BD MP Mouth Parts

If you've ever rigged a mouth for animation, you know the viseme atlas workflow: separate the mouth render into lips, teeth, and tongue, then pack them into an RGBA texture where the game engine reads each from its own channel. Doing that used to mean a five-node chain - SAM3 ×3 for the three parts, fill holes, resolve, pack channels. BD MP Mouth Parts collapses all of it into one node: it separates an isolated mouth render into lips / teeth / tongue and packs them straight into the Unreal viseme-atlas contract - R = lips, G = teeth, B = tongue, A = POM (parallax occlusion, if you wire a depth map in).

The clever part is the default engine, color, needs no model at all. It's pure HSV classification tuned on the pack's own PinkF1 viseme set: the mouth render comes in on a near-black background with magenta/purple lips, white teeth, and a salmon/rose tongue - a clean color-separation problem. Teeth are bright and desaturated (low saturation, high value); the tongue is the warm-hue region (closer to red, ~338°) gated to the eroded mouth interior so warm lip-corner highlights don't read as tongue; lips are the saturated remainder, with a saturation floor that keeps the dark mouth cavity out.

The engine choice

engine defaults to color, but flip it to sam3 when color alone fails - e.g. the lit front of the tongue reads as saturated magenta and bleeds into the lips. In sam3 mode, the color split becomes the prior: each part's color mask seeds SAM3 (bounding box + interior positive points + the other parts' centroids as negatives), SAM3 grows the true object boundary, and the result gets clipped to the mouth foreground and cleaned. If you leave model unwired, it auto-loads the official SAM3 checkpoint in-house.

Then there's the threshold dial cluster for the color engine - these are tuned for the PinkF1 palette and exposed so you can adapt to other renders:

• bg_v_min (25) - background cutoff; raise if the render has a lifted grey background.
• teeth_s_max (60) / teeth_v_min (140) - teeth = desaturated and bright.
• tongue_h_lo/hi (325/358), tongue_s_min/max (80/170), tongue_v_min (120) - the warm tongue band.
• lips_s_min (85) - the cavity guard.
• interior_frac (0.06) - the tongue must lie inside the mouth interior eroded by this fraction; kills false tongues from warm highlights.

The cleanup trio - edge_smooth, fill_holes, despeckle - seals jagged edges, fills interior holes, and drops tiny components (specular glints that read as teeth). And edge_refine (off → guided → matting → vitmatte) snaps part edges to the image, each degrading gracefully to 'off' if its backend is missing.

Outputs and install

You get lips, teeth, tongue as MASKs; rgb_packed (R/G/B, no alpha); rgba_packed (R/G/B/A with POM - the viseme cell contract); a debug_overlay tinted by part for QC; and a status string. Wire pom with a depth map (e.g. Lotus2 depth → normalize → center luma) and it rides the A channel; leave it and A = the mouth foreground, so the packed image is still usable.

It's in ComfyUI-BrainDead: Manager search "BrainDead", or clone + pip install -r requirements.txt, restart (V3-API pack - keep ComfyUI current). It needs cv2 (numpy/opencv) and, in sam3 mode, auto-downloads the SAM3 checkpoint.

The honest word: it's tuned on a specific render style, so if your mouth renders use a completely different palette, expect to touch the HSV thresholds (the debug_overlay makes that painless). And it's a separation and packing node only - saving to disk and POM derivation live in their own nodes. But for its lane - PinkF1-style visemes straight into a Unreal atlas - it replaces a six-node chain with one node and no model, which is exactly the kind of thing this pack does well.