Nano Banana - Face Swap (Mask Inpaint)

Nano Banana's celebrity/public-figure classifier fires on the original face, not just the identity you're aiming for. NanoBananaInpaintSwap is the pack's third pathway and its bluntest workaround: it detects the face, paints it solid red, and only then sends the image to Gemini - so the model literally never sees the face it would have refused to touch.

This is the node you reach for when the whole-image path dies on a refusal and the crop path dies too. Where Crop & Composite works around scene content, this one works around identity content. It's also the closest thing in the pack to classic masked inpainting: a specified region gets redrawn from a prompt while the rest of the frame stays bit-identical, because Gemini only ever sees the obscured crop.

How it works

The flow: FaceMesh landmarks build a face polygon, mask_dilate_px expands it outward, and the region is filled red. That obscured image plus your identity references go to Gemini with an edit instruction - the model redraws a face matching the refs into the red blob. Back on your side, the result is composited into the original frame with the same integration stack as the other swap nodes: LAB color match, grain match matched to cheek variance, sharpness match that only ever blurs (never sharpens), and a Laplacian or feathered composite.

The red-paint trick has a side benefit beyond evasion: because Gemini sees no original face to preserve, it has freedom to redraw bone structure, which is why the pack suggests raising mask_dilate_px when you want a more thorough head rework.

Inputs that matter

• api_key / model - Gemini key (or GEMINI_API_KEY env) and the model: gemini-3.1-flash-image-preview default, gemini-3-pro-image-preview for higher quality, gemini-2.5-flash-image for the older model.
• target_image + identity_1..identity_6 - the frame and up to six refs. Set grid_mode to auto_sheet if you're feeding a composed sheet.
• scope - face, head, or head+styling, same semantics as the rest of the pack.
• mask_dilate_px (default 25) - how far the red polygon puffs out. This is the knob for this node. Higher = more room for Gemini to redraw bone structure; too high and the red paint bleeds into hair.
• mask_feather_px (default 16) - softness of the composite edge. 16–32 is typical.
• detector, crop_size, plus the integrate stack (color_match, grain_match, sharpness_match, composite_method, lab_strength, grain_strength). match_directional_lighting stays off for cross-identity swaps.
• The v0.2 extras are all here too: dry_run for a free prompt preview, timeout_ms, ref_cap_px, auto_relax_on_refused, and the optional network input for a proxy route node.

Outputs

The standard four: image (result, red-tinted on refusal), status (OK/REFUSED:<category>/ERROR:<reason>), mask, and debug_sheet.

Install

Same pack, same steps:

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

Restart ComfyUI (or ComfyUI Manager → "NanoBanana FaceSwap"). You need a Gemini API key, and the mediapipe dependency for the detector cascade - the requirements install handles it.

Troubleshooting

• Oval "sticker" boundary around the face - the classic sign mask_dilate_px is too big and the red blob carved a visible shape. The pack's workflow notes suggest lowering it (toward 12 or below).
• Double jawline / face not tracking the original head - the opposite: too little dilation, so Gemini didn't have room to redraw. Raise mask_dilate_px.
• REFUSED even though the original face is hidden - the red blob is still a face-shaped region, and the classifier can guess. This is where the pack's Unbiased pathway or the obfuscator picks up.

Batch tip: with batch_axis, each frame or identity gets its own API call and its own retry - one refusal in a run doesn't stop the rest.