arkennemasis fal · Nano Banana 2 Edit · $0.06-0.16/image

This is the most unusual input set in the pack. Beyond the six image sockets, Nano Banana 2 Edit accepts a video, an audio file and a PDF as context for the edit - and none of the image sockets are even required, as long as one of those other inputs is there. You describe an output; it uses everything you connected to make it.

What that actually enables

Three workflows, and they're not the ones you'd guess:

Multi-reference compositing. Six image sockets, each accepting a batch and sent in socket order. That's enough to hand it a product, a room, a logo, a swatch and two references and ask for a single coherent shot - the class of job that used to be a mask, a ControlNet and an hour of fiddling.

Video as context. Connect a clip and ask for a frame, a storyboard panel, or an image matching its look. The node accepts an http(s)/data URL to a video file (downloaded and sent inline, max 15MB) or a YouTube URL, which is passed through to the model without downloading. That last part is worth reading twice: a YouTube link is the one input here that doesn't count against the size limit.

PDF as context. A link, downloaded and sent inline, 15MB cap. Point it at a spec sheet, a menu, a brand guideline, and ask for an image that follows it. It's a small, weird, genuinely useful door.

The dials

prompt is the instruction; resolution (1K default, plus 0.5K/2K/4K - billed at 0.75×/1×/1.5×/2× the $0.08 base, so 4K is 16 cents); aspect_ratio defaults to auto and supports the extreme 4:1/1:4/8:1/1:8 set; num_images 1–4, each billed; enable_web_search (+$0.015); thinking_level (minimal/high, +$0.002 at high); system_prompt for a standing style instruction; safety_tolerance 1–6; seed; limit_generations, on by default, which discards in-prompt requests for multiple variations.

Outputs

images (batchable IMAGE), image_paths (files under output/fal/fal-ai_nano-banana-2_edit/), description (the model's own text about the result), info (JSON: 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 install through ComfyUI-Manager by searching arkennemasis. Only httpx is declared - for the pack's Codex nodes - and the fal provider needs nothing beyond the standard library. Then:

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

in the .env holding your launcher.

One thing worth internalising about this pack, since this node sends so much: everything you connect here leaves your machine. Images, clips, audio and documents go to fal and on to Google. The pack is careful with your key - no key widget anywhere, so it can't end up inside a shared workflow - but your inputs are the price of the feature.

Where people get burned

• "At least one input is required" and you only typed a prompt. This node has no required image socket, but it's an edit endpoint: prompt alone isn't enough unless one of the image/video/audio/PDF inputs is connected.
• A 40MB clip. The inline inputs cap at 15MB, and an oversized one fails in pre-flight - before anything bills. Trim, or hand it a YouTube URL instead.
• Wrong order in a multi-image edit. Batches flatten in socket order: everything on socket 1 goes first.
• Six images and a 4K render is a 16-cent-plus call, not an 8-cent one. Watch the badge - the price rule here multiplies by resolution and adds for search and thinking.
• You wanted a pixel-exact face transfer. This is generative instruction editing, not a paste. The KB's identity chapter covers the tools that actually composite.
• A refusal from Google, with safety_tolerance at 6 and nowhere left to go. auto aspect plus a vague prompt is also a reliable way to get a reframe you didn't ask for - pin the ratio.