arkennemasis fal · Flux 2 Pro Edit · $0.03 + $0.015/MP

This is the editor of the Pro pair: the same API-only flagship tier, with image_1 through image_6 and an image_size that defaults to auto so the output follows what you fed it. Pricing is $0.03 for the first megapixel of output, then $0.015 per additional megapixel of input and output, rounded up - so a 1024×1024 output is $0.03 and a 1920×1080 output is $0.045.

What you're actually renting

FLUX.2 [pro] never leaves the API - that's BFL's pattern going back to Flux.1, where Pro stayed closed while Schnell shipped Apache 2.0 and Dev shipped non-commercial. There is no local equivalent at any VRAM, so the only question this node answers is whether the job is worth the per-megapixel rate. Multi-reference editing is the usual yes: character plus object plus style, conditioned together, no fine-tuning, no ControlNet stack.

Inputs

prompt is the edit instruction - write it as an action on the file. "Replace the background with a plain grey studio wall", "remove the text on the mug". Describe the scene instead and the model drifts into generating something adjacent to your picture rather than editing it.

image_1 is required, and extra sockets run to image_6; batches are allowed and counted in socket order. image_size defaults to auto here, letting the model follow the input's shape - the sane default for editing, and the reason this node behaves differently from the Klein editors in this pack, which default to straightforward square_hd. Use custom with image_size_width and image_size_height when the output shape has to be exact.

seed at -1 randomises; output_format defaults to jpeg (set png if you're going to edit the result again); safety_tolerance runs 1 (strictest) to 5 (most permissive), default 2, with enable_safety_checker alongside it, and the tooltip notes that turning the checker off requires authorisation on your fal account. There's no num_images and no step or guidance control - the Pro tier decides for you.

The four shared options are identical on every fal node in the pack: max_cost_usd (default $20) refuses the run before upload if the estimate is too high, which is genuinely useful given that input pictures count toward the bill; reuse_identical_run frees an exact repeat; extra_json passes through fal fields the node has no box for; max_concurrent (default 1) keeps paid fal calls serial rather than simultaneous.

Outputs

images gives you the edited batch for the rest of the graph, image_paths the files on disk, seed the reproducibility handle, and info the JSON record - request id, estimate, saved files, fal's full response. This endpoint's node returns a slimmer set than the Klein editors: no timings, no has_nsfw_concepts, no prompt passthrough.

Install

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

Restart ComfyUI, then:

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

into the .env beside run_nvidia_gpu.bat. Standard library only - no pip install for the fal nodes; the pack's httpx requirement belongs to its Codex/ChatGPT nodes.

Where people get burned

The swapped descriptions, first, because it's the kind of thing that wastes an hour: fal's page copy for the two Pro endpoints is effectively inverted, with the plain pro page describing editing and the pro/edit page describing text-to-image. Your only reliable guide is the schema - if there's no image socket, it isn't the editor.

Second, cost with inputs counted: a 2MP input plus a 1MP output is three billed megapixels, so roughly $0.06 per attempt, and loops make that multiply silently. Set max_cost_usd to something you'd actually be happy to spend. Third, expectations on identity - Pro is excellent at composition and multi-reference conditioning, but if a specific face must come through untouched, an identity-preserving editor is still the safer tool. And the standard caveat: your references go to fal before anything renders, so decide what you're comfortable uploading.