PrunaVAED: Drop-In Faster VAE Decoder for LTX-2.3 by Pruna AI
PrunaVAED is a drop-in replacement VAE decoder for LTX-2.3, achieving 1.7x faster decoding and ~50% lower peak VRAM with near-original visual quality.
The PrunaVAED model directly replaces the video VAE decoder in diffusers/LTX-2.3-Diffusers. The encoder and latent format remain unchanged, making it a drop-in upgrade — no workflow changes required.
Benchmark
Metrics compare videos decoded from the same latents by the LTX-2.3 VAED and PrunaVAED on an NVIDIA H100 80GB (bfloat16, batch size 1, full decode without tiling or torch.compile):
| Clip length | LTX-2.3 VAED | PrunaVAED | Speedup |
|---|---|---|---|
| 5 s @ 720p | 841.3 ms | 500.2 ms | 1.68× |
| 10 s @ 720p | 1670.7 ms | 998.0 ms | 1.67× |
Side-by-Side Comparison
The following videos compare LTX-2.3's original decoder (left) with PrunaVAED (right) on the same latents:
| LTX-2.3 Decoder | PrunaVAED |
|---|---|
ComfyUI Support
PrunaVAED is supported in ComfyUI as of commit c38171d (Support Pruna LTX VAE). The ComfyUI-LTXVideo wrapper nodes from Lightricks automatically detect and load the PrunaVAED decoder when the model file is placed in the correct directory.
Usage
Download the PrunaVAED safetensors from HuggingFace and place it in your ComfyUI models/vae/ directory. The ComfyUI LTXVideo nodes will automatically use it as the decoder — no additional configuration required.
Links
- HuggingFace Model: suryatmodulus/PrunaVAED (also at PrunaAI/PrunaVAED)
- Pruna AI GitHub: github.com/PrunaAI/pruna
- Base Model: diffusers/LTX-2.3-Diffusers
- License: LTX-2 Community License
Comments
Sign in with GitHub to join the discussion.