arkennemasis fal · Flux Krea · $0.025/MP

FLUX.1 Krea [dev] is the checkpoint Black Forest Labs shipped in July 2025 with Krea's aesthetic tuning baked in. The pitch was simple: plain Flux Dev looks plasticky at default settings, and this one doesn't. If you don't want to spend the disk and VRAM, this node runs that model on fal.ai for $0.025 per megapixel and drops the result into your graph as a normal IMAGE.

What you're actually buying

A text-to-image call. No model files, no VRAM, no local install of anything 12B. The tradeoff is the usual API tradeoff, and it's worth being honest about it: your prompt and any reference pixels leave your machine, and every run is metered. In exchange you get the Krea look in a graph that also contains your local upscaler and your masking.

The honest caveat on the model itself: it improves realism over Flux Dev, and it also runs slightly warm and slightly soft. Some read that as "filmic" rather than "CGI"; others bounce off the yellow cast and go back to Dev with a photorealism LoRA. Try it before committing a client project to it.

The inputs you actually set

prompt is the obvious one, and Flux-lineage models want natural sentences, not tag soup. image_size picks the shape; leave it on one of the presets and only reach for custom when you have a specific target, which then reads image_size_width and image_size_height (16–14142, step 16).

Three settings do most of the work:

• num_inference_steps, default 28. Range is 1–50 and there's no reason to go above the high twenties on this model.
• guidance_scale, default 4.5, range 1–20. This is your adherence dial. Nudge it up when the prompt keeps getting ignored, down when the image looks fried.
• acceleration - none / regular / high. Free speed, some quality cost. high is a reasonable first try when you're iterating on a prompt.

seed at -1 lets fal pick. Set a real number once you have a composition you like, because the node hands the seed back on its seed output - wire that somewhere or write it down, and the same prompt and seed will reproduce the image.

enable_safety_checker defaults on. Leave it on unless you know your fal account is authorized for it, because unflagged requests are checked anyway and anything flagged comes back as a black image. output_format is jpeg or png.

The four widgets at the bottom are the same on every fal node in this pack, and two of them matter: max_cost_usd (default 20) stops the run before anything is sent if the estimate is over it - set it to 0 for no cap. reuse_identical_run (on by default) will hand you back the files from an identical earlier run in output/fal/ instead of paying again. That's great for re-runs after a UI crash and terrible when you want a fresh take with the same settings - turn it off.

Outputs

images is the one you wire into Preview Image / Save Image. image_paths gives you the saved file locations under output/fal/fal-ai_flux_krea/, timings and info are diagnostics, and has_nsfw_concepts tells you whether the safety checker flagged anything.

Installing it

The pack goes in like any other:

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

Or in ComfyUI-Manager: search comfyui-arkennemasis (or Install via Git URL with the repo URL) - Manager handles the dependencies. If you install by hand, the only requirement requirements.txt actually declares is httpx, and that's for the pack's separate Codex/ChatGPT nodes; the fal provider itself talks to fal with the standard library. Portable build:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\comfyui-arkennemasis\requirements.txt

Then restart ComfyUI. The node shows up under arkennemasis/fal/Image/Flux.

You need a fal key. Put one line in the .env file of your ComfyUI install - in the portable build, the folder holding run_nvidia_gpu.bat:

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

There is deliberately no key box on the node, which is why a key can never leak into a shared workflow.

Where people get burned

The key has to be in .env. The pack reads it from there and nowhere else - an FAL_KEY exported in your shell or set as a system environment variable won't be found, and you'll get "fal: no FAL_KEY found" pointing at the path it looked in. That message is accurate; read the path.

Cost scales with the square of your resolution. The unit is the megapixel, and fal counts 1024×1024 as one. A 2048×2048 custom size is four megapixels, so four times the price, and num_images multiplies on top. The live price badge on the node follows your settings - glance at it before you queue a batch of four.

Queueing, not failing. max_concurrent defaults to 1 across every fal node in a run, so a canvas full of these runs them one at a time. If you wired up five and they're all sitting there saying they're waiting their turn, that's the default working as designed; raise it on the nodes you want side by side.