⭐ Star QwenEdit+ Conditioner

Single-image editing with Qwen-Image-Edit is old news by now - one reference, one prompt, done. The interesting workflows are the multi-reference ones: here's the character, here's the outfit, now put them together. That's where the "+" models live, and that's what StarQwenEditPlusConditioner is built for.

It's the multi-image sibling of the pack's Qwen Edit Encoder. Where the encoder takes one image (or a ready latent) and produces conditioning, this node takes up to three reference images - image1, image2, image3 - plus a prompt, and hands you a single CONDITIONING output that carries all of them. Same idea as the single-image path: the editor gets to see multiple context images at once, so the edit can respect the character and the pose and the scene, not just whichever one you crammed into a prompt.

What you set

• clip - from your Qwen text encoder loader. Required, same as any conditioning node.
• prompt - the edit instruction, multiline. "Change the shirt to match the fabric in image 2" is the kind of phrasing this setup rewards.
• vae - used to encode the reference images to latents. Required for the image path.
• image1 (the primary reference), image2 and image3 (optional). Feed it one image and it behaves like a regular Qwen-Edit conditioner; feed it three and the edit model gets the full context.

Output is CONDITIONING, wired straight into your sampler's positive input.

Why you'd reach for it

The classic multi-reference use is identity + apparel: reference A is the person, reference B is the jacket you want them wearing, and the edit prompt says "put this person in that jacket, keep both exactly as shown." Character sheets, turnarounds, and "same character, new scene" workflows all lean on having more than one image in the conditioning, because a single reference forces the model to guess everything it wasn't told.

It also does the work the pack's other conditioning node does - the sizing and encoding bookkeeping so the conditioning lines up with the latent - so you don't have to hand-wire a VAE-encode per image yourself.

Install & gotchas

Pack story, same as always: ComfyUI_StarBetaNodes is the retired beta repo and these nodes now ship in ComfyUI_StarNodes - install that via ComfyUI Manager (search Starnodes) or git clone https://github.com/Starnodes2024/ComfyUI_StarNodes into custom_nodes, then restart.

Two things to respect. First, VRAM: every reference image is encoded and held in conditioning, and Qwen-Image-Edit is a 20B model - three references is meaningfully heavier than one, so a GGUF quant and a close eye on your memory are part of the job. Second, aspect: the model understands references best when they're roughly the same aspect ratio. Feeding it a 16:9 scene and a 1:1 face in the same conditioning invites the editor to get confused about what's what. Keep them consistent and the "+" earns its keep.