arkennemasis fal · Flux Kontext LoRA Text to Image · $0.035/MP

Yes, the instruction-editing model has a text-to-image endpoint. It sounds like a typo and it isn't - Kontext [dev] is a 12B Flux-line model, and it generates from a prompt like any other Flux checkpoint. The reason this endpoint exists is narrower and more useful than "you can also do t2i": it lets you generate with a Kontext LoRA loaded.

That matters because the Kontext LoRA library is a closed world. Style LoRAs, the Zoom Out LoRA, depth-reference LoRAs - trained against Kontext, never ported to Qwen-Image-Edit or Klein. If the look you want lives in one of those files, you need a Kontext endpoint, and you need a LoRA field, and you need it in generation mode rather than editing mode. That's this node. $0.035 per megapixel, about three and a half cents.

The field that justifies the node

loras takes JSON and the tooltip notes you can list several, merged together. Empty leaves it out. The practical constraint is the one that trips everyone: these must be URLs fal can download, not files sitting in your ComfyUI models/loras folder. Check that the LoRA you want has a fetchable URL before you design a workflow around it - plenty of them only exist as a CivitAI download, and those are effectively unavailable here.

The rest of the inputs

prompt is a generation prompt, not an edit instruction. Worth saying because everything else on the endpoint is Kontext-shaped and it's easy to write "remove the person on the left" into a node with no image socket.

image_size is a combo defaulting to landscape_4_3, with custom plus image_size_width and image_size_height for exact dimensions. Unlike the editing endpoints there's no input image to match, so this is entirely your call.

num_inference_steps defaults to 30, guidance_scale to 2.5 - Flux guidance, not CFG, and the guidance-distilled caveat applies: pushing it up doesn't improve adherence, it just burns the image at more cost. acceleration at none is your speed lever. seed at -1 for a random draw, pinned when you're comparing prompt variants. num_images up to 4 per call, enable_safety_checker on (disabling requires account-level authorization at fal, and unauthorized requests are checked regardless), output_format defaulting to png.

When to reach for it instead of plain Flux

You have a Kontext LoRA and want it in generation mode. The obvious one.

You want Kontext's specific aesthetic. It came from the same lineage as Flux Dev and it renders a particular kind of image - slightly different from what Flux 2 or Klein produce, and some workflows depend on it.

You're building a character set. Kontext's architecture concatenates reference images into the model's context rather than injecting adapter embeddings, which is why it held identity across successive edits with minimal drift and why it became the backbone of turnaround-sheet workflows. Generate the sheet, then take it to a Kontext editing node in this pack to restage it, and you're in the pipeline the community built in 2025 and never fully abandoned.

Where it's the wrong call: cheap prompt-to-image. ArkFal_fal_ai_flux_dev is $0.025/MP and Flux Dev is the more established generator; this endpoint is four and a half cents more at the same size purely for the LoRA field and the family's rendering. If neither of those is your reason, use the cheaper node.

Outputs

images and image_paths (file paths under output/fal/fal-ai_flux-kontext-lora_text-to-image/), plus timings, has_nsfw_concepts, seed - the seed actually used, which is how you get back to a good roll - prompt, and info, a JSON string 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, or install through ComfyUI-Manager → Install via Git URL. The pack's only declared dependency is httpx, and that's for its Codex nodes; the fal provider talks to fal with Python's standard library. No checkpoints download and no GPU is used.

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

in the .env next to run_nvidia_gpu.bat in the portable build, or in the ComfyUI folder. No node in this pack has a key widget - a key typed into a node is saved verbatim into every workflow file you share.

Two things to know before you queue

reuse_identical_run is on by default. Same settings, same prompt, same LoRA JSON and the same saved files still on disk means you get the previous image back and nothing is billed. That's the right default for a workflow you're debugging and an annoying one when you clicked Queue hoping for a fresh take - turn it off for that.

max_cost_usd default is $20 and it stops the run before upload when the estimate is over; 0 removes the cap. At $0.035/MP a four-image batch at 1024×1024 is about 14¢, so the cap is mostly there to catch a mistake - a custom size of 4096 or a fan-out of forty branches.

And the licence note, since it applies to everything Kontext-based: the [dev] weights sit under BFL's non-commercial licence, which the v2.0 revision generalised to cover Fill, Canny, Depth, Redux and Kontext together. Fine for personal work; check it before you deliver for a client.