arkennemasis fal · Flux 2 LoRA Edit · $0.021/MP

If the KB is right that “Flux 2 dev, even at fp8 turbo, is the best local edit model by far” - and the people doing serious editing work do keep saying that - then this is the node that gets you that model without the 53GB of weights. It's the editing endpoint of FLUX.2 [dev] with a LoRA slot and up to four reference images, at $0.021 per megapixel of input and output.

The mechanism, briefly

The node is a wrapper around fal-ai/flux-2/lora/edit: it uploads your pictures to fal's CDN, submits a queued job with your prompt and settings, polls until it's done, and saves the results into output/fal/fal-ai_flux-2_lora_edit/. Anything you connect to the image sockets leaves your machine - that's the deal, not a defect.

What matters about the model underneath: [dev] does text-to-image, single-reference and multi-reference editing in one checkpoint, it composes from several reference images better than anything else open, and it's the tier that supports character sheets. It's also slow and heavy - 28 inference steps at 2.5 guidance by default, because it isn't a distilled 4-stepper.

Inputs

prompt is an instruction on the file. From the editing guides that circulate: name the action and name what must not change - "zoom out until the sitting woman's white sneakers are visible. Maintain the faces. Maintain the posture." Vague vibes produce a remixed picture rather than an edit.

image_1 is required; sockets run to image_6, but fal accepts four images maximum and counts every frame of a connected batch, numbered in socket order. This is the multi-reference case: person plus garment plus prop, all conditioning the same generation.

guidance_scale (0–20, default 2.5) is your adherence dial. num_inference_steps runs 4–50 with a default of 28 - don't knock it down to 4 expecting Klein behaviour, the model isn't step-distilled. acceleration (none / regular / high) trades some quality for turnaround; regular is the default.

loras takes a JSON list of up to three - URL, HuggingFace repo ID or local path - and empty leaves it out. enable_prompt_expansion (off) lets an LLM rewrite your prompt, which is convenient and occasionally annoying. image_size defaults to square_hd; use custom if the aspect ratio of your source matters, since nothing here preserves it automatically. num_images (1–4), output_format (png default), seed (-1 means the model decides), and the safety checker, which returns flagged results as black images.

Shared options: max_cost_usd ($20) refuses the run pre-upload if the estimate exceeds it, reuse_identical_run gives you a paid-identical re-run for free, extra_json merges raw JSON for unsupported fields, max_concurrent caps concurrent paid fal calls.

Outputs

images is the edited image batch; the rest is bookkeeping you'll actually use - image_paths, seed, prompt, timings, has_nsfw_concepts, and info with the request id, cost estimate, files and fal's full response.

Install

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

Restart ComfyUI, then add the key to the .env that sits beside run_nvidia_gpu.bat:

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

No pip step for the fal nodes - that provider is standard library only; the httpx dependency in the pack is for its ChatGPT/Codex nodes.

Where people get burned

Cost first, because it's doubled here: $0.021 per megapixel of input and output, so a 1024×1024 result from a 512×512 source is $0.042, and num_images: 4 is four output megapixels. Second, the LoRA surcharge - fal adds 50% per gigabyte for stacks over 2GB, and the node's badge doesn't know your LoRA sizes, so the estimate can understate a heavy stack by 1.5× to 2.5×.

Third, the thing the model is bad at: [dev] is strong on composition and character sheets, but if you only need one small change to a photo, you're paying for a 32B model to do a job a smaller editor does for a fifth of the price. Reach for this when the edit genuinely needs multi-reference understanding.