Nano Banana 2

Nano Banana 2 (Gemini 3.1 Flash Image) is Google's conversational image model - the one that edits from a sentence, holds character consistency across turns, and thinks before it draws when you let it. GeminiNanoBanana2V2 is the current ComfyUI wrapper for it, and it's the version you should actually use. The original GeminiNanoBanana2 node is deprecated; the V2 landed in May 2026 and fixed the real problem with the first one: the interface. Same model, but now the model selector is a DynamicCombo that reconfigures the node around your choice, and the fiddly controls moved where they belong.

Mechanics, briefly: this is a synchronous call to Google's Vertex AI through Comfy's proxy at api.comfy.org - prompt, up to 14 reference images, and optional context files go into a generateContent request, and the result comes back in one round-trip. No polling loop like the BFL nodes, no model files, no VRAM. Billed from your Comfy account credits, no bring-your-own-key (a standing community sore spot - there's a GitHub discussion for voting on it).

The model DynamicCombo is where the V2 earns its keep. Pick Nano Banana 2 (Gemini 3.1 Flash Image) and you get nested aspect_ratio (auto through 21:9, plus the extreme 8:1 / 1:8 letterbox options), resolution (1K/2K/4K - 2K/4K use Gemini's native upscaler), thinking_level (MINIMAL/HIGH - HIGH is the signature feature that improves complex edits), an Autogrow images slot for up to 14 references, and a files input for the Gemini Generate Content Input Files node. Or pick Nano Banana 2 Lite - the cheaper, 1K-only tier for quick edits that don't need the full model. That's the whole upgrade: model choice becomes a real dropdown instead of a hardcoded string.

Outside the combo, the V2 adds temperature (0–2, default 1) and top_p (default 0.95) - the classic sampling knobs, both marked advanced, both genuinely useful if you're chasing a specific style: lower temperature for tighter, more deterministic-feeling output. response_modalities flips between IMAGE and IMAGE+TEXT, and the default system_prompt is a solid "always produce an image" instruction you should leave alone.

Outputs: IMAGE, a STRING with the model's text, and thought_image - the first thinking-pass sketch, only present with thinking_level = HIGH and IMAGE+TEXT. Wire the IMAGE to Save Image; ignore the rest unless you want them.

Troubleshooting worth knowing: this model family had a stretch in early 2026 where the API consumed credits and returned text but no image - blank 13KB files, reproducibly across machines, an API-side or parsing failure rather than your setup. Fixed since, but if it recurs, update ComfyUI first. And the seed is explicitly best-effort: Google's own tooltip says deterministic output isn't guaranteed, and any parameter change can shift the result.

Is it worth the credits? For instruction editing and character-consistent multi-turn work, Nano Banana is the class leader and the V2 node is the least painful way to use it from ComfyUI. Just don't treat it as a reproduction machine - treat it as a very smart collaborator.