arkennemasis fal · Flux 2 LoRA Gallery Digital Comic Art · $0.021/MP

The endpoint's own description says "transforms images into comic book style", which is half a sentence away from what it does. Read the inputs and you'll see the truth: there is no image socket. This is text-to-image with fal's comic-art LoRA glued to Flux 2, $0.021 per megapixel. If your plan was to run a photo through it, the node will quietly ignore the picture and draw you something new.

What it actually does

One of the flux-2-lora-gallery endpoints - a fixed style LoRA, hosted on fal, wrapped by the pack as a single node. Prompt in, comic art out, files saved under output/fal/fal-ai_flux-2-lora-gallery_digital-comic-art/. Nothing local runs; the node uploads nothing and submits the prompt straight to fal's queue.

Inputs

prompt does everything, and there's a trigger word: d1g1t4l. Rare leetspeak tokens like that are standard practice for style LoRAs - the point is that it can't collide with a real word you might use by accident. Using it is what gets you the "best results" the tooltip promises. Otherwise write the shot: subject, framing, action. Flux 2's encoder is an LLM reading a chat turn, so tags and weighting syntax don't do anything here.

lora_scale (0–2, default 1) is the style strength - above 1 pushes toward heavy inks and halftone, below 1 fades to generic illustration. guidance_scale (0–20, default 2.5) and num_inference_steps (4–50, default 40) are the flux dials; 40 steps is a lot for a style endpoint, and dropping to 20 is a reasonable first test if you're iterating on composition.

acceleration is none or regular - regular is the default and the sane one. image_size defaults to landscape_4_3 plus the usual presets and custom, which is your panel shape, so set it deliberately: a comic strip is not square by default here. num_images (1–4) gives you a page's worth in one paid call. seed at -1 randomises; output_format defaults to png (keep it, you may want the alpha later); the safety checker returns flagged images as black.

The four shared options behave the same on every node in this pack: max_cost_usd (default $20) refuses a run before it's uploaded if the estimate is over, reuse_identical_run hands back an identical already-paid result, extra_json merges raw JSON for fal fields the node doesn't expose, max_concurrent (default 1) keeps paid calls from firing in parallel.

Outputs

images is your batch - wire it to SaveImage, or to a mask/upscale chain if you're compositing panels. image_paths, seed, prompt and info come along; info is the JSON receipt with the request id, cost estimate, saved files and fal's full answer.

Install

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

Restart, then add to the .env next to run_nvidia_gpu.bat:

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

The fal provider is pure standard library, so no pip install is involved for these nodes - the pack's httpx dependency is there for its Codex/ChatGPT nodes.

Where people get burned

The description mismatch is the trap. If what you wanted was your photo in comic style, this is the wrong node - this gallery's transformation endpoints (add background, apartment staging, face to full portrait, multiple angles, virtual try-on) take images; the style ones don't. For a restyle of an existing picture you want an editing node, and if the subject has to stay recognisable, prompt it like a Photoshop action rather than describing a scene.

Second: text in comic art. Flux 2 puts letters in panels whether you asked or not, and they're usually gibberish. Say "no lettering" or accept that you'll be covering it in post. Third: at $0.021 per megapixel a 1024×1024 image is $0.021 and num_images: 4 is $0.084 - cheap until you're exploring, at which point you've spent more than a LoRA on a model you can run locally would have cost you in an evening.