Forbidden Vision 🧩 Face Edit Prep

The Fixer is all-in-one: it detects, inpaints, and composites in a single node. That's great until you want to do the editing part yourself - run the crop through an external editor, a hand-drawn mask, or a node that isn't the pack's own sampler. Face Edit Prep is the "give me the raw material" half of that workflow. It runs the same detection and segmentation models as the Fixer, but instead of sampling anything it hands you a cropped face, a clean mask, and the geometry you'll need to put the result back.

What it outputs

Three things, and the third is the clever one:

• cropped_face - the face region, squared up to processing_resolution (default 1024) with crop_padding (default 1.6) of surrounding context so an editor sees lighting and style, not just skin.
• cropped_mask - the segmentation mask for that crop.
• face_info - a FACE_INFO packet that records where the crop came from in the original image, the rotation that was applied, and the mask geometry. This is the glue: you feed it to the sibling Face Edit Merge node later so it knows exactly where and how to stitch your edited face back.

You set the usual suspects: face_selection (0 = all faces, or target a specific one), detection_confidence (default 0.8), enable_segmentation (contour masks on, oval fallback off), plus the mask-processing trio of mask_expansion, sampling_mask_blur_size, and sampling_mask_blur_strength. enable_pre_upscale + upscaler_model bump small faces up to a workable resolution before cropping, which matters if you're editing a face that was 60 pixels wide in the original.

The one input that changes everything

The optional mask input. If you feed one in, detection is skipped entirely - the node crops around your mask instead of hunting for faces. That's the escape hatch when the YOLO detector misses a heavily stylized face or you want to edit a region that isn't a face at all. isolate_face is the other flag worth knowing: when on, it isolates the face from hair and background context in the crop, which is handy if your downstream edit cares only about skin.

How to use it in a workflow

Load Image → Face Edit Prep → cropped_face → (your own edit: manual inpaint,
                                              another node, an external app)
                         → face_info ─────────────┐
                         → cropped_mask ──────────┤
                                                  ▼
                         edited face → Face Edit Merge → final_image

The pair ships together, and the README's example workflows show the same detection models doing the work, so what you see in the preview is what the Fixer would have found - just with the sampling step handed to you. It's the right node when you want control over the edit itself, at the cost of wiring up a few more connections than the all-in-one Fixer.