civitai

Civitai routes generation requests through an orchestrator service that abstracts multiple backend implementations (ComfyUI, ImageGen, TextToImage) via a unified schema-based interface. The generation.router.ts exposes endpoints that validate requests against generation.schema.ts, then dispatch to orchestrator.service.ts which selects the appropriate backend based on model type and generation parameters. This enables seamless switching between generation backends without frontend changes and supports complex workflows like upscaling and inpainting through ComfyUI's node-graph architecture.

Civitai maintains a search and indexing system that ingests model metadata, descriptions, and tags into Elasticsearch for semantic and full-text search. The system uses background jobs (via the background jobs infrastructure) to asynchronously index model updates, with a search_index_update_queue_action enum tracking indexing state. Search queries hit Elasticsearch to return ranked model results with filtering by model type, base model, and creator. The architecture supports real-time index updates through a queue-based pattern that decouples model updates from search index synchronization.

Civitai implements an article system that allows creators to publish guides, tutorials, and documentation about their models. Articles support rich text formatting, image attachments, and links to associated models. The system tracks article metadata (title, author, creation date, view count) and enables discovery through search and recommendations. Articles serve as a knowledge base for the community and help creators document their models' usage and capabilities. The architecture integrates articles with the model system, enabling cross-linking and discovery.

Civitai implements authentication and session management using NextAuth or similar, with support for multiple auth providers (OAuth, email/password). The system manages user sessions, permissions, and feature flags that control feature rollout and A/B testing. Feature flags are evaluated at request time to enable/disable features per user or user cohort. The architecture integrates authentication with the database schema to track user identity, permissions, and feature access. Session management handles concurrent logins and token refresh.

Civitai implements a notification system that alerts users about relevant events (model updates, comments, bounty awards, etc.). The system respects user notification preferences (email, in-app, push) and allows users to customize notification frequency and types. Notifications are generated by background jobs that monitor for triggering events and queue notification delivery. The architecture integrates with the database to track notification state (read/unread) and user preferences. Notifications can be delivered through multiple channels (email, in-app, push notifications).

Civitai implements a cosmetic shop where users can purchase cosmetics (badges, profile themes, etc.) using Buzz. The system manages cosmetic inventory, user cosmetic ownership, and cosmetic application to user profiles. Cosmetics are displayed on user profiles and in leaderboards, serving as status symbols and incentives for engagement. The architecture integrates with the Buzz economy for cosmetic pricing and purchase tracking. Cosmetics can be limited-edition or seasonal, creating scarcity and urgency.

Civitai implements a Redis-based caching strategy that caches frequently accessed data (models, user profiles, leaderboards) to reduce database load. The system uses cache keys with TTLs (time-to-live) and implements cache invalidation patterns (tag-based, event-based) to keep caches fresh. Different data types have different cache strategies: models are cached long-term, user profiles medium-term, leaderboards short-term. The architecture integrates caching at multiple layers (API responses, database queries, computed values) to maximize hit rates.

Civitai implements a background job system (using a job queue like Bull or similar) that handles async tasks like image processing, search indexing, notification delivery, and metrics collection. Jobs are queued by the main application and processed by background workers, enabling long-running tasks without blocking user requests. The system tracks job status (pending, processing, completed, failed) and retries failed jobs with exponential backoff. Metrics are collected asynchronously and aggregated for analytics and monitoring.

Civitai implements an image ingestion and moderation pipeline that processes uploaded images through automated NSFW detection and manual moderation workflows. The system assigns NSFW levels (0-3 scale) to images based on detection results, with browsing controls allowing users to filter content by their preferred NSFW threshold. The pipeline integrates with the moderation tools system and supports a 'New Order Moderation Game' for gamified community moderation. Images are stored with metadata linking to posts and models, enabling content filtering at query time.

Civitai implements a generation queue system (Queue.tsx, QueueItem.tsx) that tracks pending and in-progress generation requests with real-time status updates. The system uses a DataGraph architecture (Generation V2) for state management, allowing the frontend to subscribe to generation status changes and receive updates as images are generated. The queue persists in the database via Prisma, and background jobs process queued requests through the orchestrator. Users can view their queue, cancel requests, and receive notifications when generations complete.

Civitai implements a Buzz economy system where creators earn Buzz tokens from model downloads, image generation usage, and community engagement. The system integrates with payment processors to allow users to purchase Buzz, and creators can withdraw earnings. The architecture includes a Creator Program that tracks creator metrics (downloads, usage, earnings) and a payment system that handles transactions. Buzz is used for platform features like generation credits, cosmetics, and bounties, creating a closed-loop economy that incentivizes content creation.

Civitai manages model versions and associated files through a versioning system that tracks model updates, file hashes, and download metadata. The CivitaiLink integration enables external tools (like ComfyUI, Automatic1111) to download models directly from Civitai using a standardized protocol. The system stores file metadata (hash, size, format) and tracks downloads per version, enabling analytics and version-specific metrics. Model versions are linked to base models and include metadata about compatibility and improvements.

Civitai implements social features including clubs (user groups), leaderboards (ranked by downloads/engagement), reactions (likes/favorites), comments, and reviews on models and images. The system tracks user engagement metrics (reactions, comments, follows) and aggregates them for leaderboards and recommendations. Clubs enable community organization around shared interests, with club-specific content and member management. The architecture uses real-time signals to update engagement metrics and feed rankings.

Civitai implements a collection system that allows users to create curated collections of models and images, organizing content by theme or use case. Collections are discoverable through search and recommendations, and can be shared with other users. The system tracks collection metadata (name, description, creator, member count) and enables collaborative curation. Collections serve as a content organization layer above individual models, enabling users to discover related models and create themed bundles.

Civitai implements a model training system that allows users to train custom models (LoRAs, embeddings) using their own datasets. The system manages dataset uploads, training job configuration, and orchestration of training workloads. Users specify training parameters (learning rate, epochs, base model) and the system queues training jobs for execution on distributed workers. Training results are stored as new model versions and can be shared with the community. The architecture integrates with the generation system to enable immediate testing of trained models.

Civitai implements a bounty system where users can post bounties (e.g., 'create a model for X style') with Buzz rewards, incentivizing creators to fulfill specific requests. The system manages bounty lifecycle (posting, bidding, completion, award), tracks bounty metadata (description, reward, deadline), and enables dispute resolution. Bounties serve as a mechanism to direct community effort toward high-demand models and features. The architecture integrates with the Buzz economy for reward distribution and creator reputation tracking.