ComfyUI-Gemini-Conversation-Canvas

ComfyUI Gemini Conversation Canvas

Multi-turn conversational image editing for ComfyUI, powered by Google Gemini's native image generation.

Unlike traditional image generation APIs that are fire-and-forget, Gemini maintains conversation context across turns — each edit builds on the previous result with full scene coherence. This node suite brings that capability into ComfyUI with persistent sessions, edit chaining, and a visual gallery.

Why Gemini Conversations?

Most image APIs (DALL-E, Midjourney, Flux) are single-shot: you send a prompt, you get an image. There's no memory between generations.

Gemini is different. Its image generation is native to the language model, so it remembers what it generated and can reason about modifications:

Turn 0: "A cozy cafe interior with warm lighting"        → initial image
Turn 1: "Add rain visible through the windows"           → same cafe, now with rain
Turn 2: "Make it evening, warm lamp light inside"         → same scene, evening mood
Turn 3: "Add a cat sleeping on the corner chair"          → cat added coherently

Each turn preserves the scene, characters, and style — no need for ControlNet, IP-Adapter, or manual inpainting.

Included Nodes

1. Gemini Session (Start)

Creates a new conversation and generates the initial image.

| Input | Type | Description | |-------|------|-------------| | prompt | STRING | Text prompt for the initial image | | session_name | STRING | Name for saving/loading this session | | model | COMBO | Gemini model selection | | aspect_ratio | COMBO | Aspect ratio (1:1, 16:9, 9:16, etc.) | | input_image | IMAGE (optional) | Source image for image-to-image editing | | api_key | STRING (optional) | Google AI API key |

Outputs: session, image, text, turn_count

2. Gemini Edit Turn

Sends an edit instruction to an existing conversation. Chainable — connect multiple in sequence for multi-step edits in a single workflow run.

| Input | Type | Description | |-------|------|-------------| | session | SESSION | From Session Start, Load, or previous Edit Turn | | instruction | STRING | Natural language edit instruction | | reference_image | IMAGE (optional) | Reference image to include with the instruction | | aspect_ratio | COMBO | Can change between turns |

Outputs: session, image, text, turn_count

3. Gemini Session (Save)

Persists the conversation to disk. Session history and all turn images are saved so you can resume later.

4. Gemini Session (Load)

Loads a previously saved session from disk. Connect to an Edit Turn node to continue the conversation.

5. Gemini Session Gallery

Outputs all turn images from a session as a batch tensor for visual comparison.

Outputs: images (batch), turn_log (text summary), turn_count

Usage

Basic: Single-Run Multi-Turn Editing

Chain multiple Edit Turn nodes for iterative refinement in one workflow execution:

[Session Start]                    [Edit Turn]               [Edit Turn]            [Save]
  prompt: "a red sports car"  →   "change to blue"     →   "add sunset sky"   →    💾
  ↓ image_0                        ↓ image_1                 ↓ image_2

Advanced: Cross-Run Session Persistence

Run 1: Generate and save

[Session Start] → [Edit Turn] → [Session Save]

Run 2: Load, add more edits, save again

[Session Load] → [Edit Turn] → [Edit Turn] → [Session Save]

Anytime: View all turns

[Session Load] → [Session Gallery] → [Preview Image]

Image-to-Image Editing

Connect an existing image to the input_image input on Session Start:

[Load Image] → [Session Start: "remove the background and replace with a beach"] → image

API Key Setup

Provide your Google AI API key via one of these methods (checked in order):

1. Direct input — paste into the api_key field on Session Start or Session Load
2. Environment variable — set GEMINI_API_KEY in your system environment
3. Key file — create a gemini_api_key.txt file in your ComfyUI root directory

Get a free API key at Google AI Studio.

Output Files

All session data is stored under your ComfyUI output directory:

| Path | Description | |------|-------------| | output/gemini_sessions/{name}/session.json | Conversation history and metadata | | output/gemini_sessions/{name}/turn_000.png | Initial generated image | | output/gemini_sessions/{name}/turn_001.png | Image after first edit | | output/gemini_sessions/{name}/turn_NNN.png | Image after Nth edit |

Installation

Clone this repository into your ComfyUI/custom_nodes/ directory:

cd ComfyUI/custom_nodes/
git clone https://github.com/jeremieLouvaert/ComfyUI-Gemini-Conversation-Canvas.git
pip install -r ComfyUI-Gemini-Conversation-Canvas/requirements.txt

Restart ComfyUI.

Dependencies

• google-genai >= 1.0.0 — Google's Gen AI Python SDK
• filelock — concurrent-safe file operations
• Pillow — image conversion
• PyTorch — already present in any ComfyUI installation

Supported Models

| Model | ID | Notes | |-------|----|-------| | Gemini 2.5 Flash Image | gemini-2.5-flash-preview-image-generation | Fast, affordable, production-ready | | Gemini 2.0 Flash Image | gemini-2.0-flash-preview-image-generation | Previous generation |

Tips

• Be specific with edit instructions. "Add a red umbrella to the person on the left" works better than "add umbrella."
• The model remembers everything. You can reference elements from previous turns: "make the cat from turn 3 larger."
• Aspect ratio can change per turn. Start in 1:1, then switch to 16:9 for a cinematic crop.
• Save frequently. Sessions persist across ComfyUI restarts, so save after important edits.
• Use the Gallery node to compare all turns side by side and verify the edit progression.

License

MIT