BD Face Mosaic Compose

Not every face-wrap job needs a mathematically correct unwrap. BD Face Mosaic Compose is the Photoshop-style alternative: it takes your 4 head-view photos, cuts them, aligns them, and pastes them side-by-side with feathered overlap into a strip. Each face stays geometrically intact - a similarity transform (rotate + scale + translate) gets the eyes horizontal and the face sized to the strip, but there's no cylindrical unwrap and no TPS deformation. The pack's own framing says it best: when the result is good enough that Qwen finalize is only doing cleanup, this is the node you want; when you need a true cylindrical unwrap, use BD Face Strip Compose instead.

How it works

For each detected view (front/left/right), a similarity transform is computed from the landmarks so the eyes are horizontal, the face height matches face_height_ratio × strip_height, and the eye-center lands at its target X column (front_center_x, left_center_x, right_center_x). Each view gets a feathered face-oval alpha and they're composited by alpha-weighted average - overlaps blend smoothly across the feather band, no visible cut lines. align_to_front (on) anchors each side view to a shared landmark (the tragion) so everything lines up at the right scale.

The rear photo is handled separately: split at rear_split_x and slotted at the strip edges. If you wire rear_mask (a head silhouette from BD_SAM3MultiPrompt or similar), each rear half is perspective-warped so its silhouette fills the slot; without it, plain cv2.resize.

The inputs you'll actually touch: the images batch (front, left, right, rear), the landmarks_batch, the four per-slot index overrides (front_index etc., -1 = auto) for when the classifier misassigns, and the strip/placement geometry (strip_width, strip_height, eye_target_y, face_height_ratio, feather_radius, and the rear knobs).

Outputs: strip_image, coverage_mask, inpaint_mask (same contract as the strip composer - feed it into the Qwen-fill subgraph), and status.

Installing

ComfyUI Manager → search "BrainDead" → install, or:

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

Restart. Requires OpenCV (cv2) and the MediaPipe landmarker bundle (auto-downloaded by BD Face Landmarks).

Where people get burned

The whole trick is alignment, so feed it photos where the subject's head is roughly the same size and pose across views - if the front is huge and the side is tiny, align_to_front can only do so much. And remember what this node is not: it won't fix perspective or wrap geometry around the head. If your side views don't overlap the front by a good third (left_center_x 0.68 / right_center_x 0.32 are the defaults that get that overlap), raise the feather and overlap before you blame Qwen.