arkennemasis fal · Flux Kontext LoRA · $0.035/image

Kontext got superseded as an editing model, and its LoRA library didn't go anywhere. That's the entire reason this node exists. The Depth reference LoRA, the Zoom Out LoRA, and a pile of Flux style LoRAs - watercolour, ink, brand looks - were trained against Kontext and were never retrained for Qwen-Image-Edit or Klein. If you want them, you need a Kontext endpoint, and this is the one with a LoRA field.

It's the same base and roughly the same price as the plain Kontext node: $0.035 per megapixel, which their tag rounds to about three and a half cents a megapixel. You're paying a cent more than ...kontext_dev for the loras input.

The one field that changes everything

loras takes JSON - the LoRAs to merge into the generation, and the tooltip notes you can list several and they'll be merged together. Empty leaves it out, and empty is perfectly valid if you're just here for a cheap Kontext.

The gotcha, and it's the one that stops people dead the first time: fal has to be able to download those files. A LoRA sitting in your ComfyUI models/loras folder is a local path; the endpoint can't see it. You need a URL - a Hugging Face repo, your own bucket, any host that serves the file. Check whether the LoRA you want has a published URL on its model page before you build a workflow around it, because the ones that only exist as a CivitAI download are, practically speaking, unavailable to this endpoint.

The rest of the node is Kontext as usual: image required (up to 14142px per side), prompt - the edit instruction - resolution_mode defaulting to match_input so the output keeps your input's shape (auto picks from a preset list matched to the input's aspect ratio), acceleration none, num_inference_steps 30, guidance_scale 2.5 (Flux guidance, not CFG), seed -1, num_images, enable_safety_checker on, output_format png here.

When a Kontext LoRA is the right call

A style you can't get any other way. Flux style LoRAs trained on Kontext are still excellent, and a two-and-a-half-cent edit is cheaper than retraining anything.

Zoom Out and depth-reference workflows. These are the two the community keeps naming when someone asks why Kontext is still installed. Zoom Out in particular does something the newer editors don't: push the composition outwards from a subject without re-inventing the subject.

Character consistency across a series. Kontext's architecture puts reference images in-context rather than through an adapter, so identity holds across successive edits with minimal drift. That predates Klein, and it still works.

Where a Kontext LoRA isn't the right call: if you're doing plain instruction editing, the newer models do it better and the LoRA-free Kontext node in this pack is cheaper. Bringing a LoRA in just slows the edit down.

Outputs

images plus image_paths (written to output/fal/fal-ai_flux-kontext-lora/), and then the diagnostics: timings, has_nsfw_concepts, seed, prompt, info - request id, cost estimate, saved files and fal's complete answer.

Install

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

Restart ComfyUI, or use ComfyUI-Manager → Install via Git URL. Only httpx is declared in requirements.txt and it's for the pack's Codex nodes - the fal provider runs on Python's standard library and downloads no weights of its own.

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

in the .env next to run_nvidia_gpu.bat (portable build) or your ComfyUI folder. The key is never a widget on these nodes, which is intentional: a key typed into a node gets saved into any workflow you export.

Money and the usual traps

$0.035/MP, so 1024×1024 ≈ 3.5¢ and a 2048-wide render ≈ 14¢; the price badge on the node follows num_images. max_cost_usd (default $20, 0 = no cap) blocks the run before anything is uploaded if the estimate is over it.

reuse_identical_run defaults on: the same image bytes, same prompt and same LoRA JSON return the file you already paid for rather than charging again. That's the behaviour you want while you're nudging the rest of the graph, and a surprise if you queued a re-run hoping for a fresh sample. Turn it off for that.

And the licence reminder, because it applies to anything built on this base: Kontext [dev] is covered by BFL's non-commercial licence, alongside Fill, Canny, Depth and Redux. Personal work, fine. Client delivery, read the current terms first.