BD Parts Builder

Segmentation gives you a pile of masks; a character parts pipeline needs parts - each region cropped into its own image with a bbox, a tag, and a place in the stack. BD Parts Builder is the node that makes that jump. You feed it masks (SAM3 per-prompt masks or Sapiens2 labels) plus the source image, and it emits a PARTS_BUNDLE where every part has its own RGBA crop, bounding box, and label, ready for the pipeline's next stages: refine already happened upstream, batch-edit and underpaint run next, export finishes.

Two input styles, and the README is explicit about which to prefer: sapiens2_labels (from a Sapiens2 segmentation node, carrying its class taxonomy) or the generic masks + mask_labels pair (e.g. SAM3 Multi-Prompt output). When sapiens2_labels is wired, masks/mask_labels are ignored. A selection_mode preset lets you filter: all (everything but Background), clothing/bodyparts (curated Sapiens2-taxonomy subsets), or specific with your own selection_list. Note the presets only understand Sapiens2's 29 classes - for SAM3 sources, use all or specific.

Inputs that shape the parts

• padding_pixels (default 5) - breathing room around each bbox, in source resolution. Give a little padding; it saves you from razor-thin crops.
• output_size - native keeps each crop's natural size; 512/1024/2048 scales the longest side.
• square_pad - pad_to_square centers each crop on a square canvas (uniform aspect, ideal for upscaler batching and T3D); none keeps aspect.
• background - alpha (default) bakes the mask into the RGBA alpha channel; white/black/hex composites onto a solid.
• min_pixels (default 64) - drops parts below this mask area.

The optional combined_mask is important: wire SAM3's combined_mask (the authoritative silhouette after negatives/vote/color filter) and every part gets intersected with it before cropping - so the "noise SAM3 already ruled out" never makes it into a part. The tooltip's warning is worth repeating: don't wire an inverted mask here, or you'll silhouette the background and zero out every real part.

Wire depth_image (e.g. BD Lotus2 Predict's raw_linear, 0=near 1=far) and each part gets a depth_median - the median depth within its mask - which the compose/export stages use for back-to-front sorting. That's the field that makes the final composite stack in the right order.

Outputs

• parts - the PARTS_BUNDLE, the pipeline's primary output.
• image_batch - the crops as one IMAGE batch, for batch-only workflows like BD Bulk Save or upscaling.
• label_list, bbox_list - parallel strings for those batch paths.

Install

In ComfyUI-BrainDead: Manager search "BrainDead", or clone + pip install -r requirements.txt, restart. Under 🧠BrainDead/Segmentation; V3-API pack, so update ComfyUI if nodes don't show.

The failure modes are mostly upstream: if your parts are mislabeled, the problem is usually the mask/label alignment (labels must be parallel to the mask batch, one per line), and if every part comes out empty, check the combined_mask isn't inverted. This is also where the pack's SAM3 node earns its keep - BD SAM3 Multi-Prompt produces aligned per_prompt_masks + labels in one node, which is exactly the shape Parts Builder expects. Feed it that, wire combined_mask and depth_image, and the parts are ready for the heavy lifting.