Generative-Media-Skills

Exposes a unified JSON Schema interface to 30+ image generation models (Midjourney v7, Flux Kontext, DALL-E 3, Stable Diffusion XL) through the muapi-cli wrapper layer. The system maps high-level generation requests to model-specific API calls via schema_data.json lookup tables, handling authentication, parameter normalization, and async polling for result retrieval without requiring developers to learn individual model APIs.

The Nano-Banana skill encodes professional design reasoning into optimized prompt templates and multi-step generation workflows. When an agent requests a logo, UI mockup, or portrait pack, the system decomposes the creative intent into structured parameters (brand guidelines, design principles, identity constraints), executes generation with reasoning-aware prompts, and applies post-processing rules specific to the domain (e.g., identity-lock for portrait consistency).

The platform utilities handle file uploads to muapi.ai cloud storage, managing authentication, chunked uploads for large files, and result file retrieval. The system supports reference image uploads (for style transfer, inpainting), source video uploads (for extension), and audio uploads (for voice cloning). Files are stored with expiration policies and accessed via signed URLs returned in generation results.

The system supports batch generation of multiple media assets in parallel through async task submission and result polling. Agents submit a batch of generation requests (e.g., 10 image variations, 5 video clips), receive task IDs immediately, and poll for results asynchronously. The system aggregates results as they complete and returns a batch result object with per-item status and metadata.

The Cinema Director skill translates high-level cinematic direction (shot type, camera movement, mood, pacing) into optimized prompts for video generation models (Seedance 2.0, Kling 3.0). The system maps directorial concepts (e.g., 'Dutch angle establishing shot') to model-specific parameter sets, manages multi-shot composition, and handles async video rendering with progress polling and result validation.

The Seedance 2 skill extends existing video clips by generating additional frames while maintaining temporal coherence and motion continuity. The system accepts a source video, target duration, and motion direction parameters, then uses Seedance 2.0's frame interpolation engine to synthesize intermediate frames that preserve object trajectories and scene consistency. Async polling monitors generation progress and validates output frame count and quality metrics.

Integrates Suno AI and other text-to-audio models through muapi-cli to generate music, voiceovers, and sound effects from text descriptions. The system supports voice cloning (map text to specific speaker identity), style control (genre, mood, instrumentation), and async audio rendering with format conversion. Audio files are polled asynchronously and returned with metadata (duration, sample rate, codec).

Exposes 19 structured generation and editing tools through the Model Context Protocol (MCP) server interface. Running `muapi mcp serve` starts an MCP server that publishes JSON Schema definitions for each tool, enabling AI agents (Claude Code, Cursor, Gemini) to discover, validate, and call generation functions directly without shell script execution. The system handles schema validation, async polling orchestration, and result streaming back to the agent.

Implements a robust async polling pattern for long-running media generation tasks. When a generation request is submitted, the system returns a task ID immediately and polls the muapi.ai backend at exponential backoff intervals (1s, 2s, 4s, 8s...) until the result is ready. The check-result.sh script handles polling orchestration, timeout management, and result validation, enabling agents to submit batch generation requests without blocking.

The edit-image.sh script enables semantic image editing through natural language prompts. Users describe desired edits (e.g., 'change the sky to sunset orange', 'remove the person from the background') and the system uses vision-language models to understand the edit intent, apply targeted modifications, and preserve unrelated image regions. Editing is performed through inpainting or outpainting depending on the edit scope.

The Workflow skill enables agents to compose complex multi-step generation pipelines as directed acyclic graphs (DAGs). Agents define nodes (generation tasks), edges (data flow), and execution parameters, then submit the workflow for orchestration. The system executes nodes in dependency order, handles intermediate result passing, and manages async polling across all nodes. Workflow results are aggregated and returned with execution traces.

The system abstracts function calling across multiple AI model providers (OpenAI, Anthropic, Ollama) through a unified schema-based registry. Each generation tool is registered with JSON Schema definitions that are compatible with OpenAI function calling, Anthropic tool_use, and Ollama native bindings. The system automatically translates between provider-specific function calling formats and executes the underlying muapi-cli commands.