⚡ LightVAE

Towards Compact and Efficient Video Autoencoders

📄 Paper · 💻 Code · 🧩 ComfyUI · ⚡ LightX2V

LightVAE accelerates video autoencoders while preserving their original latent interfaces. We provide compressed components for Wan2.1, Wan2.2 and MiniMax-H3, with native ComfyUI integration for H3.

📦 Checkpoints

Backbone	Component	Download
Wan2.1	Pro decoder	safetensors
Wan2.1	Lite decoder	safetensors
Wan2.2	Pro decoder	safetensors
Wan2.2	Lite decoder	safetensors
MiniMax-H3	Lite decoder	safetensors
MiniMax-H3	Lite encoder	safetensors

Pro uses temporal low-rank compression while retaining the original decoder depth. Lite additionally prunes blocks for Wan; H3 Lite uses block pruning.

Wan models reuse their original encoders. H3 supports independently replacing the encoder and decoder.

Choose the correct Wan VAE: Wan2.2 14B uses the Wan2.1 VAE; Wan2.2 5B TI2V uses the Wan2.2 VAE.

🧩 ComfyUI · MiniMax-H3

The Load H3 LightVAE node reuses ComfyUI’s native H3 VAE. No LightVAE package installation, TLR or custom Triton kernels are needed. The current node is H3-only.

Install

Update ComfyUI to a version with native MiniMax-H3 support. Run with the Python environment used by ComfyUI:

cd /path/to/ComfyUI

git clone https://github.com/ModelTC/LightVAE.git LightVAE
python LightVAE/comfyui/install.py --comfy-root .

python -m pip install huggingface_hub
hf download lightx2v/LightVAE \
  lightvae-lite-h3-decoder.safetensors \
  lightvae-lite-h3-encoder.safetensors \
  --local-dir models/vae/LightVAE

Already cloned? Run git -C LightVAE pull, then rerun the installer.

The installer creates custom_nodes/ComfyUI-LightVAE/ automatically. Do not clone the entire repository directly into custom_nodes.

The original H3 video VAE is required. If missing:

hf download Comfy-Org/MiniMax-H3 \
  vae/minimax_h3_video_vae_fp16.safetensors \
  --local-dir models

Connect

Restart ComfyUI and search for Load H3 LightVAE.

Load the original H3 video VAE with the standard Load VAE node, then connect it to LightVAE’s base_vae input.

• FL2V: select the Lite decoder and leave encoder = none. Keep original first/last-frame encoding unchanged; connect LightVAE to VAE Decode.
• REF2V: connect LightVAE to both reference encoding and VAE Decode. Select either or both Lite components.
• none uses the corresponding original component. Encoder-only replacement is supported with decoder = none.

Keep the audio VAE unchanged. Load Lite components through Load H3 LightVAE, not the standard Load VAE node.

The standard H3 files above and previous *-h3-*-comfy.safetensors exports are both accepted.

Workflows

Download a JSON and drag it into ComfyUI:

FL2V · Image references · Video reference

See the workflow guide for required base models and input preparation.

🐍 Python · Video Reconstruction

Example: original Wan2.1 encoder + LightVAE-Lite decoder.

git clone https://github.com/ModelTC/LightVAE.git
cd LightVAE
python -m pip install -r requirements.txt

hf download lightx2v/LightVAE \
  lightvae-lite-wan21-decoder.safetensors \
  --local-dir weights/LightVAE

hf download Wan-AI/Wan2.1-T2V-1.3B \
  Wan2.1_VAE.pth --local-dir weights/Wan2.1

python infer.py \
  --model wan21 --variant lite \
  --encoder weights/Wan2.1/Wan2.1_VAE.pth \
  --decoder weights/LightVAE/lightvae-lite-wan21-decoder.safetensors \
  --input demo.mp4 --frames 81 --height 480 --width 832 \
  --output outputs/wan21_lite_recon.mp4

More examples, benchmarks and visual comparisons are available in LightVAE. For video generation with LightVAE, see LightX2V.

⭐ Support

If you find this useful, please star LightVAE and LightX2V!