Sd Vae Ft Mse

Sd Vae Ft Mse is a fine-tuned variational autoencoder model published by Stability AI on Hugging Face. It serves as an improved decoder for latent representations in Stable Diffusion image generation pipelines.

The model replaces the original VAE decoder in compatible Stable Diffusion setups to produce higher fidelity reconstructions from latent space to pixel space. It was evaluated on the COCO 2017 validation set using 5000 images at 256 by 256 resolution and on a 10000-image subset of LAION-Aesthetics 5+ at the same resolution. These weights are distributed specifically for use with the Diffusers library.

Integration occurs through a standard installation sequence. Users run pip install -U diffusers transformers accelerate, then load the model with a few lines of Python code that instantiate a DiffusionPipeline from the repository identifier stabilityai/sd-vae-ft-mse. The provided example code demonstrates generation of an image from a text prompt using bfloat16 precision and CUDA device mapping. Notebooks for Google Colab and Kaggle are referenced as additional starting points.

The model card lists safetensors and Diffusers support along with an MIT license. It is positioned within the class of Stable Diffusion components that enhance autoencoder performance through decoder fine-tuning.