Banana Studio

Banana Studio for ComfyUI

Banana Studio is a ComfyUI custom node pack centered on Gemini image generation plus a small set of prompt-building and workflow utility nodes.

Included Nodes

• Banana Studio: generate images with Gemini image models, with optional image inputs and batch support
• Banana Prompt: build a structured prompt from named sections
• Z-Image Prompt: build an image prompt with framing, subject, lighting, style, and constraints sections
• Prompt Editor: expand prompt variables and strip inline comments
• String Concat: join strings with an optional separator
• Single Parameter Dispatcher: emit per-batch values for parameter sweeps
• Format Path: expand path templates with %date:...% formatting for downstream save nodes

Installation

Clone this repository into ComfyUI/custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/tjcccc/comfyui_banana_studio.git

Restart ComfyUI after installation.

Gemini API Key

Create a local config.ini file in the repository root:

[auth]
GEMINI_API_KEY = your_gemini_api_key_here

You can copy config.ini.example to config.ini and replace the placeholder value.

API key resolution order in the Banana Studio node:

1. config.ini
2. The node api_key input
3. Raise an error if neither is set

The api_key input remains available for temporary local testing, but config.ini takes priority when both are present.

Main Node Inputs

The Banana Studio node supports:

• Gemini image models: gemini-3.1-flash-image-preview, gemini-3-pro-image-preview, gemini-2.5-flash-image
• text prompt input
• optional image inputs for image-conditioned generation
• batch generation
• aspect ratio, resolution, temperature, top_p, thinking level, seed, and proxy controls

Node Usage

Banana Studio

Use this as the main Gemini image generation node.

• Set prompt to your text instruction.
• Set model to the Gemini image model you want to use.
• Set batch_size to generate multiple images in one run.
• Connect up to 6 optional IMAGE inputs when you want image-conditioned generation or reference images.
• Use aspect_ratio and resolution to control the output format.
• Use temperature, top_p, thinking_level, and seed when you want more control over sampling behavior.
• Leave api_key empty if you already configured config.ini.
• Set proxy only when you really need it. Large inline image uploads are more likely to fail through unstable proxies.

Outputs:

• images: generated image tensor output for downstream image nodes
• logs: text summary with generation status and token usage

Typical use:

1. Build a prompt directly or with Banana Prompt.
2. Connect that string to Banana Studio.
3. Optionally connect one or more input images.
4. Feed the returned images into preview, save, or post-processing nodes.

Banana Prompt

Use this when you want a structured cinematic prompt assembled from labeled sections.

• medium_or_tech is the main required field. Use it for medium, camera, lens, and composition context.
• Fill optional sections such as identify_reference, subject_or_presence, action_or_state, environment, clothing_body, and final_style only when needed.
• If only medium_or_tech is filled, the node returns that value directly.
• If any optional section is filled, the node formats all non-empty sections into a readable multi-section prompt.

Output:

• prompt: the assembled text prompt string

Good fit:

• character consistency workflows
• portrait, fashion, editorial, and cinematic prompt writing
• reusable prompt blocks where each section has a clear role

Z-Image Prompt

Use this when you want a separate structured image-prompt helper based on framing, subject, scene, lighting, and constraints.

• composition_or_framing is required and usually defines the shot first.
• Add subject_or_identity, wardrobe_or_appearance, environment_or_scene, lighting, mood_or_style_or_quality, and constraints as needed.
• The constraints field is useful for hard requirements like no text, no watermark, or photorealistic.
• If only the required field is filled, the node returns it directly.
• Otherwise it builds a bracketed multi-section prompt.

Output:

• prompt: the assembled text prompt string

Good fit:

• image prompting with explicit framing and lighting control
• prompts that need hard exclusions or quality constraints
• workflows outside the Banana Studio node

Prompt Editor

Use this to turn prompt code into a final prompt string.

• Put your editable prompt source into prompt_code.
• The parser removes comments, extracts variables, resolves substitutions, and returns the final prompt text.
• This is useful when you want prompt templates instead of manually rewriting full prompts each time.

Output:

• prompt: the parsed final prompt string

Good fit:

• reusable prompt templates
• variable-driven prompt authoring
• prompt text that should stay maintainable over time

Example:

# Variables
character = cinematic portrait of a young woman
location = neon street at night
style = soft rim light, shallow depth of field, highly detailed

# Prompt body
prompt = {
A realistatic photo.
{character}. She is in a {location}.
{style}.
}

Expected output:

A realistatic photo.
cinematic portrait of a young woman. She is in a neon street at night.
soft rim light, shallow depth of field, highly detailed.

String Concat

Use this to merge two text blocks before sending them into prompt-consuming nodes.

• string_a is required.
• string_b is optional.
• separator defaults to a blank line.
• Escape sequences such as \n are decoded when the input is plain ASCII text, so you can type separators like \n, \n\n, or , directly.

Output:

• a + b: the merged string

Good fit:

• joining a base prompt and a style suffix
• appending reusable negative or constraint text
• building prompts from smaller blocks

Single Parameter Dispatcher

Use this for simple batch-time parameter sweeps.

• value is the starting value.
• delta is the amount added for each next item.
• max_value clamps the sequence when increasing or decreasing.
• batch sets the total queue length.
• reset_mark forces the sequence to rebuild when changed.
• output_tag_format lets you generate a label string using placeholders: %current_value%, %index%, %delta%, %max%, %batch%

Outputs:

• dispatch_value: current numeric value for the current batch item
• index: 1-based position in the batch
• output_tag: formatted text tag using the placeholders above
• log: debug text for the current dispatched value

Good fit:

• varying guidance-like numeric inputs across queued runs
• generating per-image labels for save paths or metadata
• simple sweep workflows without a larger parameter system

Format Path

Use this when a downstream save node accepts a path or filename-prefix string and you want the path to be generated from a reusable template.

• path_template supports custom date tokens such as %date:yyyy-MM-dd%, %date:yyyyMMdd-HHmmss%, and %date:yyyy/MM/dd/ComfyUI%.
• Connect formatted_path to any save node input that accepts a STRING path or filename prefix.

Output:

• formatted_path: the formatted path string

Notes

• If you use a proxy, upload timeouts can happen before Gemini returns a response, especially when sending large inline image payloads.
• The repository keeps config.ini git-ignored so local keys do not end up in commits.
• banana_studio.py resolves the API key before making the Gemini request, so saved workflows no longer need to store the real key when config.ini is present.

Development

Relevant local validation for this repository:

python3 -m py_compile banana_studio.py gemini_service.py path_format.py format_path.py __init__.py

Publish

To publish this node to the Comfy Registry manually:

pip install comfy-cli
comfy node publish

You will need a Registry publishing API key for publisher tjcccc.

This repository also includes publish_action.yml. To enable automatic publishing on version bumps:

1. Create a repository secret named REGISTRY_ACCESS_TOKEN.
2. Put your Comfy Registry publishing API key in that secret.
3. Bump version in pyproject.toml and push to main.

Version

Current project version: 0.7.0