VibeusOstris has released Primal VAE in two variants: 2x1 and 2x2, named... - Vibeus
Ostris has released Primal VAE in two variants: 2x1 and 2x2, named for spatial compression and channel count.
The autoencoder reduces spatial resolution only 2× and produces one or two latent channels. Applying 4×4 patching then gives the familiar layouts without making the VAE perform deep spatial compression:
• 2x1 → 16 channels at 8× compression, matching the Flux.1 latent shape.
• 2x2 → 32 channels at 8× compression, matching the Flux.2 latent shape.
The diffusion transformer handles patching while the VAE stays small. Its latent is trained to have statistics aligned with Gaussian noise, avoiding a separate scaling factor, shift factor or latent-normalization step. The workflow is encode, patch, then diffuse.
The attached reconstruction comparison favors the 2x2 version over the Flux.2 VAE on three 512-pixel examples. The model card reports PSNR of 37.5 vs 37.2 dB for the studio photo, 27.1 vs 26.5 for the picnic scene, and 27.4 vs 25.9 for the interface screenshot. Both use the same information budget: 32 values for each 8×8 image block.
The developer explicitly notes that three images and PSNR alone do not constitute a benchmark. A KL VAE also pays a reconstruction cost for posterior noise that this deterministic encoder does not. The 2x1 variant trails the Flux.1 VAE on textured content and has received less training: 62k steps versus 147k for 2x2.
Both variants have roughly 80M parameters. Image dimensions must be multiples of 16. The weights and Diffusers model implementation are available under Apache 2.0; the original post links to the model card and usage details.
