arkennemasis fal · Flux Krea LoRA · $0.035/MP

Krea's aesthetic plus LoRA support, hosted, $0.035 per megapixel. Node-wise this is the Krea text-to-image endpoint with one extra box - loras - and that box is where all the confusion lives. Everything else behaves exactly like the plain Krea node.

The loras box

It's a multiline string holding JSON, and the tooltip tells you the shape: any number of LoRAs, merged together to make the final image. What it wants is a JSON list of LoRA entries, each naming the weights and a scale - the same convention fal's own API uses. Practically that means a URL to a .safetensors file fal can fetch, plus a strength. Empty leaves it out entirely, which is why the field can sit there harmlessly while you test a prompt.

The gotcha is the one you'd expect from a text field: it's not a dropdown of your local models/loras folder. These are LoRAs fal loads on its side, so your local files are invisible to it unless they're reachable over the internet. The usual pattern is registering them with fal, or pointing at a Civitai/Hugging Face URL. If you paste a local path, nothing happens or you get a request error, and you'll spend ten minutes blaming the prompt.

The rest of the node

• prompt - natural language, full sentences.
• image_size, plus image_size_width / image_size_height when it's set to custom (16–14142, step 16).
• num_inference_steps default 28 (1–50). guidance_scale default 3.5, range 0–35 - note that's lower than the plain Krea node's 4.5, which is the LoRA endpoint's own default rather than a mistake.
• num_images 1–4. The tooltip notes this is always 1 for streaming output, so don't plan on four in one call.
• seed, enable_safety_checker, output_format (jpeg/png), acceleration.

And the four widgets every fal node in this pack carries: max_cost_usd (default $20, caps the run before upload; 0 = off), reuse_identical_run (default on - identical settings and input files return the already-paid results instead of billing again), extra_json (a JSON object merged into the request last, for fal fields this node has no box for), and max_concurrent.

Outputs

images for the preview/save path, image_paths for where it wrote them (output/fal/fal-ai_flux-krea-lora/), plus timings, seed, has_nsfw_concepts, prompt (the echoed prompt) and info (request id, cost estimate, saved files, fal's full answer). That info string is the one to open when something looks odd - it's the only place the actual request id is easy to grab.

Install

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

Manager users: search comfyui-arkennemasis, or Install via Git URL with the repo URL and let it pull httpx for you. The fal nodes need nothing beyond the standard library, so a manual install works with no pip step at all. Restart ComfyUI, then look under arkennemasis/fal/Image/Flux.

Key setup is one line in the .env beside run_nvidia_gpu.bat:

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

No key box on the node - the pack reads the file and nowhere else, so setting the variable in your shell won't work.

Troubleshooting

"is not valid JSON". The node validates loras before anything is uploaded, so a trailing comma or single quotes gets you a free, unbilled error. Build the string in a text editor, not by hand-typing brackets.

Nothing changed after adding the LoRA. Either the entry doesn't resolve on fal's side, or the scale is too low to show. Raise the scale before you go hunting for other causes.

Same output twice. That's reuse_identical_run doing its job. Turn it off, or change the seed, when you want a fresh sample for the same settings.

Price creep. $0.035/MP on a custom 1536×1536 (2.25MP) is 2.25× the per-image line on the title, and num_images multiplies on top. The node's live badge has already done that arithmetic for you.