Stability-AI/generative-models

Features

• Diffusion Models - Provides a comprehensive framework for training and sampling diffusion models to generate high-fidelity images, video, and 4D assets.
• Latent Diffusion Models - Implements high-fidelity content generation by iteratively removing noise within compressed latent spaces.
• Input Standardization Conditioners - Standardizes input types like vectors and sequences through a single conditioner to guide model generation.
• Text-to-Image Generators - Produces synthetic high-resolution images based on natural language text prompts or existing image inputs.
• Media Synthesis from Text - Generates high-fidelity images and video from textual or latent descriptions using diffusion architectures.
• Generative Model Training Tools - Provides a modular environment for configuring networks and pipelines to train diffusion models.
• External Embedder Conditioning - Steers the generative process using external embedders that process text and class labels.
• Training Datasets - Provides systems for processing large-scale image and label datasets for generative model training.
• Latent Conditioning Mechanisms - Implements mechanisms for injecting spatial or semantic guidance into the latent space of diffusion models.
• Model Training Pipelines - Ships a comprehensive workflow for sourcing datasets, configuring networks, and training diffusion models.
• Modular Agent Assembly - Uses a modular system to assemble network submodules and conditioners via YAML definitions.
• Diffusion Process Conditioners - Implements a standardized conditioner to process diverse input types into vectors for steering the generative process.
• Noise Level Sampling Strategies - Provides a system for managing noise level sampling and loss weighting to stabilize generative training.
• Noise-Aware Loss Weighting - Provides modular loss weighting that decouples noise level sampling from loss calculation to stabilize training.
• Training Pipelines - Offers a modular environment for managing datasets, noise sampling, and loss weighting in training pipelines.
• Generative Training Data Pipelines - Ships a map-style data pipeline to efficiently feed large-scale image and label pairs into the training loop.
• Numerical Diffusion Solvers - Provides customizable mathematical solvers to transform noise into images via discretized time-steps.
• Model Architecture Configurations - Implements a YAML-based assembly process to define network submodules and conditioners without manual boilerplate code.
• Diffusion Sampling Methods - Provides customizable numerical solvers and discretization methods to generate final outputs from the diffusion model.
• Novel View Synthesis Engines - Transforms single input videos into consistent 4D assets by generating novel camera views and frames.
• Model Component Assembly - Uses YAML-based assembly to define and instantiate model submodules without modifying the core source code.
• AI Content Watermarking - Implements a dedicated system for embedding and detecting invisible markers in AI-generated imagery.
• AI Watermark Detection Tools - Includes a dedicated utility for detecting invisible markers embedded in images to verify model origin.
• Invisible Watermarks - Integrates subtle, invisible markers into generated pixel data to verify the origin of synthetic content.
• Synthetic Media Provenance Verification - Includes a toolset to identify embedded markers in generated images to verify the model of origin.
• Diffusion Model Research - High-resolution image synthesis using latent diffusion.
• Foundation Models - Core generative model repository from Stability AI.
• Generative Media Tools - Generative models for image and video.
• Model Acceleration - Adversarial diffusion distillation for rapid image synthesis.
• Generation - Listed in the “Generation” section of the Awesome Diffusion Models awesome list.

Star history