Back to Model
P

Primal VAE

Model Nov 4, 2026

Ostris's deterministic image autoencoder family uses shallow spatial compression and patching to produce latent layouts for diffusion research.

Visit primary link 3 sources attached
Author comparison shows Primal VAE 4x1, SD 1.5 VAE and SDXL VAE reconstructions, latents and amplified error for three images.
Ostris’s author-supplied comparison of three 512-pixel inputs. Reconstruction PSNR is an illustrative sample, not an independent benchmark or proof of generation quality. Source: Ostris, LLC / Primal VAE
multimodalself-hosted

What happened lately

Primal VAE timeline

Primal VAE provides deterministic image encoders and decoders with shallow spatial compression. The original 2x1 and 2x2 checkpoints were added October 30; a 4x1 checkpoint followed November 4. The English card documents Diffusers remote-code loading, patching helpers and author-produced reconstruction comparisons. Tensor layout matching is distinct from demonstrated compatibility with an unchanged diffusion model. The project is licensed Apache 2.0; reported reconstruction quality and latent statistics are the author’s observations, not AITrending benchmark reproductions.