arkennemasis fal · Flux Pro v1.1 Ultra Redux · ~$0.21/image

Redux is one of Black Forest Labs' Flux Tools, and it does the thing that's hard to describe and immediately useful once you've seen it: it reads a reference image and produces variations of it. Style transfer, angle changes, mood shifts, "give me this again at golden hour". You don't describe the subject in the prompt - you hand over a photo and let the image do the talking.

Local Redux exists as a Dev-tier model, so this node isn't your only route to the concept. What it buys you is the Pro base doing the work: the undistilled tier with better facial variety and no 12GB download, at roughly twenty-one cents an image. That price is the honest headline here - it's the expensive one in this family, roughly three and a half times Ultra's plain text-to-image rate, and it's the one you shouldn't be firing off in batches while you fiddle.

How the mechanism works

This is the fal-ai/flux-pro/v1.1-ultra/redux endpoint. You connect a picture, fal takes it as the reference, and the model generates in that image's visual territory. The prompt is optional here - genuinely optional, not "optional but please fill it in". Empty leaves it out entirely, which is exactly right for pure variation work; it's what you'd use when you want the model's read of the reference and not yours.

The pack itself is a wrapper, and the whole fal family shares its plumbing: check inputs, price the job, upload the reference to fal's CDN, submit one billed call, poll, download into output/fal/fal-ai_flux-pro_v1.1-ultra_redux/. The fal half of the pack uses Python's own urllib and nothing else, so installing it doesn't touch your ComfyUI environment.

Inputs worth touching

• image - required, and it's the point of the node. One picture, not a batch; a batch gets rejected with a message telling you to pick a frame first.
• image_prompt_strength - 0 to 1, default 0.1. The name is inherited from Ultra's image-prompt input, and it's loose. Given how the endpoint behaves, treat it as the reference's pull; if variations feel too shy or too mush, this is the dial before anything else.
• aspect_ratio - default 16:9. Since Redux is billed by output megapixels underneath, shape and size here have more cost consequence than they do on Ultra.
• prompt - leave empty for a straight variation. Fill it in when you're steering ("same scene, night, rain").
• num_images - up to 4. At this price, think before setting it to 4.
• seed, raw, enhance_prompt, safety_tolerance, output_format - as on every Flux Pro node in the pack.

The price, honestly

The node's title says ~$0.21/image and the estimate behind the money guard is a flat $0.21 per image. fal's own billing underneath is $0.05 per megapixel, so that tilde is doing real work: a 2K-ish output is what lands you around twenty-one cents, and a smaller one costs less. The badge and the pre-run estimate are deliberately conservative rather than exact - the info output on the node returns the real estimate that was used, which is the number to look at when you're auditing a batch.

Money guards, all shared across this pack's fal nodes: max_cost_usd (default $20) stops the run before anything is uploaded; reuse_identical_run (default on) gives back the cached result for an identical rerun instead of charging you twice; max_concurrent (default 1) keeps paid fal calls from stampeding fal's queue in one run.

Outputs

images and image_paths are the two you'll wire - the IMAGE socket drops straight into SaveImage, an upscaler, or a video node. seed comes back as an integer so you can lock a variation you liked and re-roll around it. prompt, timings, has_nsfw_concepts and info round it out; info is the JSON with the request id, cost estimate, saved files and fal's full answer.

Install

ComfyUI Manager → search comfyui-arkennemasis → install. Or the manual route:

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

Restart ComfyUI, then add your key to the .env file next to run_nvidia_gpu.bat:

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

That's the entire setup. No models to download, no CUDA wheels, no restart dance beyond the one.

Where people get burned

Two of them. First, cost: Redux is the node in this family where an idle click of Run is a real twenty cents, so don't leave num_images at 4 while you're exploring - set it to 1, find the variation style you want, then batch. Second, expectations: "Redux" reads like an upscaler or a 1:1 edit to a lot of people, and it isn't either. It generates a new image informed by the reference. If you wanted the same pixels with one thing changed, that's Fill or Erase, not this.

And the usual fal-family friction: no key means no run (with a clear message and no charge), an account with an exhausted balance just stops generating, and if ComfyUI restarts mid-job the paid request isn't lost - it's in output/fal/_requests.jsonl and the pack's free Recover Result node collects it by id for nothing.