BD Qwen Multi-Image

BrainDead ships four Qwen-Image encoders that look nearly identical and behave differently enough to matter. BD Qwen Character Edit is the generalist, BD Qwen Identity Lock is the face-first lockbox, and this one - BD Qwen Multi-Image - sits in the middle: a multi-reference encoder with a weight_mode control the others don't have. If you keep reaching for a node that just handles "a few reference images and an edit instruction" without the identity hard-sell, this is probably it.

What it is

Same job as its siblings: prompt + up to three reference images in, a CONDITIONING tensor out, ready for a Qwen-Image-Edit sampler. The default template sets the division of labor: Picture 1 is the primary reference for character identity (face structure, hair, skin tone, build), Picture 2/3 are style/pose/environment/lighting. But where Identity Lock demands Picture 1 survive verbatim, Multi-Image's template is a softer "combine elements while maintaining character consistency" - and you get a weight_mode combo to control how the references are blended.

The one control that's special

weight_mode (default diffusers) is the differentiator. It controls how the reference images are weighted when they're folded into the conditioning. The other Qwen encoders don't expose this at all - they just attach reference latents and let the template do the talking. Here you get to pick the blending behavior, which is useful when one reference is a clean identity shot and another is a mood board you want some influence from but not equal weight.

The rest of the inputs

Standard for the pack: clip (your Qwen-Image-Edit CLIP), prompt (the edit instruction), system_template (editable, and it's where you adjust the identity-vs-style balance if the default doesn't suit), optional vae (wire it - no vae, no reference latents), image / image2 / image3, plus the two resize toggles enable_resize and enable_vl_resize. Outputs: conditioning, the three resized references, and latent (VAE-encoded first reference for img2img).

How to think about the family

Pick by temperament, not by spec sheet:

• Character Edit - balanced default, good all-rounder for character work.
• Identity Lock - the face is non-negotiable; weighted strengths default above 1.0.
• Multi-Image - flexible blending via weight_mode, best when your references are doing different jobs and you want to dial their relative pull.

Installing it

ComfyUI Manager: search "BrainDead" and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/BizaNator/ComfyUI-BrainDead
cd ComfyUI-BrainDead
pip install -r requirements.txt

Restart ComfyUI; the node lives under 🧠BrainDead/Character. Remember the pack doesn't ship Qwen-Image-Edit weights - load the model as you normally would and feed its clip and vae in. And since this is an instruction-editing encoder, expect the usual behavior: it re-emits the whole frame, so on a chain of edits the face can drift even with references wired. Re-pin the identity reference when it does.