wowsuch ai
Live

ostrisModel

Primal VAE 2x2

Primal VAE 2x2 is a downloadable image autoencoder checkpoint that encodes images into a two-channel latent at 2× spatial reduction. Its latent is trained to have unit-normal statistics and can be patched to match the Flux.2 VAE latent shape.

What it does

Developers can use it to encode and reconstruct images or provide Flux.2-shaped latents to compatible diffusion workflows.

Availability
Weight file: primal_vae_2x2.safetensors
Pricing model
Not established
Deployment
Load for inference with Diffusers; example uses CUDA
Privacy
Not established
Licence
Not established
Capability
image generation
Modality
image
Access
open weightsSDK
Deployment
on-premises
Pricing
unknown
Target user
developerresearcher

Evidence

5 claims
  • The 2x2 checkpoint is available as primal_vae_2x2.safetensors. [1]
    vae = AutoModel.from_pretrained("ostris/primal_vae", variant="2x2", trust_remote_code=True, torch_dtype=torch.bfloat16).to("cuda")
    Hugging Face. Quoted text is outside CC BY 4.0 and keeps its original rights.
  • The 2x2 variant encodes to 2 channels at 2× spatial reduction. [1]
  • With patch compression 8, its latent has 32 channels at 8×, matching the Flux.2 VAE latent shape. [1]
    patch 4 the latent, you end up with 8x at 16 channels, the same as most modern VAEs, without the need for deep
    Hugging Face. Quoted text is outside CC BY 4.0 and keeps its original rights.
  • The 2x2 model has approximately 80 million parameters. [1]
    is a 12:1 reduction with no slack, and it has well under half the training steps of the 2x2 (62k vs 147k).
    Hugging Face. Quoted text is outside CC BY 4.0 and keeps its original rights.
  • Image sides need to be multiples of 16. [1]
    Image sides need to be multiples of 16.
    Hugging Face. Quoted text is outside CC BY 4.0 and keeps its original rights.

Sources

1
1
Hugging FaceOwner-controlled
https://huggingface.co/ostris/primal_vae Retrieved 2026-10-31 00:47 UTC
Verified

Versions

1
VersionEffectiveChangeChanged fieldsState
12026-10-31 00:48 UTCpublicationNoneCurrent

Open in the catalog

Original catalog content is licensed under CC BY 4.0, attributed to wowsuchai. Source quotations and linked content keep their original rights.