Browse Models
Generate images and videos using a powerful low-level workflow graph builder - the fastest, most flexible, and most advanced visual generation UI.
Forge is a platform built on top of Stable Diffusion WebUI to make development easier, optimize resource management, speed up inference, and study experimental features.
The Stable Diffusion 3.5 model family, released by Stability AI on October 22nd, 2024, represents a significant advancement in text-to-image generation technology. This family consists of three distinct models that share a common architectural foundation while serving different use cases and hardware requirements. The family includes Stable Diffusion 3.5 Large, Stable Diffusion 3.5 Turbo, and Stable Diffusion 3.5 Medium, each optimized for specific scenarios ranging from professional use to consumer applications.
The Stable Diffusion 3.5 family introduces a sophisticated Multimodal Diffusion Transformer (MMDiT) architecture that serves as the foundation for all models in the lineup. This architecture incorporates three pretrained text encoders: OpenCLIP-ViT/G, CLIP-ViT/L, and T5-xxl, as detailed in the MMDiT research paper. A notable technical innovation across the family is the implementation of Query-Key (QK) normalization in transformer blocks, which enhances training stability and simplifies fine-tuning processes.
The training dataset for the model family encompasses both synthetic and filtered publicly available data, contributing to the models' diverse capabilities and robust performance. This comprehensive training approach enables the models to generate high-quality images across various styles and contexts, from photorealistic renderings to artistic interpretations.
The flagship model of the family, Stable Diffusion 3.5 Large, features 8.1 billion parameters and is optimized for professional use at 1-megapixel resolution. It demonstrates superior performance in image quality, typography rendering, and prompt adherence, setting new benchmarks in the field of text-to-image generation. The model's architecture allows for maximum customization potential, making it particularly suitable for professional and enterprise applications.
Stable Diffusion 3.5 Turbo represents an innovative approach to model optimization through Adversarial Diffusion Distillation (ADD) technology. This variant maintains competitive quality while dramatically reducing inference time to just 4 steps, making it ideal for applications where speed is crucial. The Turbo variant demonstrates how advanced distillation techniques can preserve core capabilities while significantly improving computational efficiency.
Stable Diffusion 3.5 Medium, with 2.5 billion parameters, bridges the gap between professional and consumer applications. It features an improved MMDiT-X architecture and supports multi-resolution generation from 0.25 to 2 megapixels, while maintaining a modest 9.9GB VRAM requirement. This makes it particularly suitable for individual creators and small businesses working with consumer-grade hardware.
The entire model family demonstrates remarkable improvements in several key areas compared to previous generations. These improvements include enhanced image quality, superior typography rendering, and better understanding of complex prompts. The models show particular strength in generating diverse representations of people and features without requiring extensive prompt engineering, as evidenced in the official release announcement.
For optimal results, the models generally perform best with specific parameter settings, such as 28 inference steps with a guidance scale of 3.5 for the Large variant. The Turbo variant's ability to generate high-quality images in just 4 steps represents a significant advancement in inference efficiency, while the Medium variant offers flexibility in resolution scaling to accommodate various use cases.
The entire Stable Diffusion 3.5 family is released under the Stability Community License, which provides a balanced approach to accessibility and commercial use. The license permits free use for research purposes, non-commercial applications, and commercial use by entities with annual revenue below $1 million. Organizations exceeding this threshold require an Enterprise License, ensuring sustainable development while maintaining broad accessibility.
Stability AI has implemented comprehensive safety measures across the model family, incorporating data filtering and safeguards to mitigate potential harms. The models feature safety-by-design principles, as detailed in the company's safety guidelines. While these measures provide a foundation for responsible use, developers are encouraged to implement additional safeguards based on their specific use cases and requirements.
The Stable Diffusion 3.5 family represents a significant step forward in making advanced text-to-image generation accessible to a broader range of users while maintaining high standards of quality and performance. The family's tiered approach, offering variants optimized for different use cases and hardware capabilities, demonstrates a mature understanding of market needs and technical constraints. As detailed in the research paper, the architectural innovations and performance improvements established by this model family are likely to influence future developments in the field of AI-powered image generation.
The success of the Stable Diffusion 3.5 family in balancing performance, accessibility, and safety considerations sets a new standard for AI model development and deployment. The family's comprehensive approach to addressing various user needs while maintaining consistent quality across variants suggests a promising direction for future model development in the field of generative AI.