Open-Generative-AI

Generates images from text prompts by routing requests through a unified MuapiClient that abstracts 50+ image generation models (Flux, DALL-E, Midjourney, Stable Diffusion variants). The ImageStudio component dynamically renders UI controls (resolution pickers, style selectors, guidance scales) based on each model's input schema defined in the models.js registry, eliminating hardcoded form logic and enabling new models to be added without frontend changes.

Generates videos from text prompts or image inputs by submitting requests to Muapi backend and polling for completion status via a job ID. The VideoStudio component manages the generation lifecycle: submission → polling loop (with configurable intervals) → result retrieval. Supports 30+ video models including Kling, Sora, Veo, and Runway, with model-specific parameter schemas (duration, aspect ratio, motion intensity) rendered dynamically. Pending jobs are persisted in localStorage and can be resumed across browser sessions.

Provides unrestricted access to image and video generation models without applying content filters, safety checks, or moderation policies. The application does not implement NSFW detection, prompt filtering, or output validation; all generation requests are passed directly to Muapi backend models without modification. This design prioritizes user freedom and creative expression over content moderation, making it suitable for unrestricted artistic and experimental use cases.

Provides a MuapiClient class that abstracts all communication with the Muapi backend, exposing unified methods for image generation (generateImage), video generation (generateVideo), lip-sync (generateLipSync), and job polling (pollJobStatus). The client handles request formatting, response parsing, error handling, and retry logic. It supports multiple model families (Flux, DALL-E, Midjourney, Kling, Sora, etc.) through a single interface, eliminating the need for model-specific API clients. All requests include the x-api-key header from localStorage for BYOK authentication.

Uses Tailwind CSS utility classes for styling all UI components across web and desktop shells, providing a consistent design system with responsive breakpoints (mobile, tablet, desktop) and dark mode support. The styling system is defined in tailwind.config.js and applied via PostCSS (postcss.config.js). All studio components (ImageStudio, VideoStudio, etc.) use Tailwind classes for layout, spacing, colors, and typography, enabling rapid UI iteration and consistent theming across platforms.

Generates lip-synced video animations by accepting an audio file (MP3, WAV) and a reference video or image, then using Muapi's lip-sync models to align mouth movements with audio phonemes. The LipSyncStudio component handles audio upload, model selection (supporting multiple lip-sync architectures), and parameter tuning (sync intensity, mouth shape variation). Results are persisted in generation history with audio metadata for reproducibility.

Generates cinematic video sequences by combining a prompt builder (CinemaPromptBuilder) that structures narrative, camera movement, lighting, and composition into optimized prompts, with an asset library (CinemaAssetLibrary) containing pre-built cinematography templates (Dutch angle, tracking shot, crane shot, etc.). The Cinema Studio routes these structured prompts to video models optimized for cinematic output, with support for multi-shot sequences and scene composition. Prompts are engineered to maximize model understanding of camera techniques and visual storytelling.

Implements a BYOK authentication model where users provide their own Muapi.ai API key via an AuthModal component, which is then stored in localStorage and used in the x-api-key header for all subsequent API requests. No user accounts, billing, or backend authentication are managed by the application; the API key is the sole credential. Key is persisted across browser sessions and can be cleared via settings. This design eliminates backend infrastructure requirements and gives users full control over API usage and billing.

Persists all generation results (images, videos, metadata) in localStorage under the muapi_history key, creating a local index of past creations that survives browser restarts. Each generation record includes job ID, model used, input parameters, output URL, and timestamp. Users can browse history, filter by model or date, and reuse parameters from past generations without re-entering them. Pending jobs are tracked separately and can be resumed by polling their job IDs.

Implements a monorepo structure where core generation logic and UI components (ImageStudio, VideoStudio, LipSyncStudio, CinemaStudio) are packaged as a reusable React component library in packages/studio, which is then consumed by three independent frontend shells: a Next.js web application (app/), a Vite/Vanilla-JS development environment (src/), and an Electron desktop application (electron/). Each shell uses the same MuapiClient and studio components but provides different distribution and deployment models (cloud-hosted, local development, cross-platform desktop).

Maintains a centralized models.js registry that defines all available image, video, lip-sync, and cinematic models with their input schemas (resolution options, style categories, guidance scales, etc.). Each model entry includes metadata (name, provider, capabilities) and a JSON schema describing required and optional parameters. The studio components (ImageStudio, VideoStudio, etc.) parse these schemas at runtime and dynamically generate UI controls (dropdowns, sliders, text inputs) without hardcoding model-specific logic. Adding a new model requires only updating models.js; the UI adapts automatically.

Provides an UploadPicker component that manages image and video uploads for image-to-video and reference-based generation tasks. Uploaded files are cached locally in localStorage (with file hash-based deduplication) to prevent redundant uploads of the same file. The component tracks upload history, displays previews, and manages file size validation before submission. Cached uploads are reused across multiple generation requests, reducing bandwidth and API overhead.

Packages the Open Generative AI application as a cross-platform Electron desktop application (electron/ directory) that runs on macOS, Windows, and Linux. The Electron main process handles window management, file system access for saving generated images/videos, and OS-level integrations (native menus, drag-and-drop, system tray). The Vite/Vanilla-JS frontend (src/) is bundled with Electron and communicates with the Muapi backend via the same MuapiClient used by web shells. Desktop app provides offline-capable UI and direct file system access without browser sandbox restrictions.