ComfyUI - fal.ai Nano Banana Pro Nodes 🍌

Custom ComfyUI nodes for Nano Banana Pro (a.k.a. Nano Banana 2 / Gemini 3 Pro Image) via the fal.ai API.

Two nodes included:

| Node | Description | |------|-------------| | Nano Banana Pro (Text to Image) 🍌 | Generate images from text prompts | | Nano Banana Pro (Image Edit) 🍌 | Edit existing images with natural language |

Installation

1. Clone or copy this folder into your ComfyUI custom_nodes directory:

   cd ComfyUI/custom_nodes
   git clone <this-repo> comfyui-fal-nano-banana-pro
   

2. Install the dependency (only requests — PIL/torch/numpy are already in ComfyUI):

   pip install requests
   

3. Set your fal.ai API key as an environment variable before launching ComfyUI:

   export FAL_KEY="your-fal-api-key-here"
   

   Get a key at: https://fal.ai/dashboard/keys

4. Restart ComfyUI. The nodes will appear under fal.ai/Nano Banana Pro.

---

Safety Tolerance / Censorship — How It Works

This is likely the reason you're here, so here's the full picture:

Text-to-Image (fal-ai/nano-banana-pro)

• No safety_tolerance parameter is exposed on this endpoint.
• Content filtering is applied server-side by a combination of Google's Gemini model safety systems and fal.ai's infrastructure.
• You cannot reduce or disable filtering for text-to-image via the API.
• If a prompt is filtered, you'll typically get an empty images array or an error.

Image Editing (fal-ai/nano-banana-pro/edit)

• Exposes safety_tolerance as an API-only parameter (not in fal.ai's playground UI).
• Scale of 1 to 6:
  • 1 = Most strict (blocks the most content)
  • 6 = Least strict (blocks the least content)
  • Default = 4
• This is included as a dropdown in the ComfyUI node.
• Important: Even at level 6, the underlying Gemini model still has Google's built-in safety filters. The safety_tolerance parameter adjusts fal.ai's additional filtering layer on top of the model — it does not remove the model's own refusals.

What this means in practice

• Setting safety_tolerance to 6 on the edit node will let more borderline content through fal.ai's filter, but Google's model-level safety can still refuse to generate certain content.
• The text-to-image node has no equivalent knob — you're subject to whatever the default filtering is.
• If you get empty results, check the description output — it sometimes contains the model's refusal reason.

---

Node Parameters

Text to Image

| Parameter | Type | Default | Notes | |-----------|------|---------|-------| | prompt | string | — | Required | | num_images | int | 1 | 1–4 | | aspect_ratio | enum | 1:1 | 10 options | | output_format | enum | png | png/jpeg/webp | | resolution | enum | 1K | 1K/2K/4K (4K costs 2x) | | seed | int | -1 | -1 = random | | limit_generations | bool | false | Force exactly 1 image per generation round | | enable_web_search | bool | false | Let the model search the web (+$0.015) | | timeout | int | 300 | Max seconds to wait |

Image Edit

All of the above, plus: | Parameter | Type | Default | Notes | |-----------|------|---------|-------| | image | IMAGE | — | Required, primary input image | | image_2 | IMAGE | — | Optional second reference image | | image_3 | IMAGE | — | Optional third reference image | | safety_tolerance | enum | 4 | 1 (strict) → 6 (permissive). API-only. |

---

Pricing (fal.ai)

• $0.15 per image at 1K/2K
• $0.30 per image at 4K
• +$0.015 if web search is enabled

---

License

MIT