arkennemasis fal · Nano Banana Pro Edit · $0.15-0.30/image

The Pro edit node takes the instruction-editing idea and points it at the job local pipelines are worst at: combining several references into one believable image. image_1 is required; image_2 through image_6 are optional, each socket accepting a batch. $0.15 an image, $0.30 at 4K.

Where the cheap edit node is a one-photo recolour machine, this is the node you use when the answer needs a product from a photo, a background from another photo, a logo that has to be rendered exactly, and enough resolution that the label text is still legible when the client zooms in.

The multi-reference move

Each socket accepts a batch, and images are sent in socket order - everything on socket 1 first. So the grammar of a multi-reference prompt is positional: "place the object in image 1 on the table in image 2, keeping the label legible and the lighting of image 2." Ambiguity about which reference means what is the most common cause of a disappointing result, and it's entirely on you to remove it.

This is also where the KB's identity chapter meets the closed model. For stills, adapter stacks lost this job: instruction-driven editing is what people reach for, and a good model with several references beats a PuLID-plus-mask pipeline on anything that isn't a pixel-exact face transfer. If you do need pixel-exact - the actual face from the actual photo composited in - this isn't the tool; it's generative instruction editing, not a paste.

The dials

prompt - the instruction, and the clearer you are about which reference is which, the better. resolution - 1K (default), 2K, 4K, with fal's multiplier taking 4K to 2× the base. aspect_ratio - defaults to auto, so the model may follow your instruction or your input; pin it when the shape is the point. num_images 1–4, billed each. enable_web_search (+$0.015) for current real-world references. system_prompt for a standing style or "no extra text" rule. safety_tolerance 1–6, default 4, under Google's IMAGE_SAFETY regime - and the January 2026 tightening on famous IP is real. seed, with -1 for model's choice. limit_generations on by default, capping a round at one image.

Outputs

images (batchable IMAGE - feed an upscaler, a save node, or a video model's first frame), image_paths (files under output/fal/fal-ai_nano-banana-pro_edit/), description (the model's own text about the result - surprisingly good as a caption), info (JSON receipt: request id, cost estimate, saved files, fal's answer).

Install

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

Restart ComfyUI, or search arkennemasis in ComfyUI-Manager. The only declared dependency is httpx, and it's there for the pack's Codex nodes - the fal side is standard library only, so this install won't disturb your existing packages. Then:

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

in the .env next to run_nvidia_gpu.bat.

And remember what you're sending: five reference photos, a prompt, and Google's moderation layer on the other end. No key widget exists on any node in this pack, on purpose, but your images are the input by definition.

Where people get burned

• Reference ambiguity. "Image 1" and "image 2" only mean something if your prompt uses those words. Batches flatten in socket order, so count the pictures, not the sockets.
• You forgot which socket was which. Label upstream - reuse or a description output - before you wire five references into a prompt you'll debug at 30 cents a try.
• Math: five 4K images is $1.50. num_images × resolution multiplier × $0.15. The badge tracks it live; trust it over your intuition.
• auto reframed the composition when you asked for something cinematic. Pin aspect_ratio.
• You wanted a 15MB-plus video or a whole PDF as context. Those inputs live on the Nano Banana 2 Edit node, not this one.
• A refusal you can't negotiate with. That's Google, safety_tolerance included.