Nano Banana 2 (fal.ai)

Nano Banana is Google's own image model, and it's closed - no weights, no downloads, no local runs. It lives behind the Gemini API, and this node reaches it through fal.ai's hosted fal-ai/nano-banana-2 endpoint. So FAL_NanoBanana2 is the "I want Google's current fast model inside my ComfyUI graph without leaving it" button: you type a prompt, a Google model renders it, and a IMAGE tensor lands back in your workflow.

Which model, exactly? Nano Banana 2 is Gemini 3.1 Flash - the February 2026 speed-quality hybrid that the community promptly clocked as fast and, for realism, second only to Nano Banana Pro. The crowd also clocked something else: this is a heavily filtered Google product. Aggressive safety gating and an invisible SynthID watermark ship with every image. It's technically lovely and creatively boxed-in, which is the trade you're making by staying in this walled garden.

How it works

The node sends your prompt to fal-ai/nano-banana-2 via fal.subscribe, waits, and converts the returned images into a batched tensor. No local compute, no model files, no GPU strain - just an API call that costs you money and a few seconds of Google's TPUs.

The inputs that matter:

• prompt (required) - all other knobs are optional.
• aspect_ratio - a long list, from 1:1 up to ultrawide 21:9 and even 8:1/1:8. Default is auto, which lets the model pick.
• resolution - 0.5K, 1K, 2K, or 4K; default 1K. This is the biggest cost lever, since higher resolutions bill higher.
• safety_tolerance (1–6, default 4) - how much the built-in filter lets through. Lower is stricter; expect more refusals the further down you go.
• thinking_level (none, minimal, high; default none) - lets the model "reason" about the prompt before generating. Costs latency; high is where the quality/instruction-following juice is.
• enable_web_search - a genuinely odd power feature: lets the model ground generation on live web results. Leave it off unless you have a specific use.
• num_images (1–4), seed (-1 random), output_format (png/jpeg/webp), private_mode (tells fal not to store your I/O).

One IMAGE output, same as every node in this pack - straight into a Preview, Save, or the next stage.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/elle-trudgett/fal-ai-comfyui-nodes
pip install -r fal-ai-comfyui-nodes/requirements.txt

or ComfyUI Manager → ComfyUI fal.ai Nodes. No models to download - the model is Google's. You do need a fal key: export FAL_KEY="..." (or a config.ini with [API] / FAL_KEY in the node folder for ComfyUI Desktop), get it at fal.ai/dashboard/keys, restart.

Gotchas

• "FAL_KEY not set" - the pack's universal first run error; set the env var or config.ini and restart.
• The filter will refuse things. This is Google's safety stack front and center; if you need to generate content this model blocks, no setting here overrides that - safety_tolerance only adjusts how much the model itself tolerates.
• It's metered. 4K + multiple images = a real per-run bill. Check the resolution before you queue a batch.
• Errors come back as fal.ai error: ... with fal's message unwrapped for you.
• Don't expect it to match an uncensored open model's range. Powerful, fast, "unimaginative" is the honest community verdict - great for polish and realism at speed, less so for pushing boundaries.

Wire it next to your open-weight stuff and use it where speed and polish beat control. Just keep the watermark rule in mind if you publish the output.