Skip to content

[Feature] Qwen2D VAE Support #1610

@Hypi122

Description

@Hypi122

Feature Summary

Qwen2D VAE Support for image models. Uses 3x less vram and is 2.5x faster.

Detailed Description

Model: https://huggingface.co/Anzhc/Qwen2D-VAE

Author claims it's "Fully compatible with image models using Qwen/Wan VAEs.", but personally I've confirmed that it at least works for Anima.
On my particular hardware (GTX 960 4GB) vae decode time went from ~10s -> 2.99s on 512x512 image. Although I presume that is mostly (or at least partially) because I didnt need to do any offloading with this 2d vae. And memory for vae went from 139.84 MB to 48.28 MB
Images are pixel perfect between those 2 vae.
I've obtained those numbers from using my vibecoded patches, as I do not understand this code in the slightest I will not be making PR for it.

Alternatives you considered

No response

Additional context

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions