arkennemasis fal · Flux 2 Flex · $0.05/MP

At $0.05 per megapixel, Flex is four times the price of the plain Flux 2 node - and rounded up to the nearest megapixel, so even a 512×512 test renders as a full dollar-cent amount. That's not a bargain, and it isn't meant to be. Flex is the tier where Black Forest Labs exposes the control knobs and pushes typography: real inference steps you can raise, a guidance value you can dial, and a safety tolerance slider. If you want the model to render a poster that says exactly what you typed, this is the endpoint that's built for it.

What's different from the other Flux 2 nodes

The full Flux 2 node gives you steps and an acceleration dial but no safety tolerance. Flash gives you neither. Flex gives you num_inference_steps, guidance_scale and safety_tolerance on a 1-5 scale - and it's the branch of the family whose pitch is text rendering: signs, labels, packaging, typographic layouts.

It also returns fewer outputs than its siblings. There's no timing blob and no NSFW-concepts field on this node - just the images, their paths, the seed and the info JSON. Nothing lost that matters, but don't go looking for a port that isn't there.

The inputs

prompt - sentences. This model's text encoder reads your prompt as an instruction, not a bag of tags, and prompt weighting syntax is a no-op. For typographic work, say the words you want on the sign explicitly and in order.

num_inference_steps - 28. Fal bills per megapixel, not per step, so raising this costs you seconds, not dollars. On a $0.05/MP endpoint, that's a good deal and you should use it.

guidance_scale - 3.5 here, higher than the 2.5 the plain Flux 2 node ships with.

safety_tolerance - 2 by default, 1 strictest through 5 most permissive. Separate from enable_safety_checker, which is the blunt on/off. If a legitimate render keeps coming back refused and you have the account authorisation to loosen things, this is the dial that actually moves.

image_size - presets with landscape_4_3 as the default, or custom plus image_size_width / image_size_height.

output_format - defaults to jpeg here, not png. Change it if you're going to keep editing the result.

seed (-1), plus the pack's usual max_cost_usd, reuse_identical_run, extra_json and max_concurrent.

Outputs

images is the IMAGE output - batch if num_images is above 1. image_paths is a STRING giving you where they were written, seed is an INT you can lock, and info is a JSON string holding the request id, the cost estimate, the saved files and fal's raw answer.

Install

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

Restart ComfyUI; the node appears under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL does the same. The pack pins one dependency, httpx, for its ChatGPT/Codex nodes - fal needs nothing beyond the standard library, so your Python environment is untouched.

Then, in the .env alongside run_nvidia_gpu.bat:

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

Where people get burned

The rounding. Billing rounds up to the nearest megapixel on both sides, so a 0.26MP test render costs the same as a 1MP one. Do your prompt-dialling on the cheap nodes; spend Flex's money on the render you're actually shipping.

Expecting CFG behaviour. Do not port your SDXL habits over. Guidance here is in the 3-ish range, not 7, and on this flow-matching family wild CFG values don't sharpen adherence so much as cook the image.

Assuming more references means more cost - that's Flex Edit's territory, and there the input side is capped at 1MP regardless of how many pictures you attach.

Repeated runs returning identical files. reuse_identical_run defaults on, so an identical rerun is served from output/fal instead of billed. That's your safety net against paying twice; flip it off when you want a new take from the same prompt. And if a run stops before anything uploads, check max_cost_usd - the estimate is compared to the cap before the paid call, so tripping it is free.