arkennemasis fal · Flux Pro Kontext Multi · $0.04/image

The regular Kontext Pro node takes one reference image. This one takes up to six, numbered, and lets your instruction talk about them - "put the person from image 1 into the room from image 2, wearing the jacket from image 3". It's fal's own description that calls it experimental, so treat it that way, and it's priced like Pro at $0.04 per image.

When you want it

The obvious case is compositing references: subject from one shot, background from another, a prop from a third. The second case is consistency across a set - feed the same reference into the same socket for every generation in a run and you get the model's in-context identity handling applied to the whole series rather than a single picture. Kontext's structural advantage is that reference images go into the model's context instead of arriving as adapter embeddings, which is why it holds a subject better than the IP-Adapter school does.

The costs of that: six images is a lot of pixels for the model to reconcile, and cross-referencing instructions get less reliable the more sockets you fill. Two or three references with a clear division of labour works. Six competing ones gives you mush and a bill.

The six sockets

image_1 is required; image_2 through image_6 are optional. The tooltip for all of them is the same, and it's worth reading: each socket may carry a batch, and the pictures are sent in socket order, with the first picture of socket 1 counting as number 1. So "image 2" in your prompt refers to a position in the flattened list, not to a socket number - connect a batch of four to image_1 and your "image 3" is the third picture of that batch.

The rest of the node is Kontext Pro's shape: prompt (the instruction), aspect_ratio ((not set) by default, or 21:9 through 9:21), guidance_scale 3.5 (1–20), seed, num_images 1–4, output_format, enhance_prompt (off by default), and safety_tolerance default 2 (1 strictest, 6 most permissive).

The shared four sit at the bottom as they do on every fal node here: max_cost_usd (default 20, evaluated before upload), reuse_identical_run (on - same settings and same pixels returns the cached result rather than billing again), extra_json, max_concurrent (default 1 across the whole run).

Outputs

images, image_paths into output/fal/fal-ai_flux-pro_kontext_multi/, plus seed, prompt, has_nsfw_concepts, timings and info.

Install

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

Or ComfyUI-Manager → comfyui-arkennemasis. Restart after. The node lands under arkennemasis/fal/Image/Flux, and the key is one line in .env next to run_nvidia_gpu.bat:

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

Practical warnings

Numbering surprises. Because batches flatten into the list, the picture the model thinks is "number 2" may not be the one you meant. Keep batches out of these sockets unless you're deliberately indexing into them.

Unused sockets still count. A reference you connected but never referenced in the prompt still occupies a slot in the model's context and can bleed into the output. Disconnect what you don't need.

Experimental means experimental. fal labels this endpoint experimental, so its behaviour can change underneath you. Pin your expectations with seed and keep a copy of anything you rely on.

Paying per image, per take. Four images at $0.04 is 16 cents a queue click. Drop to one while you're working out which socket is which.