arkennemasis fal · Nano Banana Edit · $0.0398/image

Six image sockets, one text box, $0.0398 an image. Describe the change; the picture comes back changed. No mask, no denoise strength, no ControlNet, no inpainting checkpoint.

This is the node that made half the classic identity toolkit optional. The KB's identity chapter is blunt about it: the adapter approach - IP-Adapter, InstantID, PuLID - stopped being how people do character consistency, because instruction-driven editing took the job, and Kontext/Qwen-Edit/Klein are where a new consistency workflow starts. Nano Banana Edit is the closed-model version of that same move, and it's four cents a call, which is why "just describe it" beats "build a mask pipeline" for most one-off edits.

How the sockets work

image_1 is required; image_2 through image_6 are optional. Each socket accepts a batch, and the images are sent in socket order - the first picture on socket 1 counts as image number 1, and so on. That ordering is the whole interface to multi-image edits: "put the object in image 1 into the room in image 2" only works if you wired them in that order, and it doesn't matter which socket holds what as long as you say so in the prompt.

A single image is the normal case: one picture in, one instruction. "Remove the street sign, keep everything else." "Make it golden hour." "Change the label text to GOLDEN BANANA." This model reads text in the image well enough that in-image text edits are a genuine use case rather than a party trick.

The rest of the settings

• prompt - the instruction. Say what changes and, if the model gets creative elsewhere, say what must stay.
• aspect_ratio - defaults to auto here, meaning the model decides from the prompt and the input. That's the trap in the pack's general advice: if you ask for "a wide cinematic version", the model may reframe your photo instead of recolouring it. Pin the ratio when the shape matters.
• num_images - 1 to 4 per request, each billed.
• safety_tolerance - 1 to 6, default 4. Same Google IMAGE_SAFETY regime as the rest of the line; the famous-IP restrictions tightened in January 2026.
• seed - -1 lets the model choose.
• limit_generations - on by default, and it disregards in-prompt requests for multiple variations, capping a round at one image.
• output_format - png, jpeg or webp.

Outputs

images - an IMAGE output that can carry a batch. image_paths - the saved files under output/fal/fal-ai_nano-banana_edit/. description - fal's returned text about the result. info - the JSON receipt with request id, estimate and fal's raw answer.

Install

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

Restart, or use ComfyUI-Manager (search arkennemasis). Only httpx is declared, and it belongs to the pack's Codex nodes - the fal provider is standard-library only. Then:

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

next to run_nvidia_gpu.bat.

Where people get burned

• It changed more than you asked. Instruction editing is a negotiation. Name the things that must not move, and pin aspect_ratio so a "cinematic" request can't turn into a reframe.
• Multi-image edits came out in the wrong order. Batches are flattened in socket order: socket 1's pictures ship first. Split the batch upstream if you need a specific pairing.
• You wanted an exact pixel-level face swap. Instruction editing is generative, not a post-process transfer. If you need the actual pixels of a specific face composited in, that's a different tool in the KB's identity chapter.
• One image instead of four. limit_generations. And raise num_images.
• A refusal. That's Google's filter, and safety_tolerance only nudges it.