arkennemasis fal · Nano Banana Lite Edit · ~$0.05/image

The node for "change one thing about this picture"

Every local editing stack you've used needs a mask: paint the shirt, write a prompt about a shirt, hope. This node doesn't. You give it an image and a sentence - "swap the wall colour to deep green, keep the furniture and lighting identical" - and the closed model decides what to touch.

That's why it's worth wiring into a local graph. It's not replacing your inpainting setup; it's the escape hatch for edits where building the mask would take longer than the edit is worth, and for the classes of edit (text on objects, relighting, colour changes, keeping a face recognisable) where these Gemini-backed endpoints are simply better than anything you have local.

The inputs that matter

Six image sockets, one prompt, and that's the shape of it:

• image_1 through image_6 - all optional, each may carry a batch. They're sent in socket order, so the first picture on socket 1 is "number 1" to the model. This is how you do multi-image edits: person on image_1, jacket on image_2, "put the jacket on the person".
• prompt - the instruction. Talk to it like a person briefing an editor, and say what to keep as well as what to change. "Only change X" is the phrase that saves you a re-roll.
• aspect_ratio - default auto, and it supports the extreme ratios (4:1, 1:8, etc.). On an edit this matters more than it looks: auto can let the model reframe, and a reframed edit is usually not what you wanted. If you need the output to match your input's shape exactly, the pack ships a neighbouring utility node for exactly that job (arkennemasis Match Aspect), which measures the image being edited and pins the generator's output to that shape.
• num_images - 1 to 4, billed per image.
• output_format - png, jpeg, webp.
• seed - -1 by default; set one if you want the same edit from the same inputs twice.
• system_prompt - a system instruction, empty by default.
• safety_tolerance - 1 to 6, default 4. Google's moderation with a bit of slack.
• limit_generations - on by default; forces one image per prompting round and ignores in-prompt "make me several" instructions.
• thinking_level - (not set), minimal, high. Optional reasoning pass before drawing.

Shared guards: max_cost_usd (default $20, checked before upload), reuse_identical_run (no double-billing identical inputs, survives restarts), extra_json (raw JSON merge for missing fields), max_concurrent (default 1 paid fal call at a time graph-wide).

How it works, and what it costs

The node uploads your pictures to fal's CDN (free - only the submit bills), calls google/nano-banana-lite/edit, and writes results into output/fal/google_nano-banana-lite_edit/. Pricing is token-based and the node's ~$0.05/image badge is an approximation for a simple one-image edit; image tokens on the output side are the expensive ones ($37.50 per million), so multi-image inputs and long instructions move the number. Trust the pre-run estimate over the badge when the run matters.

By the way - the fal endpoint's own output is fixed 1K. Editing a 4K photo doesn't give you a 4K edit back.

Outputs

images (the edited IMAGE batch), image_paths, description (fal's text for the result), info (request id, cost estimate, saved files, fal's raw answer).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install -r comfyui-arkennemasis/requirements.txt

Restart ComfyUI. (ComfyUI-Manager's Install via Git URL with the repo URL works too.) The only declared dependency is httpx; the fal provider itself uses Python's standard library, so nothing exotic gets pulled in.

Key line in the .env next to your launcher:

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

Where people get burned

Reference order is not negotiable, and batches flatten. Six sockets, each accepting a batch, means the model's "Image 3" could be the second picture on socket 2 - they're numbered in socket order, first picture first. Keep one image per socket until you've confirmed which is which.

The model redraws more than you asked. These endpoints are generative editors, not masked inpainting. Identical faces drift, backgrounds quietly change, and resolution is whatever the endpoint decided. Say "preserve exactly" in the prompt and expect to reroll a couple of times.

auto aspect ratio can re-frame your picture. If the composition must survive, pin the shape.

Moderation applies at Google's end. safety_tolerance loosens; it does not unlock. Nothing local can patch a filter that lives in the model.

Small new packs deserve a read. This one is a fresh, single-author Node pack of API wrappers - the category that once shipped a credential stealer. The key handling here is done sensibly (env file, no key widgets), but reading a network-touching pack before you run it is habit worth keeping.