BD Parts Batch Edit (Qwen)

Here's the problem the BrainDead parts pipeline solves: SAM3 segments a character into parts, but the masks only show the visible pixels. The pants under the holster, the shirt under the jacket - those parts are incomplete. To rebuild them you'd normally loop part-by-part through Qwen Image Edit manually, and on a ten-part character that's ten separate babysitting sessions. BD Parts Batch Edit (Qwen) does the whole loop inside one node: every part in a PARTS_BUNDLE gets its own Qwen Image Edit pass in a single execution, and the bundle comes out with every part rebuilt.

The mechanism is worth understanding because it explains the defaults. Each part's RGBA is composited onto neutral gray, and its prompt is built by dropping the part's tag into prompt_template (the default is tuned for low-poly stylized characters: "Recreate the complete {tag} in the same low-poly game style, sharp clean edges, fill any occluded or missing regions..."). The result is encoded, sampled, and decoded with the original alpha reapplied so edges stay clean.

The two modes

• flatten_redraw (default) - flattens the part on a white background, encodes it, latent-upscales by latent_upscale_factor (1.25 default, the high-res-fix trick), samples, tonemaps, and decodes. Produces a clean object render, then cuts it to the part outline so the composite assembles without white backgrounds.
• true_inpaint - latent-space inpaint with a prefilled source crop; the noise mask covers only enclosed holes, so Qwen regenerates just the missing regions and preserves what's visible. Best for completing obstructed parts. Requires source_image wired so Qwen sees full context (pants with the gun on top).

This node is built around the Qwen Image Edit 2509 fp8 stack with the 4-step Lightning LoRA - the defaults match that setup exactly: steps=4, cfg=1.0, sampler=euler, scheduler=simple. It also applies the internal patches (shift=3.0, CFGNorm 0.85, Reinhard tonemap 2.0) that the manual recipe uses. If you're on a different stack, adjust accordingly; model_sampling_shift and cfg_norm_strength at 0.0 disable the internal patches if you've already wired them externally.

Inputs that matter most

• model, clip, vae - wire your Qwen Image Edit model stack.
• prompt_template - where {tag} gets interpolated per part. Change it per art style.
• target_pixels (default 1024x1024) - working resolution for VAE encode; 1536² is Qwen's max native and ~1.4× slower.
• alpha_after_edit (default auto_from_white_bg) - detects white as transparent in Qwen's output and keeps the full generated shape (no re-cropping to the SAM3 mask, which would re-introduce holes). The other options (fill_holes, original_part, etc.) trade that for different crop behaviors.
• skip_tags / min_pixels - skip parts by name or tiny crops.

Outputs: parts (the edited bundle), a summary, and image_batch of the rebuilt crops.

Install & the honest caveats

In ComfyUI-BrainDead: Manager search "BrainDead", or clone + pip install -r requirements.txt. This node needs the Qwen Image Edit model (a 20B MMDiT - big, and fp8 + Lightning LoRA are there for a reason). It's a V3-API pack, keep ComfyUI current.

Where people get burned: the internal patches are load-bearing. Disable model_sampling_shift without having ModelSamplingAuraFlow wired externally and the Lightning noise schedule doesn't match the upscaled latent - you get noise stripes instead of a clean render. And if a part's crop is tiny, the node skips it (below min_pixels) and keeps the original, which is correct behavior that reads as "nothing happened" the first time. The true_inpaint mode is the one to reach for when occlusion matters (pants under holster); flatten_redraw is the general-purpose rebuild.