Introduction

The LightX2V team at ModelTC just shipped LightVAE, a set of slimmed-down video VAE decoders for Wan2.1, Wan2.2 and MiniMax-H3, plus a lighter H3 encoder. The checkpoints are on Hugging Face (lightx2v/LightVAE) under Apache-2.0, and the code, a ComfyUI custom node and six ready-made workflows live on GitHub (ModelTC/LightVAE). The ComfyUI exports landed on October 12, so you can wire this into a local graph today.

Why care? After your sampler finishes, the VAE still has to turn latents back into frames. On long, high-res clips that last step is slow, and you sit there waiting on it. LightVAE goes after that wait without touching your diffusion model.

What LightVAE actually is

LightVAE compresses existing pretrained video VAE decoders but keeps their original latent interface. Your model's latents stay exactly the same; only the decoder doing the unpacking gets lighter.

Two flavors:

  • LightVAE-Pro: temporal low-rank (TLR) convolutions at the original decoder depth, for the best reconstruction fidelity. Wan2.1 and Wan2.2 only.
  • LightVAE-Lite: Pro plus extra block pruning for more speed. Available for Wan2.1, Wan2.2 and MiniMax-H3.
  • H3 bonus: a separate Lite encoder for H3, so reference-image and reference-video encoding can get faster too.

What stays the same:

  • Wan encoders stay original. LightVAE only swaps the decoder.
  • H3 audio VAE is untouched. LightVAE is a video codec, not an audio one.
  • Latents keep their format, so no manual rescaling and no retraining.

One detail that will save you a headache: Wan2.2 14B uses the Wan2.1 VAE (16 latent channels), so it pairs with the wan21 decoders. Only the Wan2.2 5B TI2V model uses the Wan2.2 VAE and the wan22 files.

How much faster?

These are the team's own ComfyUI numbers: NVIDIA H200, BF16, 768×1344, batch 1, 125 frames for Wan / 124 for H3, median of three runs. They time the VAE.decode step only, not the whole generation.

Wan2.1

  • Original VAE: 6.844 s (18.26 FPS)
  • LightVAE-Pro: 3.754 s (33.30 FPS), 1.82× faster
  • LightVAE-Lite: 2.330 s (53.64 FPS), 2.94× faster

Wan2.2

  • Original VAE: 6.886 s (18.15 FPS)
  • LightVAE-Pro: 4.420 s (28.28 FPS), 1.56× faster
  • LightVAE-Lite: 2.999 s (41.68 FPS), 2.30× faster

MiniMax-H3

  • Original VAE: 6.303 s (19.67 FPS)
  • LightVAE-Lite: 3.574 s (34.69 FPS), 1.76× faster

So on a 125-frame Wan2.1 clip, the decode drops from almost seven seconds to just over two. The repo is clear that these are decode speedups, not full-generation speedups. Your sampler steps take as long as they did before; you just get the frames back sooner. The README also shows side-by-side reconstruction videos (original vs Pro vs Lite) if you want to judge quality yourself.

Run it in ComfyUI

The -comfy.safetensors files need the team's Load LightVAE (TLR) node. Don't load them with ComfyUI's standard Load VAE node. Install it from your ComfyUI folder, using the same Python environment that runs ComfyUI (H3 Lite decoder shown here):

cd /path/to/ComfyUI

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

hf download lightx2v/LightVAE \
  lightvae-lite-h3-decoder-comfy.safetensors \
  --local-dir models/vae/LightVAE

To grab every ComfyUI export at once:

hf download lightx2v/LightVAE --include "*-comfy.safetensors" \
  --local-dir /path/to/ComfyUI/models/vae/LightVAE

Restart ComfyUI and search for Load LightVAE (TLR). Then:

  • Wan: keep your normal Load VAE and plug it into LightVAE's base_vae. Pick a Pro or Lite decoder, leave encoder = none, and send LightVAE's VAE output to your encode/decode nodes.
  • H3, decoder only: pick the Lite decoder, set encoder = none, and connect the VAE output to VAE Decode. Keep your original first/last-frame encoding wiring.
  • H3, full Lite: pick both Lite files and feed the VAE output to the reference-encoding node and VAE Decode. No original video VAE is needed.
  • Precision, backend, device: leave them on auto. Triton kernels kick in on compatible NVIDIA GPUs; otherwise it falls back to plain PyTorch.

The six bundled workflows

Drag any of these JSON graphs from the repo's workflows folder into ComfyUI. They need only ComfyUI core plus the LightVAE node:

  • Wan2.1 1.3B T2V: 832×480, 81 frames, 16 FPS (the team's suggested first run)
  • Wan2.2 5B T2V: 1280×704, 121 frames, 24 FPS
  • Wan2.2 5B I2V: 1280×704, 121 frames, 24 FPS
  • Wan2.2 14B I2V: 640×640, 81 frames, 16 FPS, two experts
  • H3 reference image: 1344×768, 124 frames, 24 FPS, video + audio
  • H3 reference video: 1344×768, 124 frames, 24 FPS, video + audio

Prefer Python?

Use the regular checkpoints, the ones without -comfy in the name. Here's a Wan2.1 Lite reconstruction with the original encoder:

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

You get an MP4 plus a JSON with PSNR and timings. There's also a benchmark.py for comparing Base, Pro and Lite on your own GPU. To plug LightVAE into a full generation pipeline outside ComfyUI, the team points to their LightX2V framework.

Good to know before you swap

  • Decode only: LightVAE doesn't shrink the DiT or text encoder. The H3 workflows still load a big BF16 model and a 32B text encoder, so you may still need CPU offload and plenty of system RAM.
  • Hardware: speed was measured on an H200. The Triton path was also compile-checked for Ampere/Ada GPUs, but other GPU families haven't been tested on real hardware.
  • Testing so far: all six workflows passed graph validation, but only Wan2.1 T2V and H3 image-reference were run end to end.
  • Custom sockets: nodes that expect a WANVAE input aren't drop-in compatible. Stick to the standard VAE nodes.
  • Paper: a LightVAE paper is referenced, but no public link was up when we checked. The GitHub repo also doesn't have a license file yet; the Apache-2.0 license comes from the Hugging Face model card.
  • Different from earlier H3 decoders: this is ModelTC's own release, not the community H3 light-decoder conversions that went around earlier this month.

Conclusion

LightVAE is a small fix for a real annoyance: waiting on the final decode after your sampler is already done. Up to 2.94× faster on Wan2.1, 2.30× on Wan2.2 and 1.76× on H3, all under Apache-2.0, with a ComfyUI node and six workflows ready to go. If you render long Wan or H3 clips locally, start with the Wan2.1 1.3B workflow, decode once with the original VAE and once with Lite on the same seed, and compare them yourself.

—Aurelia ♡