BD Face Strip Compose

When you need a head-wrap texture that's geometrically correct - a true cylindrical unwrap where the side of the head maps to the side of the strip - BD Face Strip Compose is the node. It stitches your 4 head-view photos (front, left, right, rear) into a single horizontal strip via a landmark-anchored cylindrical unwrap: each detected view gets TPS-warped (thin-plate spline) so its 478 MediaPipe landmarks land in strip space, a cosine-bell confidence keeps each side dominant in its own slice, and the rear photo gets split and slotted onto the strip edges.

This is the "I need it to actually unwrap correctly" option, versus BD Face Mosaic Compose, which is the faster "good enough for Qwen to finish" stitch. If your result just needs a cleanup pass from Qwen, use the mosaic; if you need a real longitude/latitude unwrap, this is the one.

What comes out

Three outputs and a diagnostic:

• strip_image - the stitched composite (default 4096×1024).
• coverage_mask - where the strip has any data at all.
• inpaint_mask - where Qwen should paint: low-coverage areas plus the seams between the rear slots and the adjacent side views. This is the key output for the Qwen-fill subgraph that fixes what the photos couldn't cover.
• status - per-view diagnostic, so you can see which view landed where.

The inputs are mostly slot assignment and geometry. front_index / left_index / right_index / rear_index (all -1 = auto) let you force which photo goes to which slot when MediaPipe's yaw classifier misfires. strip_width / strip_height set the canvas. Rear handling: rear_split_x (where to split the rear photo), rear_extent (fraction of strip per half), rear_confidence, rear_blend_band, rear_flip_lr, rear_seam_width (the inpaint-flagged band on each side of the rear slots - set 0 to disable), and rear_mask (a head silhouette from e.g. BD_SAM3MultiPrompt that gets perspective-warped to fill the slot). bell_sigma controls how wide each view's confidence bell is (0.22 keeps each side dominant in its own slice), and canonical_mode defaults to cylindrical - the atan2(x,z)/latitude projection that makes the strip a real unwrap.

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. Needs the mediapipe model bundle (auto-downloaded by BD Face Landmarks) and scipy for the TPS warp.

Where people get burned

The rear seam is where this node lives or dies - if you don't provide a rear_mask, the rear halves get a plain cv2.resize and the seam shows; feed it the head silhouette and the warp fills the slot properly. And don't skip the inpaint_mask: the strip will have low-coverage gaps, and that mask is the whole mechanism for getting Qwen to fix them. Force the slot indices if the classifier misassigns - that's what they're there for.