Fal Image-to-Image

Style transfer, product editing, relighting, Flux-fill-style inpainting - the fal image-to-image catalog is enormous, and Fal Image-to-Image is the single node that opens it all. It's one of eleven nodes in the ComfyUI-Fal-Gateway pack, and if you've ever stared at a "one Python class per model" fal wrapper and watched the dropdown never grow, this is the anti-that. The model list is pulled live from fal's OpenAPI catalog, so the ~400 I2I endpoints show up as one type-ahead dropdown, newest releases included.

Inputs and outputs

Three required inputs, and the one you'll actually think about is the image:

• image - wire a LoadImage (or anything producing an IMAGE tensor) into this. That's your source frame.
• model_id - the dropdown. Families like [bria] Fibo Edit, [alibaba] Qwen Image 3 Image Editing, plus the FLF-capable models that double as plain I2I when you only wire a start image.
• prompt - multiline. Leave it empty for models that are pure edits; the node only sends a prompt you actually type, and some endpoints don't need one at all.

Outputs are the standard image triple: image (a normal IMAGE tensor - wire to SaveImage), image_url (the fal-hosted result URL), and info (a JSON string with the full result dict, timings and all).

How it works

Your source tensor gets uploaded to fal's CDN, the job runs against the endpoint you picked, and the result is downloaded and decoded into a tensor you can keep piping around the graph. The "edit" models on fal are mostly instruction-tuned, which is worth internalizing: with something like Qwen-Image-Edit, the prompt is the edit. "Make the background snowy", not "snow, winter, landscape, 8k". That's a genuinely different muscle from local img2img at low denoise, and people who treat these like SD img2img get muddy results and blame the node.

Install

Same story as every node in this pack - not in the Manager registry yet, so clone it in:

cd ComfyUI/custom_nodes
git clone https://github.com/modbender/ComfyUI-Fal-Gateway
pip install -r requirements.txt   # into your ComfyUI venv
export FAL_KEY=your_key           # or fill config.ini

Restart ComfyUI and the whole pack appears under Fal-Gateway. The dependency list is modest - fal-client, aiohttp, opencv-python, Pillow, numpy, pydantic - nothing exotic, no giant model downloads.

Gotchas

• Every call costs money, and editing models bill per job. The title-bar cost badge (green → red) is your friend; glance at it before you queue a batch of 40 relight variations.
• Fresh nodes sometimes show <no models available> for a few seconds while the catalog fetch runs. Right-click → "Fal-Gateway: refresh catalog cache" if it's stuck, and restart to pick up brand-new models.
• No key means no run - the node raises "FAL_KEY not set" with the fix in the message. Get one at fal.ai/dashboard/keys; it's the same key for every node in the pack.

One honest take: for a single sharp edit on an image you already like, local instruction-editing (Qwen-Edit class) is cheaper per iteration. Where this node wins is breadth - every closed or huge editing model at once, from a machine that never had to hold one.