arkennemasis fal · Flux 2 Max Edit · $0.07 + $0.03/MP

The editing half of the Max tier: same flagship model, an image_1 socket, and pricing that counts your inputs as processed megapixels. $0.07 for the first, $0.03 for each additional. It's the most expensive node in this pack's Flux 2 range, and the reason to pay it is multi-reference consistency - the one capability the community concedes nothing local matches, at least not at this quality.

What it does

A wrapper over fal-ai/flux-2-max/edit: your reference images are uploaded to fal's CDN, the job goes to fal's queue, and the results land in output/fal/fal-ai_flux-2-max_edit/. No local model, no ControlNet, no loops - one call, one billed result set.

Like the rest of the Max tier, it hands you almost no knobs: no step count, no guidance scale, no num_images. That's the tier's pitch, and it cuts both ways - fewer ways to get it wrong, and no way to trade quality for speed when you're iterating.

Inputs

prompt is the edit instruction. Flux 2 responds best to commands on the file rather than scene descriptions: "remove the laptop from the desk", "zoom out until the full shoes are visible. Maintain the faces." Describe the subject and the model reinvents it.

image_1 is required; sockets run to image_6, and per the tooltip each socket may carry a batch, sent in socket order. Multi-reference conditioning is where Flux 2's reputation lives - character plus object plus style in one generation, without fine-tuning. If you're feeding several references, send them deliberately: the first picture of socket 1 is number 1.

image_size defaults to auto on this node, which means the model decides the output shape based on what you fed it - a nicer default than the Klein nodes' fixed square_hd, and the right one for editing. custom hands you image_size_width and image_size_height if you need to pin the shape. seed at -1 randomises.

output_format defaults to jpeg. For an edit you plan to grade or re-edit, set png. safety_tolerance is a 1–5 combo, 1 strictest and 5 most permissive, default 2, with enable_safety_checker alongside; the tooltip is explicit that disabling the checker requires fal account authorisation.

The standard four: max_cost_usd (default $20, 0 = off) stops the run before upload when the estimate exceeds it - here that pre-flight check is the most valuable field on the node, since a 4MP request with three references is over half a dollar; reuse_identical_run fingerprints settings and the bytes of your pictures so an exact repeat is free; extra_json injects fal fields with no box on the node; max_concurrent (default 1) serialises paid fal calls.

Outputs

images and image_paths for the graph and the disk, seed for reproducibility, and info - the JSON with request id, estimate, saved files and fal's full answer. Note this node's output set is slimmer 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, and put the key in the .env next to run_nvidia_gpu.bat:

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

No pip step for the fal nodes - standard library only.

Where people get burned

The bill, because inputs count: a 2MP reference plus a 2MP output is four processed megapixels, so $0.07 + three times $0.03 = $0.16 for a single image. Four references and a large output gets you toward half a dollar per attempt, and there is no batch option to spread that across.

Second: no dials. If an edit comes back wrong you can change the prompt, the seed, the size or the references - that's the entire surface. Third: the same identity caveat as every Flux editor. Max is strong on composition and multi-reference work, but if the requirement is that a specific face survives untouched, a dedicated identity-preserving edit model is the more reliable call, even though it looks less impressive on a test shot.