arkennemasis fal · Flux SRPO · $0.025/MP

SRPO is a preference-optimised Flux dev variant - the ["dev]" in the endpoint name is literal - tuned for aesthetics rather than instruction-following. On fal you get it as fal-ai/flux/srpo for $0.025 per megapixel, which lands a 1 MP image at two and a half cents.

Is it worth knowing about? Mildly, and honestly. The people running it locally report that the difference against the base is subtle - one of the few community write-ups on it is someone comparing SRPO Q5 at 8.4GB and Q8 at 12.7GB and concluding the gap is small. What they do report with more confidence is behaviour: it responds better to short prompts, one to three lines, and it leans painterly, handling artistic styles almost like a different model. There are also model-mix threads pairing it with Krea and other checkpoints for photographic output, which tells you where it sits - an aesthetics-first base that people blend rather than a workhorse.

So the honest framing: run it hosted if you don't want a 12 GB quant and 28 steps on your own card, and steal the prompting habit regardless of where you run it. If your Flux prompt is a paragraph of clause-stacked detail, SRPO is not the model to fix that with.

The inputs

• prompt - required and, per the above, shorter is better here.
• image_size - presets (landscape_4_3 default, square_hd, square, portrait presets) or custom with image_size_width/image_size_height. Billing is per output megapixel, so this is the price control: square_hd is 1 MP and two and a half cents, 1536×1536 rounds up to three megapixels and seven and a half cents.
• num_inference_steps - default 28, range 1–50. SRPO is not a step-distilled tier; the default is where it was tuned.
• guidance_scale - default 4.5, range 1–20. Note that's above the classic Flux 3.5, and the standing advice about Flux's same-face look - drop it toward 2 for people - applies here too.
• acceleration - none / regular / high.
• num_images (up to 4), seed, output_format, enable_safety_checker.
• Shared across the pack's fal nodes: max_cost_usd (default $20, checked before upload), reuse_identical_run, extra_json, max_concurrent.

Outputs

images (IMAGE - into SaveImage, an upscaler, an edit node), image_paths (the file as a string), seed as an INT so you can hold a run you liked, and prompt, timings, has_nsfw_concepts, info. The info output is JSON: fal's request id, the cost estimate, the files saved and fal's full answer.

Install

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

Manager equivalently: search comfyui-arkennemasis and install. Restart ComfyUI, then:

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

in the .env file beside run_nvidia_gpu.bat. The fal nodes in this pack install nothing - the wrapper is standard-library HTTP - so this is a genuinely two-minute setup, and there is no key box on the node because a typed key would be written into every workflow you share.

Notes from actually using it

If SRPO's output feels like base Flux, that's consistent with what the community says about it - treat it as a flavour dial, not a different model, and don't buy a workflow redesign on its account. Aesthetic tuning also tends to mean you should stop helping: the "ultra realistic, 8K, masterpiece" boilerplate that SD 1.5 needed actively hurts Flux-family output, and with an aesthetics-tuned variant you're fighting the tuning rather than helping it.

The safety checker returns black images instead of erroring when it flags something, and disabling it requires account authorisation on fal - unauthorised requests get checked regardless. Black output at this price is moderation, not breakage. reuse_identical_run (on by default) will hand back the cached file for identical inputs, including after a restart, so if you want a genuinely different take from the same settings, either change the seed or switch it off. And if ComfyUI restarts mid-job, the money isn't gone: requests are logged to output/fal/_requests.jsonl and the pack's free Recover Result node collects the finished job by id.