Draw Things

Generates images from natural language prompts by executing Stable Diffusion and FLUX models directly on Apple Silicon devices using Metal GPU acceleration, eliminating cloud dependency and network latency. Models are downloaded once and cached locally, enabling offline generation after initial setup. The Metal acceleration framework optimizes tensor operations and memory bandwidth for M-series chips, delivering generation times measured in minutes per image on consumer hardware.

Enables users to train custom Low-Rank Adaptation (LoRA) modules locally on Apple Silicon devices by fine-tuning base models (Stable Diffusion, FLUX) on user-provided image datasets. Trained LoRAs are stored locally and can be applied during inference to customize model outputs without retraining the full base model. The training process uses gradient descent optimization on-device, with inference applying LoRA weights as low-rank matrix multiplications during the diffusion process.

Supports multiple image generation models (Stable Diffusion, FLUX, and others) with UI-based model selection, enabling users to switch between models for different generation tasks without restarting the app. Each model is downloaded and cached separately, and the app manages model loading and memory allocation. Implementation uses abstraction layer for model inference to support multiple architectures.

Provides native UI implementations across iOS, iPadOS, and macOS using platform-specific frameworks (SwiftUI, UIKit) rather than cross-platform abstractions, enabling optimized UX for each platform. The unified codebase shares inference logic while maintaining platform-specific UI patterns and capabilities. iOS/iPadOS versions leverage touch input and mobile-optimized layouts; macOS version uses keyboard shortcuts and desktop-optimized workflows.

Offers free local image generation on Apple Silicon devices with limited cloud compute hours (Lab Hours), with optional paid tier (Draw Things+) providing higher cloud compute quotas and custom LoRA cloud inference. Free tier enables full local inference without payment; cloud features are optional and quota-based. Pricing model uses monthly Lab Hours allocation rather than per-request billing.

Distributes the application through Apple App Store for iOS/iPadOS/macOS with direct download option as fallback when App Store is unavailable or inaccessible. App Store distribution enables automatic updates and seamless installation; direct download provides alternative installation path for users in regions with App Store restrictions or experiencing connectivity issues.

Applies ControlNet conditioning to text-to-image generation, allowing users to guide model outputs using structural constraints (edge maps, pose skeletons, depth maps, etc.) provided as input images. ControlNet modules are loaded alongside base models and inject spatial conditioning into the diffusion process, enabling precise control over composition, pose, or layout without full inpainting. Implementation uses cross-attention mechanisms to blend ControlNet embeddings with text prompt embeddings during denoising steps.

Enables users to edit specific regions of images by masking areas and regenerating only masked regions using the diffusion model, preserving unmasked content. The infinite canvas feature allows expanding the image boundaries and filling new regions with model-generated content. Inpainting uses masked diffusion, where the model only denoises masked pixels while keeping unmasked pixels fixed, enabling seamless blending of edited and original content.

Transforms images into different artistic styles (animation, painting, sketch, etc.) by using the source image as a conditioning input to the diffusion model. The process encodes the source image into latent space and applies style guidance through text prompts or style presets, regenerating the image in the target style while preserving composition and content. Implementation uses image encoding followed by conditional diffusion with style-specific prompts or LoRA weights.

Converts static images into short video clips by generating frame sequences that extend or animate the source image content. The process likely uses frame interpolation or latent space animation to create smooth transitions between generated frames. Implementation details are undocumented, but the feature appears to create basic animations rather than full video generation from scratch.

Manages the download, storage, and lifecycle of image generation models (Stable Diffusion, FLUX, ControlNets, LoRAs) from Draw Things servers to local device storage. Models are downloaded once and cached locally, enabling offline inference after initial setup. The system tracks model versions, manages storage quotas, and provides UI for model selection and deletion. Implementation uses HTTP downloads with resume capability and local filesystem storage with version tracking.

Provides optional cloud inference capability for users who want faster generation or higher quality outputs than local device allows, with quota-based billing using 'Lab Hours' currency. Free tier includes limited monthly Lab Hours; paid tiers (Draw Things+) increase quota. Cloud inference uses Draw Things' managed servers and supports custom LoRA inference, which is not available in free local tier. Implementation uses account-based authentication and cloud API for offloading generation requests.

Enables users to visualize apparel or character designs on virtual models or in different contexts by applying generated or uploaded designs to reference images. The mechanism is undocumented but likely uses image-to-image transformation or ControlNet-guided generation to adapt designs to different poses, body types, or character models. Allows rapid prototyping of apparel concepts without physical samples.

Provides UI controls for fine-tuning text-to-image generation through prompt input and numerical parameters (guidance scale, sampling steps, seed, etc.). Guidance scale controls how strongly the model follows the text prompt; sampling steps control diffusion quality and speed; seed enables reproducible generation. Implementation uses standard diffusion parameter exposure through UI sliders and text input fields.