arkennemasis fal · Flux Pro v1.1 Redux · $0.05/MP

Redux is BFL's image-variation adapter - it encodes a picture with SigLIP and conditions generation on that instead of (or alongside) text. This node runs it on the FLUX1.1 [pro] endpoint, the closed premium tier, at $0.05 per megapixel. Same idea as the Krea Redux node, but here the prompt box exists and can be left empty: the tooltip says so plainly.

What you're actually doing with it

Two patterns, and neither is editing. First, variation: feed it a picture, get back something recognisably of that picture - same subject, same palette, new composition. Second, style transfer: use the reference purely to carry a look into whatever the prompt describes, which is the use the community rated most highly for Redux in the local Flux Tools era.

If you want a targeted change - remove the hat, fix the hand - Redux is the wrong tool and always will be. It doesn't follow instructions about parts of an image; it absorbs the whole thing. The Kontext nodes are for that.

One number worth carrying over from the local Redux experience: in a ComfyUI graph where Redux and a prompt both feed the sampler, the community found anything much above a tiny weight makes Redux overwhelm the text and hand you a straight variation. That dial doesn't exist on this node - the tradeoff between image and text is decided by the endpoint - so if your prompt seems ignored, you're seeing the model's own weighting, not a setting you misconfigured.

Inputs

• image - the reference. Its tooltip mentions matching the dimensions of a mask, which is leftover schema text on fal's side; there is no mask socket on this node.
• prompt - optional. "Empty leaves it out" is the tooltip's own wording, so pure variation means leaving it blank.
• image_size, default landscape_4_3, with image_size_width / image_size_height for custom (16–14142, step 16).
• num_inference_steps 28 (1–50), guidance_scale 3.5 (1–20), seed (-1 random), num_images 1–4.
• output_format (jpeg/png), safety_tolerance default 2 (1 strictest, 6 most permissive), enhance_prompt off by default.

The shared four on every fal node here: max_cost_usd (default $20, blocks the run before upload, 0 = off), reuse_identical_run (on - identical pixels and settings come back from output/fal/ instead of billing again), extra_json, max_concurrent.

Outputs

images → preview/save; image_paths → output/fal/fal-ai_flux-pro_v1.1_redux/; seed, prompt, has_nsfw_concepts, timings, and info with the request id and cost estimate.

Install

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

ComfyUI-Manager alternative: search comfyui-arkennemasis. Restart ComfyUI, then find it under arkennemasis/fal/Image/Flux.

# .env, in the folder holding run_nvidia_gpu.bat
FAL_KEY=your-key-from-fal.ai/dashboard/keys

The honest caveats

$0.05/MP is the priciest per-megapixel node in this Flux batch and Redux is a variation tool, not a problem solver - it's easy to spend a surprising amount the first evening just seeing what it does with an image. Keep num_images at 1 while you explore.

Pro is closed weights. There's no local Redux-on-Pro equivalent, which is the whole reason an API node is the only route.

The output is a new image, not your image, recoloured. Unmasked pixels are not preserved - that's the difference between this and the masked inpaint nodes, and it's why you shouldn't use Redux where the original has to survive untouched.

Nothing happens on a re-run? That's reuse_identical_run. Nothing changed, so nothing was billed, and the same files came back.