V1-5-pruned-emaonly

Based on the Latent Diffusion model structure. Setting it Up

: Refers to Stable Diffusion version 1.5, which was fine-tuned for 595k steps at a native resolution of 512x512 on the

: This is the Stable Diffusion version 1.5, a highly popular and versatile base model for AI generation. v1-5-pruned-emaonly

This is the most crucial differentiator. EMA stands for .

By choosing the emaonly variant, the user is opting for the most stable, consistent, and predictable version of Stable Diffusion 1.5. It removes the option to swap to non-EMA weights, streamlining the inference process. Based on the Latent Diffusion model structure

Stable Diffusion 1.5 often produces grey/washed-out images if used without a proper VAE (Variational Autoencoder). You need to download a matching VAE (e.g., vae-ft-mse-840000-ema-pruned ) and place it either as:

If you are running on a GPU with 4GB to 6GB of VRAM (e.g., Nvidia GTX 1050 Ti, 1660, or RTX 3050), the 2GB pruned-emaonly model fits perfectly, leaving room for the VAE and latent space processing. If you tried to load a 7GB full model, you would crash with out-of-memory (OOM) errors. EMA stands for

When AI models are trained, researchers often keep two sets of data: the raw, active weights and a smoothed-out version called the EMA weights. The EMA version is generally better for generating high-quality images but isn't useful if you want to continue training the model yourself. By providing only the EMA weights and "pruning" out the rest, developers created a file that is significantly smaller—around 4GB compared to the original 7GB+—without sacrificing the quality of the final image. Why Use v1-5-pruned-emaonly?