttl / ttl_value and outputFormat (JPG/PNG/WEBP) widgets are shared pack plumbing - you can ignore them. Everything runs in the cloud, so there's no local model to download.

Install

The whole pack installs at once, since these are all nodes in Runware/ComfyUI-Runware:

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

Restart ComfyUI. Easier still: open ComfyUI Manager, search "Runware", install, restart. The dependencies are light (runware-sdk, pillow, soundfile) - no CUDA, no gigabytes of model files. You do need an API key: grab one from the Runware dashboard, then paste it in ComfyUI Settings → Runware API key, set the RUNWARE_API_KEY env var, or run runware auth login once.

Gotchas

Every run is a paid cloud inference, and the node's title bar shows the cost plus a NSFW flag after each run - worth glancing at if you're batch-running a folder of faces. And remember the output is a grayscale mask image; if a node downstream wants a MASK socket, convert first. That trips people up more than anything else.