TextEncodeQwenImageEditPlus

The non-Plus Qwen edit encoder handles one reference image. This is the version for when one image isn't enough - you've got a character in shot A, a jacket in shot B, a background from shot C, and you want the model to reconcile all of them. TextEncodeQwenImageEditPlus is Qwen-Image-Edit's multi-reference encoder: up to three image inputs (image1, image2, image3), all described and embedded together with your edit instruction.

The mechanism is where it gets clever. For each image you connect, the node resizes it down to about 384×384 (the vision-LLM's per-image budget), appends a "Picture N:" vision token to the prompt, and runs the whole thing through a chat template that instructs the model to describe the key features of each image, then explain how your text instruction should modify them. That describe-then-instruct framing is what makes multi-reference work: the encoder doesn't just see three images, it reasons about what in each image is relevant to your edit. Each image is also VAE-encoded into a reference_latents list (at the higher ~1MP resolution, so the latents keep detail the vision tokens dropped). Output: one CONDITIONING carrying the merged text-vision understanding plus all reference latents.

Inputs: clip, prompt, and the optional vae + up to three image1–image3 slots. That's the whole surface. Connect only image1 and you've basically got the non-Plus node with a fancier prompt template - the Plus name earns itself when you actually use two or three.

It's core (comfy_extras/nodes_qwen.py), ships with ComfyUI, no install. Same checkpoint and loader family as the regular Qwen edit nodes - the Plus variant is a workflow difference, not a different model file.

The real-world notes matter here because this is the node people reach for with character-consistency workflows: three references of the same character from different angles and the model holds identity across an edit that a single-image encoder would drift on. But "consistency with the original where appropriate" is doing a lot of work - the template tells the model to keep what's relevant, so if your prompt doesn't tie the character's features to the edit, the references read as loose inspiration rather than constraints. And the honest limit from the community: the more you stuff in, the more the model re-emits - pixels you never mentioned still move. Keep the edit instruction specific about which reference carries what, and you'll get the good multi-ref behavior instead of a collage.