Nano Banana Pro (BudgetPixel)

Nano Banana Pro is the flagship Google image model - Gemini 3 Pro Image - and in the BudgetPixel pack it's the one you pull out when the image has to be right the first time. Google's positioning for this tier is "professional asset production": it reasons about the prompt before generating, handles precise instruction-based edits, and composes from up to nine reference images. In practice that means it's the node for hero assets, client-facing work, and complex edits where the base models tend to fumble - at a price. 120 credits per image at 1K or 2K, 180 at 4K.

It's a hosted API node like the rest of the pack: no weights, no local GPU, no download. The prompt, references and instructions go to the API, the reasoning-heavy pipeline runs server-side, and the image comes back as a core IMAGE tensor in your graph. There is no local path for Nano Banana at all - this is the only way to run it inside ComfyUI.

The inputs

• prompt (required) - description or edit instruction.
• reference_images - up to nine, for multi-reference composition and precise editing. Normal IMAGE connections, so you can feed it locally generated work.
• size - 1K / 2K / 4K. Default is 2K, which is interesting: unlike the main Nano Banana 2 node (default 1K), Pro's default sits at the higher tier. 1K and 2K both bill 120; 4K bills 180.
• aspect_ratio - ten options, 1:1 default.
• num_images - 1–4.

Output is images into Preview Image or Save Image.

Pro vs the rest of the Nano Banana family

The honest framing: Pro is not three times better than Nano Banana 2 at 1K - but the gap shows up exactly where it's supposed to. Long, complex, multi-clause prompts get obeyed more faithfully. Edits that require holding several constraints at once ("keep the lighting, the pose, and the background, change the coat to red") hold up better. And the reference composition is more disciplined. If you're generating ten variations of a product shot for a mood board, Lite at 40 credits is the sane choice. If a specific hero image is going in front of a client, the 120 credits buys you a materially lower chance of the embarrassing redo.

One gotcha: because Pro "reasons" before rendering, it's slower than the Flash-tier siblings even on a good day. Budget for that if you're batching.

Install and issues

Install via ComfyUI Manager (search "BudgetPixel") or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Restart, find it under Add Node → BudgetPixel → Image, and set your API key via BUDGETPIXEL_API_KEY or a budgetpixel.json file - no key widget, by design.

Pack-wide caveats: a recent ComfyUI for core types, and Stop doesn't cancel a submitted render - the server-side job keeps running and bills. And because it's a Google model, the content filters are baked in; anything the hosted Pro refuses, stays refused, with no local weights to modify.