Qwen Image 2512 LoRA · Image Editing

The Qwen Edit LoRA library is the reason this node exists. Camera-angle changes, relighting, next-scene, expression swaps - a whole genre of small LoRAs got trained on Alibaba's editor, and the thing that separated them from every other base was the licence: Apache 2.0 carries none of the anti-circumvention terms that made NSFW LoRAs for Flux Kontext a violation, so people built openly. This node runs the 2511 revision of that editor with up to three of those LoRAs attached, on one to three source images.

It's also the drift-fixed revision. The 2511 card's own headline claims are "mitigate image drift" and "improved character consistency", which tells you what everyone complained about with the earlier versions. So if your mental model of Qwen Edit is "it re-emits the whole frame and my subject slowly mutates", this is the node where that improved.

The inputs

• prompt - say what to change. The tooltip tells you the house style directly: refer to your sources as Picture 1, Picture 2 and Picture 3. Model-default vagueness is what gets you a blend of everything you fed in.
• image_urls - one to three source images, one autogrow socket each. The result follows the first one, so order matters: your primary subject goes first.
• loras (optional) - up to three LoRA weight files from the SpicyAPI LoRA node, each with its own strength. This is where the camera-angle and relighting LoRAs go.
• aspect_ratio - defaults to model default, which means it follows the first source image. Leave it alone unless you want to change shape; that's a nicer default than the fixed 1:1 the 2.1 nodes use.
• output_format - jpeg, png or webp.
• seed - same seed and settings, same result. It ships fixed so a re-queue is a re-use, not a second charge.

Outputs

images and task id. No alpha channel on this one, so it's a straight edit-and-save node: Save Image, or a detailer if you're pushing the 2511 output further.

A LoRA for this node is not a LoRA for the others

Worth internalising before you download anything: the pose, angle and expression LoRAs people share for the Edit line are trained on the editor, and they don't transfer to the Qwen-Image base at all - a camera-angle LoRA is meaningless on a text-to-image node, because there's no input camera to move. Match the LoRA to the lineage.

Feeding it: SpicyAPI LoRA node, an https:// link that resolves straight to a .safetensors and needs no login, plus a strength. Chain up to three.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, add a key from spicyapi.ai/console/keys in Settings > SpicyAPI, restart again so the catalogue comes from your account. No GPU, no downloads, no extra Python packages. SPICY_API_KEY for servers.

Practical notes

About one megapixel by default. If you leave aspect_ratio on model default, the result matches the first source image's shape at roughly 1MP. Bigger source images don't come back bigger, so if you need more, plan an upscale pass rather than assuming the resolution travels.

Three images is the limit, not a target. Each source is another constraint, and prompts that reference them positionally work; prompts that don't produce a composite of everyone.

Per image, around $0.03, and edits come in passes. You'll run this three or four times per image to get a set. Set Max cost per run in the SpicyAPI settings - the node quotes before submitting and blocks anything above the ceiling with nothing charged. Remember that cancelling only stops you waiting: the job completes server-side and is billed if it succeeds. Keep the task id when a run fails; 40310 is an unverified email and 40004 is a parameter combination no deployment serves.