Qwen Image 2 Edit (fal.ai)

Qwen-Image-Edit became the open community's default instruction editor back in late 2025 - the model you hand a picture and a sentence, and the thing happens. Qwen Image 2 folds that editing ability into a single 7B model, with native 2K output, and it's API-only. FAL_QwenImage2Edit is how you touch it from ComfyUI: up to three input images plus an instruction, sent to fal's fal-ai/qwen-image-2/edit endpoint, edited result returned as a normal IMAGE.

The headline is that this is the new architecture - the unified one the community keeps hoping will get open weights someday. Via fal you don't wait. The trade is the usual one for this pack: hosted, metered, no local compute, and the model's edit behavior is tuned by a prompt rather than diffusion dials.

How it works

Your connected image tensors are uploaded and passed as image_urls - the first is the subject, image2/image3 are optional references. The prompt drives the edit, the endpoint regenerates the frame, and the result downloads back as an IMAGE tensor.

The inputs that matter:

• image (required) and prompt (required) - everything else is optional.
• image2 / image3 - optional reference images for multi-image edits, up to three total.
• num_images - note this one goes to 6, unlike the v1 edit node's 4. If you want variations on an edit, this is the knob.
• enable_prompt_expansion (default true) - the model rewrites your edit instruction internally before acting. Leave it on for short instructions; turn it off when your wording is precise and you want it obeyed literally.
• image_size presets (square_hd default) or custom (512–2048px), negative_prompt, enable_safety_checker (default on), seed, output_format, private_mode.

One IMAGE output. Each image port takes the first frame of its tensor - no batch processing per port.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/elle-trudgett/fal-ai-comfyui-nodes
pip install -r fal-ai-comfyui-nodes/requirements.txt

or ComfyUI Manager → ComfyUI fal.ai Nodes. No model downloads. Set your key - export FAL_KEY="..." or a config.ini with [API] / FAL_KEY in the node folder for ComfyUI Desktop - from fal.ai/dashboard/keys, then restart.

Gotchas

• "FAL_KEY not set" is the standard first-run error; env var or config.ini, restart.
• No steps or guidance knobs - the 2.0 unified model doesn't expose them. Your prompt and the expansion toggle are the tuning surface.
• The full-frame re-emission caveat still applies. Like every generative editor, pixels you didn't mention can drift; chain enough edits and identity wobbles. The v1 Edit's drift work mostly carried over, but it's not a surgical tool.
• Metered and upload-heavy. Every connected image is uploaded and every generation billed; six variations of a three-image edit is a real charge.
• Errors come back as fal.ai error: ..., unwrapped for you.

If you liked the v1 editor but want the newer model's behavior and 2K output without running a 20B model, this is the version to reach for. If you need pixel-exact edits, keep a mask-based local workflow around too.