arkennemasis fal · Flux Pro v1 Fill Finetuned · $0.06/MP

Take Flux Pro Fill, add a finetune, charge one extra cent a megapixel. That's the node. It's the pair-member to Ultra Finetuned in this pack, and it answers a specific need: you've trained a concept on fal and you want it to appear inside a masked region of an existing photo, not on a blank canvas.

The use case is narrower than it sounds and worth spelling out. Somebody wants the product rendered sitting on the client's actual countertop. Somebody wants the character to hold a specific object. Straight Fill can do those if the prompt is good enough; Fill Finetuned does them when the thing being inserted is a concept that only exists in your head, your dataset and your finetune.

The two fields that aren't in plain Fill

• finetune_id - required. The identifier of your trained model on fal. As a string. There is no file picker, because there is no local file: this is BFL/fal-hosted finetuning, not the CivitAI LoRA in your models/loras folder. If that's what you have, use the local Flux Fill workflow instead - importing it here won't work.
• finetune_strength - 0 to 2, defaulting to 0. Zero means off. People install this node, run it once, see no difference, and conclude the endpoint is broken; it's the default value. Raise it until the concept reads, bearing in mind the tooltip's own honesty that the optimal setting depends on the finetune and the prompt.

Everything else is Fill: prompt (describe what goes in the hole and what surrounds it), image and mask (both required - you paint the region in ComfyUI, the pack converts it into the RGBA form fal expects and resizes it to the image), seed, num_images (up to 4), output_format, safety_tolerance, enhance_prompt, plus the shared max_cost_usd, reuse_identical_run, extra_json and max_concurrent.

Money

$0.06 per output megapixel, rounded up, multiplied by num_images. fal counts a megapixel as 1024×1024, and the pack's estimate and the live price badge are built from the same rule, so the number you see before running is the number the guard checks. That guard - max_cost_usd, default $20 - fires before the upload, so a run you decide against costs nothing. reuse_identical_run (on by default) returns a cached result for identical inputs as long as the files are still in output/fal, including after a restart, which makes iterating on a downstream node free.

Outputs

images (IMAGE) and image_paths (STRING) are the working pair. Then seed as an INT, prompt, timings, has_nsfw_concepts, and info - the JSON with the fal request id, estimated cost, saved files and fal's full response. Files go to output/fal/fal-ai_flux-pro_v1_fill-finetuned/.

Setup

Install the pack - comfyui-arkennemasis from ComfyUI Manager, or:

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

Restart ComfyUI, then create or edit the .env in the folder holding run_nvidia_gpu.bat:

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

No models to download and no Python dependencies for this node - the fal wrapper runs on the standard library alone.

Gotchas

The big one is conceptual, and it's why this node has a low ceiling on "just try it": without a finetune on fal, there's nothing to point finetune_id at, and no way to fake it. Second, mask discipline applies exactly as it does to plain Fill - paint past the object, feather the edges, and composite the result back over the original rather than trusting one pass. Third, cost sneaks up through num_images, not through megapixels; four candidates on a 2 MP edit is roughly half a dollar, which is fine, and four candidates on a 4 MP edit every time you tweak a word is not.

Housekeeping you get for free in this pack: a failed run stops the other fal nodes in the same graph from starting (so a failure mid-batch doesn't keep billing), Cancel on the ComfyUI canvas cancels the request on fal too, and a mid-run restart is recoverable - the free Recover Result node picks up the finished job by id from output/fal/_requests.jsonl.