AI BriefingSegmind unveils lightweight SD models
Key point
Segmind has released SD-Small and SD-Tiny, lightweight Stable Diffusion models with reduced parameters achieved through knowledge distillation.
Details
Segmind has open-sourced the weights and training code for its compressed SD-Small and SD-Tiny models to improve the efficiency of generative AI models.
These models apply Knowledge Distillation techniques to reduce the number of parameters by 35% and 55% respectively compared to the existing model, while being designed to preserve image quality as much as possible.
Key technology and training information:
- Block-removal KD: This approach trains the student model after removing some of the UNet layers.
- Multiple loss functions: In addition to Latent loss, Feature-level loss, which matches the block outputs of the teacher model and the student model, was used to minimize quality degradation.
- Training environment: Training was conducted using Realistic-Vision 4.0 as the teacher model and the LAION Art Aesthetic dataset.
The training code is available on GitHub, and the pretrained checkpoints can be found on Hugging Face.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.