Nano Banana 2

Nano Banana 2 - Google's Gemini 3.1 Flash Image model, the one that made the "Nano Banana" nickname stick - is one of the most genuinely impressive API image models around: it edits from a sentence, keeps characters consistent across turns, and does it with a reasoning pass first. This node, GeminiNanoBanana2, is the ComfyUI wrapper for it that landed in February 2026. And here's the thing to know up front: it's deprecated. Marked is_deprecated in core, superseded by the V2 node of the same name that shipped in May 2026 with the DynamicCombo interface. If both show up in your search, reach for GeminiNanoBanana2V2 - same model, better UI.

If you've read the partner-node preamble anywhere else in this family, the mechanics will be familiar but the plumbing is different: this node doesn't run a diffusion poll loop like the BFL ones. It's a synchronous call to Google's Vertex AI through Comfy's proxy - your prompt (and up to 14 reference images via the Batch Images node, or files from the Gemini Generate Content Input Files node) go into a generateContent request, and the model's response comes back in one shot. Still billed from your Comfy account credits, still no bring-your-own-key, still nothing on your GPU.

What this node has that the V2 folds away: aspect_ratio (a real dropdown from 1:1 to 21:9, default auto which matches your input image), resolution (1K/2K/4K, where 2K/4K invoke Gemini's own native upscaler), and thinking_level (MINIMAL or HIGH). That last one is the Nano Banana signature: HIGH makes the model think before it draws, which meaningfully improves multi-step edits at the cost of speed. The default system_prompt is a well-written "you are an image engine" instruction - leave it unless you know why you're changing it.

Outputs: IMAGE, a STRING (the model's accompanying text), and a thought_image - the first sketch from the model's thinking process, only present with thinking_level = HIGH and response_modalities = IMAGE+TEXT.

Real troubleshooting, from the community: this node's generation has a known flakiness where the API returns text and consumes tokens but the image payload comes back empty or blank - people reported it failing almost every run for a stretch in early 2026, reproducibly across machines, with no local cause. It's an API-side or parsing issue, not your setup. When that happens, there's no local fix; update ComfyUI and try again, and know you're not alone. The seed is best-effort, and per the tooltip Google says outright that determinism isn't guaranteed.

Worth using? The model, absolutely - it's the best conversational editor Google ships. This specific node, only if you need the version with the older fixed inputs; otherwise the V2 is the maintained path.