arkennemasis fal · Flux 2 LoRA Gallery Ballpoint Pen Sketch · $0.021/MP

Most of the fal LoRA Gallery nodes are transformations - give them a picture, get the picture back restyled. This one isn't. It's text-to-image: you type a prompt, you get a ballpoint pen sketch, and it costs $0.021 per megapixel. Whoever wrote the endpoint's description ("ballpoint pen sketch drawing style") left the word text-to-image out, which has confused a fair number of people into connecting an image that the node simply ignores.

What it is

fal hosted a flux-2-lora-gallery series: one endpoint per trained style, each a fixed LoRA on Flux 2. The pack turns each into a ComfyUI node with the LoRA's strength exposed. Here, that's lora_scale; the rest of the node is a Flux 2 generation with sensible defaults - 40 steps, guidance 2.5, regular acceleration.

Inputs

prompt is the whole job. There's a documented trigger word - b4llp01nt - and it's worth typing carefully: rare alphanumeric tokens are how style LoRAs avoid colliding with ordinary words, and the leetspeak is deliberate, not a typo. b4llp01nt, a crow perched on a fire hydrant, cross-hatched shading.

lora_scale (0–2, default 1) is style strength. Past 1 the pen gets aggressive - heavy, scribbly, high-contrast; below 1 you drift toward generic line art, at which point you should ask yourself why you're paying for the LoRA at all.

num_inference_steps defaults to 40 and runs 4–50, guidance_scale (0–20) sits at 2.5, and acceleration (none / regular) trades speed for quality. image_size defaults to landscape_4_3 with the usual preset list plus custom; num_images (1–4) generates a batch in a single paid call, which is how you'd use this - four sketches, pick one. seed at -1 lets the model choose; output_format defaults to png, and the safety checker returns flagged output as black images.

Worth knowing about Flux 2 generally: the prompt is read by an LLM encoder as a chat turn, so masterpiece, best quality, ultra detailed is dead weight and ((sketch:1.4)) is disabled syntax on this encoder. Sentences describing the subject and the framing do the work.

The shared four: max_cost_usd ($20 default) blocks a run before upload, reuse_identical_run reuses an identical paid result, extra_json injects fal fields the node doesn't expose, and max_concurrent (default 1) serialises paid fal calls across your graph.

Outputs

images - a normal IMAGE batch for save nodes, upscalers or a contact sheet - plus image_paths, seed, prompt and info, the JSON with request id, estimate and fal's response.

Install

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

Restart, then:

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

in the .env next to run_nvidia_gpu.bat. No pip install needed for the fal nodes; the pack's httpx requirement is for its ChatGPT/Codex side.

Where people get burned

The misread description above is the big one - if you wire an image into this node, nothing happens to it; there's no image socket at all. Second, the trigger word: mistype it and the LoRA still fires (it's baked into the endpoint), but you're not getting the "best results" the tooltip promises. Third, the economics. $0.021 a megapixel is fine for a sketch; it's silly for twenty sketches of the same subject, so set num_images to 4 and pick, rather than re-running the node while you search.

And the usual cloud caveat, since it applies to every node in this gallery: your prompt goes to fal, billed to your account, decided by fal's safety checker. If you want a sketch style you can iterate on for free at 3am, train a LoRA for a model you can run - that's exactly what the LoRA in this endpoint is, and Klein 4B trains on an 8GB card.