AINovice2005/pruna-vaed-modular-diffusers

---

Pruna VAE for Modular Diffusers

This repository provides a Modular Diffusers implementation of the PrunaVAED. It packages the pruned decoder implementation together with the custom loading logic required to use the model as a reusable Modular Diffusers component.

The repository exposes a custom LoadPrunaVAE block that downloads and instantiates PrunaAutoencoderKLLTX2Video, making it available as components.vae within a Modular Diffusers pipeline.

Features

• Custom PrunaAutoencoderKLLTX2Video implementation
• Compatible with trust_remote_code=True
• Self-contained repository containing both the implementation and model weights
• Intended for composition inside Modular Diffusers pipelines

Intended Use

This repository is designed to be consumed by Modular Diffusers pipelines and reusable pipeline blocks. It is particularly useful for:

• Video inference
• Modular Diffusers experimentation
• Custom video generation pipelines
• Research on pruned video VAEs

Usage

from diffusers.modular_pipelines import ModularPipelineBlocks

blocks = ModularPipelineBlocks.from_pretrained(
    "AINovice2005/pruna-vaed-modular-diffusers",
    trust_remote_code=True,
)

pipeline = blocks.init_pipeline()
pipeline.load_components()

vae = pipeline.vae

The loaded component is an instance of:

PrunaAutoencoderKLLTX2Video

and can be used anywhere a compatible LTX-2 Video VAE is expected.

The VAE decoder reconstructs frames from compressed latent representations. Depending on the downstream workflow, applying a dedicated image or video upscaler after decoding can noticeably improve perceived sharpness and fine detail, particularly for outputs intended for display at higher resolutions.

Acknowledgements

This work builds upon the Pruna VAE implementation released by PrunaAI and adapts it for use as a reusable Modular Diffusers component.