Nano Banana Pro (Replicate)

Nano Banana Pro is Google's closed Gemini 3 Pro Image model - the strongest closed image generator around, and one you physically cannot own. No weights, no download, no local fork. If you want it in a ComfyUI workflow, you call it, and that's exactly what this node does. In mid-2026 it's also the single most-requested closed model in the community, so this is probably why you're here.

It deserves the hype more than most. Native 4K output, up to 14 reference images for character consistency, and genuinely good text rendering in multiple languages. The catch, besides price, is the moderation: Google's IMAGE_SAFETY filtering is aggressive and it follows the model, not the node. There is no abliterating a closed model, and no "looseness" slider that actually works. Set expectations accordingly.

How it works

Same engine as every node in Replicate Select: the pack reads a catalog entry for google/nano-banana-pro, packages your inputs, uploads reference images as PNG, calls the Replicate API, and returns the result as an IMAGE tensor you can wire straight into SaveImage or downstream detailers. From the graph's point of view it's just another image generator - the cloud round-trip is the only difference.

The inputs that matter

• prompt (required) - a plain text description. Google's "thinking mode" handles the reasoning internally; you don't get a thinking-dial, you just get better adherence than most models.
• image_input - up to 14 reference images for transformation or subject consistency. This is the feature people pay for.
• resolution - 1K / 2K / 4K. 4K is native and impressive, and also where the per-image price climbs into the cents-to-dimes range. Draft at 1K.
• aspect_ratio - 11 choices, default match_input_image.
• output_format - jpg or png. JPG is the default and fine for most; use png if you're going to re-edit.
• safety_filter_level - the three options are block_low_and_above (strictest), block_medium_and_above, and block_only_high (most permissive, and even it blocks some prompts). Try block_only_high first - you'll still hit refusals, just fewer of them.
• allow_fallback_model - quietly reroutes to bytedance/seedream-5 when Nano Banana Pro is at capacity. Nice for unattended batch runs; the downside is you don't get to choose which model answered.

Single output, named image, an IMAGE tensor.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", then restart. One dependency (replicate>=1.0.7), no model files. Set your token as an environment variable before starting ComfyUI:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• The filter, not the node, decides. The community's standing complaint is that Nano Banana is powerful but "unimaginative" - a chunk of that is the safety layer rerouting or refusing things you'd expect to pass. If a prompt keeps dying, it's Google's policy, and no pack can fix it.
• Capacity waits. It's popular, so "model is at capacity" errors happen. That's what allow_fallback_model is for.
• Cost math. Per-image pricing adds up faster than you think when you're iterating on a batch. Do your exploration at 1K/2K and treat 4K as the final render.