Google Gemini Pipeline

Highly optimized Google Gemini pipeline with advanced image and video generation capabilities, intelligent compression, and streamlined processing workflows.

Features

• Optimized asynchronous API calls for maximum performance
• Intelligent model caching with configurable TTL, refreshed at once when model valves change
• Streamlined dynamic model specification with automatic prefix handling
• Smart streaming response handling with safety checks
• Retries of temporary API errors for streaming and non-streaming requests
• Advanced multimodal input support (text and images)
• Unified image generation and editing with Gemini 2.5 Flash Image Preview
• Image editing across turns: earlier generated and uploaded images of a saved chat are sent with the edit request (the newest kept within the image limit, each image once; the pipeline reads only files of the requesting user)
• Nano Banana image models (gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-nano-banana-2.1)
• Extra image generation model IDs configurable without a code change
• Interim thought images skipped, so each generated image is uploaded once (the last one is kept if no final image arrives)
• Native tools and URL context left out for image models, which do not support them
• Search grounding left out for image models without Search support (gemini-2.5-flash-image, gemini-3.1-flash-lite-image)
• Intelligent image optimization with size-aware compression algorithms
• Automated image upload to Open WebUI with robust fallback support
• Generated images and videos linked in the answer for API clients
• Optimized text-to-image and image-to-image workflows
• Non-streaming mode for image generation to prevent chunk overflow
• Progressive status updates, each closed when a request ends, fails or is stopped
• Consolidated error handling and comprehensive logging
• Seamless Google Generative AI and Vertex AI integration
• Advanced generation parameters (temperature, max tokens, etc.)
• Configurable safety settings with environment variable support
• Military-grade encrypted storage of sensitive API keys
• Intelligent grounding with Google search integration
• Vertex AI Search grounding for RAG
• Native tool calling through Open WebUI's tool loop (built-in, workspace, MCP, OpenAPI, terminal and direct tools, tool approval)
• Thought signatures carried across tool rounds and turns
• Tool calls returned to API clients as OpenAI tool_calls (streaming and non-streaming)
• Google Search grounding together with function calling on Gemini 3
• URL context grounding for specified web pages
• Unified image processing with consolidated helper methods
• Optimized payload creation for image generation models
• Configurable image processing parameters (size, quality, compression)
• Flexible upload fallback options and optimization controls
• Configurable thinking levels for Gemini 3 models with model-specific validation
• Thinking summaries stripped from replayed history to save tokens (configurable)
• Configurable thinking budgets (0-32768 tokens) for Gemini 2.5 models
• Configurable image generation aspect ratio (1:1, 16:9, etc.) and resolution (1K, 2K, 4K)
• Model whitelist for filtering available models
• Additional model support for SDK-unsupported models
• Video generation with Google Veo models (Veo 3.1, 3, 2)
• Configurable video generation parameters (aspect ratio, resolution, duration)
• Asynchronous video generation with progressive polling status updates
• Automatic video upload to Open WebUI with chat file attachments
• Image-to-video generation support for Veo models
• Negative prompt and person generation controls for video
• Token usage reporting for streaming and non-streaming responses
• Usable as task model for title, tag and follow-up generation

Tools

• Grounding with Google search with google_search_tool.py filter
• Grounding with Vertex AI Search with vertex_ai_search_tool.py filter

GitHub

Open-WebUI-Functions

License

Apache License 2.0