arkennemasis fal · Flux 2 Edit · $0.012/MP

People who edit for a living still say Flux 2 Dev at fp8 is the best local edit model by a mile - and they also say it's too heavy to use. Both are true. This node is the way to get the first without the second: the multi-reference editing strength of the 32B model, on someone else's card, for about a cent and a half an edit.

What it does

Instruction-based editing. You hand it one to four pictures and a sentence describing the change, and it re-renders the file with that change applied. The multi-reference part is the real differentiator - this is the model that composes from several images at once, which is why it handles character sheets and "put this object in that scene" tasks that smaller editors fumble.

The inputs that matter

prompt - the whole game, and the most-messed-up field on the node. Describe the action on the file, not a scene to imagine. Direct, technical, post-processing language works far better than a vibe: remove blur and noise, fix exposure and color profile, shadow recovery. If you describe the subject instead of the edit, the model slides into generator mode and invents a new person. And when it's a scene edit, name what must not move: "Zoom out until the sneakers are visible. Maintain the faces. Maintain the posture."

image_1 (required) through image_6 - sockets, but fal's endpoint only uses the first four. Each socket can carry a whole batch, and they're sent in socket order, so the first picture on socket 1 is reference number 1. Which one you put where matters; that's the ordering.

image_size - square_hd by default, with custom driving image_size_width / image_size_height. For a 1024×1024 output, that's 1MP.

num_inference_steps (28), guidance_scale (2.5), acceleration (none/regular/high), enable_prompt_expansion (off - this one rewrites your careful edit instruction, so leave it alone), output_format, enable_safety_checker (flagged images come back black).

What it costs, exactly

This endpoint bills per megapixel of input and output, and inputs get resized to 1MP. So a 1024×1024 render from a 512×512 reference is 1MP in plus 1MP out: about $0.024. That's the number to keep in your head, because the node's price badge follows your image_size and num_images, not your reference image. Four references are still 1MP of input - you're paying for one, not four.

Outputs

images is the edited image (a batch if num_images > 1) - straight into SaveImage or a comparison board. image_paths gives you the file on disk, seed lets you reproduce a take, prompt echoes the instruction, timings is fal's timing JSON, has_nsfw_concepts reports the checker's read, and info carries the request id, estimate, saved files and fal's full answer.

Install

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

Restart, and look under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL → the same repo works if you'd rather not touch a terminal. requirements.txt is one line, httpx, and that's for the pack's ChatGPT nodes; fal runs on Python's standard library, so nothing else to install.

Key goes in the .env next to run_nvidia_gpu.bat:

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

Where people get burned

Identity drift is the documented weakness of this model, and it's legitimate: it does not hold a face the way Qwen-Image-Edit does. The result looks more real, but the character's features and expression shift. If face preservation is the job, that's a Qwen edit model, not this one. Elsewhere the community's fix is a dedicated consistency LoRA, which is what the LoRA-flavoured variants in this pack are for.

A warm colour shift on edits is reproducible on this model regardless of precision - the output comes back slightly warmer than the source. Match it back with a colour/histogram node against the original if you're cutting edits into a sequence.

Moving parts get re-rendered with a new, plausible version. The prompt fix is explicit negatives: keep the faces, keep the posture, don't change the background.

Repeated identical runs return cached files because reuse_identical_run is on by default. That's a feature - it also means a wasted run is never charged twice - but flip it off when you want a different take from the same instruction. And keep max_cost_usd modest while you're dialling in a prompt; the check happens before upload, so a run that trips the cap is free.