Official announcements or software releases from relevant AI platforms

BitsFusion Achieves 7.9X Compression of Stable Diffusion v1.5 to 1.99 Bits with Improved Performance

Jun 7, 2024, 07:20 AM

BitsFusion is a new weight quantization method that compresses the UNet of Stable Diffusion v1.5 from 1.72 GB (FP16) to 219 MB (1.99 bits), achieving a 7.9X compression ratio. This method not only reduces the model size significantly but also improves its performance. The paper introduces a mixed-precision strategy to achieve this compression, transitioning from FP32, making diffusion-based image generation models more efficient and capable of synthesizing high-quality content.