invoke-ai/invokeai

Invoke is a leading creative engine for Stable Diffusion models, empowering professionals, artists, and enthusiasts to generate and create visual media using the latest AI-driven technologies. The solution offers an industry leading WebUI, and serves as the foundation for multiple commercial products.

Structural Verdict

How Data Flows Through the System

User interactions in the web UI trigger Redux actions that build node graphs, which are sent to the FastAPI backend where the Invoker executes them through the invocation system, ultimately generating images stored via the ImageService

1. User Input — User configures generation parameters in React UI components
2. Graph Building — Frontend builds execution graphs from UI state using graph builders
3. API Request — Redux RTK Query dispatches graph to FastAPI /queue/enqueue endpoint
4. Session Processing — SessionProcessor queues and executes invocation workflows
5. Model Loading — ModelManagerService loads AI models based on invocation requirements
6. AI Inference — Invocations execute diffusion pipeline steps using loaded models
7. Image Storage — Generated images are saved via ImageService and metadata stored
8. WebSocket Updates — Progress and results sent to frontend via WebSocket events

System Behavior

How the system actually operates at runtime — where data accumulates, what loops, what waits, and what controls what.

Data Pools

Feedback Loops

• Generation Progress (polling, balancing) — Trigger: Generation starts. Action: WebSocket sends progress updates to frontend. Exit: Generation completes.
• Model Loading Retry (retry, balancing) — Trigger: Model loading fails. Action: ModelManager retries loading with fallback strategies. Exit: Model loads successfully or max retries exceeded.
• Session Queue Processing (polling, balancing) — Trigger: Sessions exist in queue. Action: SessionProcessor picks up and executes next session. Exit: Queue is empty.

Delays & Async Processing

• Model Loading (async-processing, ~variable (seconds to minutes)) — Generation requests wait for models to load into memory
• AI Inference (async-processing, ~variable (seconds to minutes)) — Users wait for diffusion process to complete
• Queue Processing (queue-drain, ~variable) — Multiple users' requests are processed sequentially

Control Points

• Model Precision (env-var) — Controls: Whether models run in fp16 or fp32 precision
• Seamless Axes (runtime-toggle) — Controls: Which axes (x/y) to apply seamless tiling. Default: empty list
• Control Weight (threshold) — Controls: Strength of ControlNet guidance in generation. Default: 1.0
• Queue Prepend (feature-flag) — Controls: Whether to prioritize new requests over queued ones. Default: true

Technology Stack

Key Components

• InvokeAIUI (component) — Main React application component that orchestrates the entire UI invokeai/frontend/web/src/app/components/InvokeAIUI.tsx
• ApiDependencies (service) — Initializes and provides all backend services to FastAPI invokeai/app/api/dependencies.py
• Invoker (service) — Core orchestration service that executes invocation workflows invokeai/app/services/invoker.py
• ControlNetInvocation (class) — Handles ControlNet model invocations for guided image generation invokeai/app/invocations/controlnet.py
• MainModelDefaultSettings (class) — Configuration class for default model parameters like resolution and CFG scale invokeai/backend/model_manager/configs/main.py
• ModelManagerService (service) — Manages AI model loading, caching, and lifecycle invokeai/app/services/model_manager/model_manager_default.py
• ImageService (service) — Handles image CRUD operations and metadata management invokeai/app/services/images/images_default.py
• SessionProcessor (service) — Processes queued generation sessions and manages execution state invokeai/app/services/session_processor/session_processor_default.py
• buildControlLayersGraph (function) — Constructs the node graph for control layer-based image generation invokeai/frontend/web/src/features/nodes/util/graph/generation/buildControlLayersGraph.ts
• cropImageModalApi (module) — State management for the image cropping modal using nanostores invokeai/frontend/web/src/features/cropper/store/index.ts

Configuration

pins.json (json)

• python (string, unknown) — default: 3.12
• torchIndexUrl.win32.cuda (string, unknown) — default: https://download.pytorch.org/whl/cu128
• torchIndexUrl.linux.cpu (string, unknown) — default: https://download.pytorch.org/whl/cpu
• torchIndexUrl.linux.rocm (string, unknown) — default: https://download.pytorch.org/whl/rocm6.3
• torchIndexUrl.linux.cuda (string, unknown) — default: https://download.pytorch.org/whl/cu128

invokeai/app/api/routers/boards.py (python-pydantic)

• board_id (str, unknown) — default: Field(description="The id of the board that was deleted.")
• deleted_board_images (list[str], unknown) — default: Field(
• deleted_images (list[str], unknown) — default: Field(description="The names of the images that were deleted.")

invokeai/app/api/routers/images.py (python-pydantic)

• width (int, unknown) — default: Field(..., gt=0)
• height (int, unknown) — default: Field(..., gt=0)

invokeai/app/api/routers/images.py (python-pydantic)

• image_dto (ImageDTO, unknown) — default: Body(description="The image DTO")
• presigned_url (str, unknown) — default: Body(description="The URL to get the presigned URL for the image upload")

Science Pipeline

1. Load base model — ModelManagerService loads diffusion model weights into GPU memory invokeai/app/services/model_manager/model_manager_default.py
2. Process control inputs — ControlNet processes control images to guidance tensors [(batch, height, width, channels) → (batch, height, width, control_channels)] invokeai/app/invocations/controlnet.py
3. Diffusion denoising — UNet performs iterative denoising guided by prompts and controls [(batch, channels, height//8, width//8) → (batch, channels, height//8, width//8)] invokeai/app/invocations/
4. VAE decode — VAE decoder converts latent tensors to RGB images [(batch, 4, height//8, width//8) → (batch, 3, height, width)] invokeai/app/invocations/

Assumptions & Constraints

• [warning] Assumes control images match expected dimensions for the loaded ControlNet model but no explicit validation (shape)
• [info] Hardcoded default resolutions (512x512 for SD1, 768x768 for SD2, 1024x1024 for SDXL) without runtime validation (value-range)
• [warning] Assumes IP-Adapter model and image encoder model are compatible but no explicit compatibility check (dependency)

Related Fullstack Repositories

Frequently Asked Questions

Analyzed on March 31, 2026 by CodeSea. Written by Karolina Sarna.