This is a weekly bulletin from Replicate covering open-source AI models, tools, and research. It highlights the release of Stable Diffusion 3 Medium, a new image generation model with a 2B parameter…
Replicate published a guide showing how to push a custom version of Stable Diffusion 3 to their platform using Cog, an open-source tool for packaging ML models in Docker containers.
Hugging Face Accelerate introduces a precision-related change and a concept guide to harmonize FSDP and DeepSpeed backends, enabling users to seamlessly switch between them.
Hugging Face introduces the RLOO (REINFORCE Leave One-Out) Trainer in TRL, a new online RLHF training algorithm that is more memory-efficient and faster than PPO.
Stable Diffusion 3 Medium (2B parameters) is now available on the Hugging Face Hub and integrated with the Diffusers library, including memory optimizations and training scripts.
OpenAI appoints retired U.S. Army General Paul M. Nakasone to its Board of Directors, bringing cybersecurity experience and joining the Board's Safety and Security Committee.
The blog post from LG AI Research presents two research papers on neural network compression presented at ICLR 2024: LESS, which combines filter pruning and low-rank decomposition for vision tasks,…
Sakana AI releases a research report on using LLMs to discover preference optimization algorithms, including the state-of-the-art DiscoPOP algorithm, and open-sources the codebase and model…