pruna-vaed-mlx / README.md
duckdbot's picture
Add PrunaVAED MLX decoder (mlx-forge pruna-vaed)
a4fc87f verified
|
Raw
History Blame Contribute Delete
1.48 kB
---
license: apache-2.0
library_name: mlx
tags:
- mlx
- ltx-2.3
- vae
- video
- apple-silicon
base_model:
- PrunaAI/PrunaVAED
- Lightricks/LTX-2.3
---
# pruna-vaed-mlx
MLX conversion of [PrunaAI/PrunaVAED](https://huggingface.co/PrunaAI/PrunaVAED) — a pruned LTX-2.3 **video VAE decoder** for Apple Silicon.
Converted with [mlx-forge](https://github.com/dgrauet/mlx-forge) (`pruna-vaed` recipe): Conv3d weights transposed to MLX channels-last; `per_channel_statistics` injected from stock [dgrauet/ltx-2.3-mlx](https://huggingface.co/dgrauet/ltx-2.3-mlx).
## Important
This is **not** a drop-in replacement for stock `vae_decoder.safetensors`. Channel widths and skip `conv_in` projections differ. Load with a Pruna-aware decoder graph (e.g. ltx-ws `VideoDecoderPruna` / ltx-2-mlx port of patched diffusers `LTX2VideoDecoder3d`).
## Files
| File | Role |
|------|------|
| `vae_decoder_pruna.safetensors` | Decoder weights (`vae_decoder.*` prefix) + per-channel stats |
| `vae_decoder_pruna_config.json` | Construct-time channel / upsample schedule |
| `split_model.json` | mlx-forge metadata |
## Usage (ltx-ws)
```bash
# Auto-download into ./models/ when --vae-decoder pruna
python server.py --vae-decoder pruna
# or: LTX_WS_VAE_DECODER=pruna
```
## Re-convert
```bash
mlx-forge convert pruna-vaed --output ./models/pruna-vaed-mlx
mlx-forge validate pruna-vaed ./models/pruna-vaed-mlx
```
## License
Follow upstream licenses for PrunaVAED and LTX-2.3.