arkennemasis fal · Qwen Image 3 Edit · $0.04-0.075/image

Here's the annoying part of the local editing stack: Qwen-Image-Edit is the open instruction editor everyone actually uses, and it's frozen at 2511 - qwen.md is blunt about it, "the successor went closed", with Qwen-Image-2.0 shipping API-only. So the newest Qwen editing capability exists, and it's a call, not a download. This node is that call.

ArkFal_alibaba_qwen_image_3_edit edits an image from one to three reference pictures plus a natural-language instruction, preserving identity and detail while applying the change. It's $0.04 per image at 1K, $0.075 at 2K, and it lives in the fal branch of the pack.

When it beats doing it locally

Locally, Qwen-Image-Edit 2511 remains the pick for anything where a face has to survive the edit - the KB calls that its clearest remaining advantage. So this node isn't a replacement; it's the escape hatch. You reach for it when the GPU is tied up, when you want six variants without a six-minute render each, or when you're on a laptop and it's the difference between editing the picture and not.

You should also know what you're buying into, and that's the honest frame from external-api-nodes.md: your reference images and prompt leave the machine, the provider's moderation applies and can't be patched out, and in this specific pack's category a node that holds a key and calls the network by design is exactly the shape of thing you should look at once before its first run. This is a new pack with a large surface - that's a reason to read it, not a reason to avoid it.

What you actually set

• prompt - the instruction. Up to 5,000 characters, Chinese or English. You refer to the references as "image 1", "image 2", "image 3", so ordering is meaningful.
• image_1 - required. The model takes one to three references, so image_2 and image_3 are the ones you'll actually use; the extra optional sockets exist because each socket can carry a whole batch, sent through in socket order. Keep it at three references. Files want to be 384–2048px per side, max 10MB each, and JPEG/PNG/WebP with no alpha - transparency is flattened to RGB before upload.
• image_size - square_hd by default. Pick one of fal's presets, or custom and use image_size_width / image_size_height (16–14142 in steps of 16, total pixels between 512² and 2048²).
• num_images - 1 to 6, so you can generate a set of takes in one call. Cost scales with it.
• enable_prompt_expansion - on by default, and it runs your prompt through an LLM first. That's usually a win for vague prompts and a loss for surgical ones. If you wrote "change the jacket to red, keep everything else", turn it off; the rewriter loves to add detail you didn't ask for.
• negative_prompt, seed (-1 lets the model pick), enable_safety_checker (on; disabling needs account authorization and unauthorized requests are checked anyway), output_format.

Plus the shared fal controls: max_cost_usd (a $20 default cap that stops the run before anything is uploaded), reuse_identical_run (on - the same inputs return the already-paid result, even after a restart), extra_json and max_concurrent.

Outputs are images (an IMAGE batch), image_paths, the seed fal used, and info with the request id, cost estimate and fal's full answer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart ComfyUI (or use Manager and search arkennemasis), then:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

into the .env next to run_nvidia_gpu.bat. No key widget on the node, by design. No models to download either - fal holds the weights, which is the whole point.

Small print worth reading

The cost estimate this node shows is computed at the 2K rate, so a 1K run gets quoted at nearly double what fal will actually charge. Don't panic at the badge, and don't set max_cost_usd so tight that it blocks an honest run.

Results go to output/fal/alibaba_qwen-image-3_edit/. If a batch comes back at mixed sizes, the IMAGE output carries only the pictures matching the first one - all of them are still saved to disk, which is the pack quietly doing the right thing.

Last thing: the request is legal, and the outcome isn't guaranteed. Instruction editors drift identity, lighting and colour a little on every pass; if the face matters, that's still Qwen-Image-Edit 2511's home turf.