arkennemasis fal · Flux 2 Klein 4B Base LoRA · $0.016/MP

Text-to-image on the un-distilled Klein 4B, with up to three LoRA weights applied on top. It's the companion to the base edit LoRA node: same idea, generation instead of editing. $0.016 per megapixel, which is nearly double the plain 4B base node, so the LoRA has to be earning its keep.

The honest framing: base Klein is the checkpoint people train against, and this node is how you sample from one of those fine-tunes without owning the hardware to run it. The 4B bases train in about 12GB of VRAM in bf16 and 8GB in int8, which is the most accessible fine-tuning target Black Forest Labs has ever shipped - and it's Apache 2.0, so what you make with it is yours.

The LoRA box

loras is a JSON text box, not a dropdown of your local files. That's not laziness on the pack's part: fal has to download the weights itself, so what goes in the box is fal's own LoRA list format - up to three entries - pointing at weights that are reachable by URL rather than sitting in your ComfyUI/models/loras folder. Check the model page for the exact JSON shape. Leave it empty and you get plain base 4B.

Two rules that will save you a wasted run. Size matters: 4B and 9B LoRAs are not interchangeable, and mixing them produces garbage, not a subtle degradation. Base versus distilled matters: most published Klein LoRAs are trained on a base checkpoint, which is exactly why this node exists rather than a LoRA on the distilled node.

The inputs

prompt - sentences with clear ownership and spatial relations. The encoder is a Qwen3 LLM reading a chat turn, and prompt weighting is disabled for it, so tag soup and (emphasis:1.4) both underperform plain language.

negative_prompt - a field that exists on this node and not on its distilled sibling, because base runs a real guidance pass. Use it for what you don't want; it works here.

guidance_scale - 5, which is where the community runs base Klein (4-5). num_inference_steps - 28 on the node against the 50 that base really wants, and since fal bills per megapixel rather than per step, raising it costs time and not money.

acceleration (none / regular / high, default regular), image_size (presets or custom with the width/height boxes), num_images, output_format (png default), enable_safety_checker - flagged renders return black squares rather than errors, and turning it off needs fal account authorisation.

Outputs

images is the IMAGE batch, image_paths the saved files as a STRING, timings fal's timing JSON, has_nsfw_concepts the checker's verdict, seed an INT you can lock, prompt the string the model received, and info a JSON blob 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 ComfyUI; the node sits under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL with the same repo is the no-terminal path. The pack's only declared dependency is httpx, for the ChatGPT/Codex half of it - fal needs nothing beyond the standard library.

Key, in the .env by run_nvidia_gpu.bat:

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

Where people get burned

Expecting base to look like a finished model. It's a training checkpoint. At low step counts it's soft and the anatomy is worse than the distilled 4B; that isn't a bug, it's what you're paying $0.016/MP for and why most people use it to validate a fine-tune rather than to ship output.

Wrong-size LoRAs - the single biggest waste of money on nodes like this. Read the LoRA card's base model line before you paste the URL.

LoRAs that need a trigger word. Many do, and the node won't tell you; if a LoRA seems to do nothing, your prompt is missing its keyword.

Cached reruns. reuse_identical_run defaults on, so an identical second run comes back from disk instead of being billed again - a nice guard against double-charging, confusing when you wanted variety. max_cost_usd (default $20) is checked before anything is uploaded; set it to a dollar or two while you're experimenting and an over-budget run costs nothing at all.