---
title: LynnReal-Onmi-light-vae
canonical_url: "https://www.modelscope.cn/models/AI-ModelScope/LynnReal-Onmi-light-vae"
md_url: "https://www.modelscope.cn/models/AI-ModelScope/LynnReal-Onmi-light-vae.md"
repository: AI-ModelScope/LynnReal-Onmi-light-vae
last_updated: 2026-09-22
license: other
parameters: 1.9B
tensor_type:
  - F32
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 88
stars: 0
tags:
  - "arxiv:2609.15863"
---

# LynnReal-Onmi-light-vae

> LynnReal-Onmi-light-vae - AI-ModelScope 在 ModelScope 开源的模型。LynnReal Lightweight Video VAE

AI-ModelScope/LynnReal-Onmi-light-vae 是 ModelScope 魔搭社区上的 1.9B 参数机器学习模型，采用 other 许可。

- **Repository**: AI-ModelScope/LynnReal-Onmi-light-vae
- **License**: other
- **Parameters**: 1.9B
- **Tags**: arxiv:2609.15863
- **Downloads**: 88
- **Stars**: 0
- **Last updated**: 2026-09-22

Source: https://www.modelscope.cn/models/AI-ModelScope/LynnReal-Onmi-light-vae

---

# LynnReal Lightweight Video VAE

Code: [LynnReal-AI/LynnReal-Omni](https://github.com/LynnReal-AI/LynnReal-Omni) · [Paper](https://arxiv.org/abs/2609.15863) · [HF Paper](https://huggingface.co/papers/2609.15863)

## 🆕 Latest updates

- **2026-09-18 · ComfyUI node pack fix** — the Flash workflows no longer stop with
  `RuntimeError: aimdo memory compile error` on machines where comfy-aimdo's DynamicVRAM is active
  (a cu13x torch, e.g. an RTX 5090). Update
  [`comfyui/custom_nodes/ComfyUI-LynnReal`](https://huggingface.co/stdstu123/LynnReal-Onmi-beta-0.1/tree/main/comfyui/custom_nodes/ComfyUI-LynnReal)
  and re-run.
- **2026-09-17 · Light VAE in ComfyUI** — the distilled 26-block decoder now loads through the
  `LynnRealH3VAELoader` node in the
  [main bundle](https://huggingface.co/stdstu123/LynnReal-Onmi-beta-0.1/tree/main/comfyui)
  (272/16 tiles), and is the decoder the Flash three-step workflows use.

Full history: [CHANGELOG.md](https://huggingface.co/stdstu123/LynnReal-Onmi-beta-0.1/blob/main/CHANGELOG.md)
in the main repository.

A complete codec with the unchanged MiniMax H3 encoder and a distilled 26-block
video decoder. Latent channels, normalization, spatial scale and temporal compression
match the official VAE. Deployment uses the decoder EMA parameters.

`config.json` and the indexed Safetensors shards follow the Diffusers component
layout. `decode_config.json` records the release decoder settings. Load with
`model.light_vae.LightVAE.from_pretrained("weight/light-vae")`, or pass
`--light-vae weight/light-vae` to `script/sample.py`. Native decoding tiles are
the default; precision and tile geometry must accompany speed/quality comparisons.

The codec changes reconstruction, not the denoiser latent interface. The same
latent need not decode to identical RGB under the official and lightweight VAE.
See `LICENSE` and the release `NOTICE` for MiniMax H3 terms and attribution.