BD Parts Builder 2 (Qwen 2.1)

Decompose a character head into asset layers with Qwen Image 2.1 prompts - no SAM, frame-true. For every item the presence vote found: VISIBLE - 'extract only the X exactly where it is' (RGBA). Trusted for WHERE, not for pixels: each source pixel goes to the faithful edit that matches it, and the layer is the SOURCE's own pixels, so the visible layers rebuild the source exactly. Items whose edit does not match the source (hidden, or restyled) own nothing and keep only a complete layer. COMPLETE - 'decompose the head into layers, output only the X layer': the whole object, hidden parts included. Split into front and _back (pixels over bare skin are behind the head). Paint order comes from the source: on two complete layers' overlap, the one the source shows is in front (depth only breaks ties). depth_median in both bundles encodes that order for BD_PartsExport. base_image = head without items (the voted accessories removed), frame-proven with seed retries; feed it to BD_PartsExport base_image with parts to get a PSD that rebuilds the source.

By BizaNator·Created 8 months ago·Updated about 13 hours ago· 15

image
model
clip
vae
head_mask
depth_image

parts
complete_parts
base_image
layers_preview
report

◄presence►

◄seed20260931►

◄steps40►

◄cfg1.0►

◄sampler_nameeuler►

◄schedulersimple►

◄make_completetrue►

◄make_basetrue►

◄colour_retrytrue►

◄base_attempts3►

◄fidelity_min0.50►

◄explained_max60►

Category🧠BrainDead/Segmentation

Inputs (19)

Name	Type	Default	Description
image	IMAGE		Headshot (e.g. HeadMasters MASTER 1). Its RGB is used as given.
model	MODEL		Qwen Image 2.1 through QwenImage21Cache.
clip	CLIP		qwen3vl_8b, CLIPLoader type qwen_image.
vae	VAE		Qwen Image 2.1 VAE (RGBA).
presence	STRING		The vision model's JSON answer (AILab_QwenVL with BD Parts Vocabulary's vlm_prompt), or a typed list: 'hair, eyes, mouth, hat'.
vocabulary	STRING	hair \| hair \| hair \| no eyebrows \| eyebrows \| eyebrows \| no eyes \| eyes \| eyes \| no mouth \| mouth \| mouth and lips \| no beard \| beard \| beard \| no moustache \| moustache \| moustache \| no glasses \| glasses \| glasses \| yes sunglasses \| sunglasses \| sunglasses \| yes hat \| hat or cap \| hat \| yes helmet \| helmet \| helmet \| yes headband \| headband or bandana \| headband or bandana \| yes headphones \| headphones \| headphones \| yes earrings \| earrings \| earrings \| yes piercings \| piercings \| piercings \| yes	From BD Parts Vocabulary. key \| ask \| noun \| accessory.
seed	INT	202609310–18446744073709550000	Every edit uses this seed; a frame-proof retry uses seed+1, seed+2 ...
steps	INT	401–200	Qwen Image 2.1 model card: 40 for edits.
cfg	FLOAT	1.00–20	—
sampler_name	COMBO	euler	44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
scheduler	COMBO	simple	9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3
make_complete	BOOLEAN	true	Also make the COMPLETE layers.
make_base	BOOLEAN	true	Also make the head without items.
colour_retry	BOOLEAN	true	One retry with the measured colour named ('the white sunglasses') for an ACCESSORY whose edit does not match the source. Never for hair/brows/eyes/mouth: next to an accessory their measured colour is the accessory's.
base_attempts	INT	31–8	Head without items: seeds tried (seed, seed+1, ...) until it is frame-true.
fidelity_min	FLOAT	0.500–1	Share of an edit's pixels that must match the source (< 40 levels) for it to own pixels.
explained_max	FLOAT	601–255	A pixel is explained by an edit when their colours are within this many levels.
head_maskopt	MASK		Where the head is (1 = head), e.g. the image's alpha. None: non-white pixels.
depth_imageopt	IMAGE		BD Lotus2 Predict map or raw_linear - BRIGHTER = NEARER. Breaks paint-order ties only.

Outputs (5)

Name	Type	Description
parts	PARTS_BUNDLE	—
complete_parts	PARTS_BUNDLE	—
base_image	IMAGE	—
layers_preview	IMAGE	—
report	STRING	—