Fooocus

Implements an AsyncTask worker system that decouples image generation from the web UI thread, allowing users to interact with the interface while generation proceeds in background. The AsyncTask class holds generation parameters and tracking data, while a dedicated worker function processes tasks from a queue and provides real-time progress updates to the Gradio UI without blocking user interactions. This architecture enables responsive UI feedback during computationally expensive diffusion sampling.

Implements intelligent prompt expansion that automatically enriches user input prompts with contextually relevant descriptors before feeding them to the diffusion model. The system uses CLIP embeddings and a curated vocabulary (stored in extras/expansion.py) to suggest and inject quality-enhancing terms like lighting conditions, artistic styles, and composition details. This reduces the cognitive load on users to write detailed prompts while improving output quality through consistent enhancement patterns.

Implements a web-based user interface using Gradio (webui.py) that provides interactive controls for all generation parameters, style selection, image modification options, and real-time progress feedback. The UI is organized into logical sections (Image Generation Panel, Image Modification Features, Styles and Presets) with dropdown selectors, sliders, text inputs, and image preview areas. The interface updates asynchronously as generation progresses, providing live feedback without blocking user interactions.

Provides a configurable sampling system that supports multiple diffusion sampling algorithms (Euler, DPM++, LCM, etc.) with algorithm-specific parameters (steps, CFG scale, noise schedule). The sampling process is abstracted into a pluggable architecture (ldm_patched/contrib/external.py) that allows users to select different samplers for different generation characteristics. Each sampler has different speed/quality tradeoffs, enabling optimization for specific use cases (fast iteration vs high-quality output).

Implements automatic model discovery, downloading, and caching that manages the lifecycle of large model files (SDXL, VAE, LoRAs, etc.). The system checks for required models on startup, downloads missing models from configured sources (Hugging Face, CivitAI, etc.), and caches them locally to avoid re-downloading. Model paths are configurable, enabling users to organize models across multiple storage locations (e.g., fast SSD for active models, slow HDD for archives).

Implements Perpendicular Negative Guidance (ldm_patched/contrib/external_perpneg.py), an advanced guidance technique that uses negative prompts more effectively by projecting negative guidance perpendicular to positive guidance in embedding space. This prevents negative prompts from conflicting with positive prompts and improves adherence to the primary prompt intent. PerpNeg is optional and can be toggled per generation, providing an alternative to standard negative prompt handling.

Implements Self-Attention Guidance (ldm_patched/contrib/external_sag.py), a technique that enhances semantic coherence by modifying self-attention maps during diffusion sampling. SAG amplifies attention to semantically important regions, improving object definition and reducing artifacts. This is particularly effective for complex scenes with multiple objects or fine details. SAG is optional and can be toggled per generation.

Provides a preset system (stored in presets/*.json and sdxl_styles/sdxl_styles_fooocus.json) that applies curated style templates to user prompts, automatically injecting style-specific descriptors and parameter configurations. Each style (anime, realistic, semi-realistic, etc.) contains both prompt modifiers and recommended sampling parameters (steps, CFG scale, sampler type). The system composes user prompts with style templates at generation time, enabling one-click style application without manual parameter tuning.

Implements mask-based inpainting that allows users to selectively regenerate regions of an image by providing a mask and modified prompt. The system works in latent space (using VAE encoding/decoding) rather than pixel space, enabling efficient editing with reduced memory overhead. The inpainting pipeline preserves masked regions while diffusing unmasked areas according to the new prompt, supporting use cases like object replacement, style transfer on regions, and iterative refinement.

Provides image upscaling that operates in two stages: latent-space enhancement using the diffusion model with a higher resolution target, followed by optional post-processing refinement. The system encodes the input image to latent space, runs a modified diffusion sampling at the target resolution with the original prompt, then decodes back to pixel space. This approach preserves semantic content while adding detail, avoiding the artifacts of naive pixel-space upscaling.

Implements LoRA integration that allows users to load and blend multiple LoRA adapters into the base Stable Diffusion XL model at inference time. LoRAs are small, specialized model weights that modify the base model's behavior for specific styles, subjects, or concepts. The system uses a model patcher architecture (ldm_patched/modules/model_patcher.py) that composes LoRA weights with the base model using low-rank matrix operations, enabling efficient multi-LoRA blending with weighted contributions.

Integrates IP-Adapter (Image Prompt Adapter) and BLIP (Bootstrapping Language-Image Pre-training) to enable image-to-image generation where an input image provides visual conditioning without requiring text description. IP-Adapter uses CLIP vision embeddings from the input image to guide the diffusion model, while BLIP automatically generates descriptive captions for the image. This enables users to generate variations of an image or transfer its style to a new prompt without manually describing the visual content.

Integrates face restoration models (such as GFPGAN or similar) that detect and enhance faces in generated images, improving facial detail, clarity, and aesthetic quality. The system runs face detection on the generated image, extracts face regions, applies restoration models to enhance them, and blends the restored faces back into the original image. This post-processing step is optional and can be toggled per generation, improving quality for portrait and character generation workflows.

Implements CLIP patching (ldm_patched/ldm/modules/attention.py, ldm_patched/modules/clip_vision.py) that modifies the CLIP text encoder and vision encoder to improve semantic understanding and prompt-to-image alignment. The patching system allows injection of custom attention mechanisms, embedding transformations, and guidance strategies that enhance how the diffusion model interprets prompts. This enables more nuanced control over semantic guidance without modifying the base diffusion model.

Implements a flexible configuration system (args_manager.py) that merges settings from multiple sources with defined precedence: built-in defaults → config.txt file → preset files → command-line arguments. Each configuration source can override previous values, enabling users to customize behavior at multiple levels without modifying core code. Presets (stored in presets/*.json) provide pre-configured bundles for different use cases (anime, realistic, LCM, etc.), reducing the need for manual parameter tuning.