arkennemasis fal · Flux LoRA Image to Image · $0.035/MP

Picture in, restyled picture out, with a pile of hosted LoRAs available to do the restyling. $0.035 per megapixel. Think of it as the cheap version of a custom fine-tune: instead of training on your reference set, you borrow someone else's LoRA and let the strength dial decide how much of it you get.

The two controls that matter

strength defaults to 0.85 with a range of 0.01–1, and the tooltip is refreshingly honest about the ends: 1.0 completely remakes the image, 0.0 preserves the original. That's your "how much of my photo survives" dial. Most useful work happens before either extreme - around 0.5–0.75 you keep the composition and lose the texture, which is what style transfer actually means in practice.

loras is a multiline JSON string of LoRA entries - each with the model file and a scale - merged together to make the final image. Empty leaves them out. It loads on fal's side, so it takes URLs; your local models/loras folder is not involved. This catches almost everyone the first time.

Everything else is standard: prompt for the target look, image_size (default square_hd) with image_size_width / image_size_height for custom, num_inference_steps 28 (1–50), guidance_scale 3.5 (0–35), seed, num_images 1–4, enable_safety_checker, output_format, acceleration. And the four pack-wide widgets: max_cost_usd (default 20, 0 = off), reuse_identical_run (on), extra_json, max_concurrent.

Outputs

images for your preview/save chain, image_paths for the files under output/fal/fal-ai_flux-lora_image-to-image/, then seed, has_nsfw_concepts, timings, prompt and info.

Install

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

Manager users can search comfyui-arkennemasis and let it install the lot, or use Install via Git URL with the repo address. Restart ComfyUI when it's done. The fal nodes themselves need no pip packages - requirements.txt declares httpx for the Codex/ChatGPT side of the pack.

Key, one line in the .env in the folder that holds run_nvidia_gpu.bat:

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

Then look under arkennemasis/fal/Image/Flux. There's no key field on the node, deliberately.

Where it trips people up

A batch comes in, one picture goes out. These image sockets want a single picture. Split batches upstream.

The square crop, again. image_size defaults to square_hd and the endpoint is told a size, so it isn't inheriting your input's aspect ratio. Set custom dimensions matching the source if you want the framing you had.

Runaway LoRA. A strong LoRA at 0.85 will erase the photograph's own character, including faces. Drop the scale or the strength one at a time so you know which caused it - and remember that two LoRAs at 0.8 each don't behave like one at 1.6.

The same result twice. reuse_identical_run served you the cached files because nothing about the request changed - including the LoRA scale, which is part of the identity check. Nudge anything real, or turn the option off for a fresh take.

Cost you didn't budget. $0.035 × megapixels × num_images. Iterating on a strength value is real money; change one variable at a time and keep num_images at 1 while you search.