arkennemasis fal · Flux 2 LoRA · $0.021/MP

Here's the honest pitch for this one. FLUX.2 [dev] is a 32B transformer with a 24B Mistral-3 encoder - call it 53GB of weights, and the reason the community mostly rejected it for local use. People who do have the hardware still call it the best open editor and one of the best text-to-image models around. This node rents it at $0.021 a megapixel and lets you stack LoRAs on it, which is the combination you cannot buy locally unless you own a workstation card.

What it is

One thin wrapper over fal's fal-ai/flux-2/lora endpoint, generated from a JSON spec in the pack. Your prompt goes out, the images land in output/fal/fal-ai_flux-2_lora/, and the node behaves like any other generator with an IMAGE output. Everything about the local-vs-cloud tradeoff in the KB applies: performance without VRAM, and your prompt and reference images leaving your machine.

This is [dev], not klein - so the default step count is different. num_inference_steps runs 4 to 50 with a default of 28, and guidance_scale sits at 2.5. That's the real, non-distilled workflow, not the 4-step trick.

Inputs

prompt first. Flux 2's encoder is a vision-language model reading a chat turn: write sentences, name what owns what, skip the tag-soup flourishes and the (word:1.4) weighting syntax. guidance_scale (0–20, default 2.5) is how hard the model is pushed toward your prompt - raise it a little if the composition ignores you, but Flux 2 does not want SDXL's 7.

loras is the reason you're here: a JSON list, maximum three, each entry a URL, HuggingFace repo ID or local path. Empty leaves the field out. acceleration is a three-way combo - none, regular, high - with regular as the default balance of speed and quality; high is the one to try when you're iterating and the final render matters less.

enable_prompt_expansion is off by default and worth knowing: flip it and an LLM inflates your prompt before generation. Good for lazy one-liners, bad when you want exactly what you typed. image_size defaults to landscape_4_3 with the usual preset list plus custom for explicit width and height, and num_images (1–4) generates a batch in one call. output_format defaults to png here. The safety checker is on, and it returns unsafe output as black images rather than failing the run.

Then the four shared ones: max_cost_usd ($20 default) is checked before anything is uploaded; reuse_identical_run hashes the settings and the bytes of any inputs so a repeat run is free; extra_json merges a raw object into the request for fields with no box on the node; max_concurrent keeps billed calls serialised across the graph unless you raise it.

Outputs

images for the graph, image_paths for the files, seed and prompt back out for reproducibility and logging, timings and has_nsfw_concepts from fal, and info - a JSON blob with the request id, the cost estimate and fal's whole answer. That's your receipt.

Install

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

Restart, then one line in the .env next to run_nvidia_gpu.bat:

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

The fal provider is standard-library only, so nothing to pip install for these nodes. There is no key box on the node by design - a key in a widget gets saved into the workflow file and travelled with every screenshot.

Where people get burned

The price is not the badge. $0.021 per megapixel is the base rate, and fal adds 50% per gigabyte on LoRAs over 2GB - a 2–3GB stack multiplies the whole request by 1.5×, up to 2.5× for 4–5GB. The node's estimate and live badge don't know how big your LoRAs are, so a heavy stack costs more than the number shown. Check fal's dashboard after a serious session, not just the node.

Second: Flux 2 Dev never got a big LoRA ecosystem, precisely because training it wants something like an RTX 6000 Pro, so the LoRA you find is more likely to be a Klein LoRA - wrong architecture, silent no-op. Third: 28 steps at 2.5 guidance on a full-strength model is slow by API standards; if you're exploring composition, draft on something cheaper and finish here.