☁️BizyAir NanoBananaPro (Official Parameters)

Nano Banana Pro is Google's flagship image model - Gemini 3 Pro Image under a cuter name - and normally you'd need a Google API key and a billing account to touch it from ComfyUI. This node skips that: it's BizyAir routing your request through SiliconFlow's cloud, so you authenticate once with a BizyAir key and pay in BizyAir credits instead of setting up Google Cloud billing. Same model, one less account to manage.

What Nano Banana Pro actually gives you

This isn't a diffusion model you're conditioning with a KSampler - it's a hosted multimodal model you talk to. You write a prompt, optionally hand it up to several reference images, and it reasons about the request before generating (Google calls this "thinking mode") and hands back a 4K-capable image. It's genuinely strong at multi-language text rendering and following detailed instructions, and correspondingly heavily filtered - don't expect it to be your NSFW workhorse.

The inputs that matter

• prompt - plain-language instructions, multiline. Since this model reasons about the prompt rather than just pattern-matching tokens, longer and more specific descriptions genuinely help.
• images (optional) - reference images. Bump inputcount (1–14) to add more reference slots and click "Update inputs" to refresh the node - that 14-image ceiling is Nano Banana Pro's actual documented limit, not an arbitrary BizyAir cap.
• aspect_ratio - eleven choices from 1:1 to 21:9, or auto.
• resolution - 1K, 2K, 4K, or auto. 4K is the headline feature; it also costs more and takes longer.
• temperature / top_p - standard LLM sampling knobs (0–1 each). Leave both at 1 unless you're chasing more deterministic output; this model listens to instructions closely enough that you rarely need to fight it with sampling settings.
• seed and max_tokens - seed for reproducibility, max_tokens caps the model's internal reasoning/response length (default 32768, the ceiling).

Outputs: image (what you came for), string (the model's accompanying text response - useful if you asked it to explain a change or describe what it did), and bizyair_model_name, which just echoes which backend served the request.

Installing it

Through ComfyUI Manager: search BizyAir, click Install, restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

Restart ComfyUI. The install gets you the node; using it needs a separate step the README is explicit about - set your BizyAir API key on first use (there's a "click to login" flow in the node itself to register/authenticate). Skip that and the node just fails at execution time with no image, not a helpful error about missing credentials.

Where people get tripped up

This node only speaks to other BizyAir nodes cleanly - its output is a plain IMAGE, so wiring it into a stock save/preview chain is fine, but if you're trying to chain it into another BizyAir cloud node (say, feeding its output as a reference into another BizyAir editing node), keep the whole segment on BizyAir types rather than routing through local ComfyUI processing in between, or you'll get type-mismatch errors on connect.

The bigger practical issue is cost and quota, not wiring: 4K output at thinking mode reasoning is the expensive end of BizyAir's pricing, and it's easy to burn through credits fast if you're iterating on aspect ratio or prompt wording at full resolution. Drop to 1K or auto while you're dialing in the prompt, then bump to 4K once you're happy - same pattern as testing at low steps before committing to a final render. And because this is Google's model wearing BizyAir's plumbing, the content filtering is Google's, not BizyAir's - a prompt that gets flagged will get flagged the same way it would through Google's own API, credits or no credits.