BD Qwen Character Edit

If you've been fighting character consistency the 2025 way - IP-Adapter, a face LoRA, a ControlNet, and a prayer - here's the 2026 answer: stop generating and start editing. Feed Qwen-Image-Edit one good image of your character plus a sentence, and it returns the same character in a new pose, outfit, or scene, because it's looking at the reference the whole time. That's the approach the community landed on, and BD Qwen Character Edit is the BrainDead pack's front door to it.

What it actually is

This is not a sampler and not a model loader. It's a conditioning encoder: you give it a clip, a prompt, and up to three reference images, and it hands you a CONDITIONING tensor ready for a Qwen-Image-Edit sampler. The "character consistency" isn't a separate trick - it's engineered into the system prompt it builds and the reference latents it attaches.

How it works

The node does three things in one pass. First, it builds the Qwen chat template: a system prompt (fully editable, see below) that tells the model which features are sacred - face shape, eye shape and spacing, hair, skin tone, tattoos - plus a user turn embedding each reference image as <|vision_start|><|image_pad|><|vision_end|> vision tokens. Second, it encodes your reference images through the VAE into latents and attaches them to the conditioning as reference_latents, so the diffusion model literally has your character's pixels in context. Third, it returns a latent (the first reference, VAE-encoded) so you can img2img if you want.

Those two toggles are worth knowing: enable_resize normalizes each reference to roughly a megapixel before VAE encode (keeps latency sane), and enable_vl_resize downscales to the ~384² the vision-language tokenizer expects. Leave both on unless you know why you're turning them off.

The inputs that matter

• clip - your Qwen-Image-Edit CLIP. Load it the same way you'd load it for any Qwen-Edit graph.
• prompt - the edit instruction ("change the outfit to a red jacket", "turn the character to face left").
• system_template - the multiline text box with the entire character-preservation doctrine in it. This is the real tuning knob. Want stricter identity? Strengthen the wording. Want it to allow clothing changes? Say so.
• vae - optional, but wire it: without it there are no reference latents, just the vision tokens.
• image, image2, image3 - up to three references. More isn't automatically better; one clean, high-quality shot beats three muddy ones.

Outputs: conditioning (into the sampler), image1/image2/image3 (the resized references, handy for preview), and latent.

Wiring it

Load Checkpoint / Qwen-Image-Edit → clip → BD Qwen Character Edit → conditioning → KSampler → VAE Decode
                                     vae ↗                                   latent ↗

The heavy lifting isn't this node - it's the Qwen-Image-Edit model file, which the pack does not ship. The conditioning it emits is only as good as the checkpoint behind the clip.

Installing it

ComfyUI Manager: search "BrainDead" and install; requirements.txt is handled automatically. Manual route:

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

Restart ComfyUI and the nodes appear under the 🧠BrainDead/Character category. This pack uses the newer ComfyUI V3 API, so keep ComfyUI itself reasonably current.

Common issues

The honest limit of instruction editing is structural: Qwen-Edit re-emits the whole frame, so on a long chain of edits the face can drift even with a locked reference. If you're doing 10-pose character sheets, expect to re-pin the identity reference occasionally, and bolt a mask back on when you need unmasked pixels to survive exactly. If results ignore your reference entirely, the usual suspects are a missing vae (no reference latents) or a mismatched CLIP/tokenizer for the Edit model - these nodes are picky about the checkpoint behind the wire.