---
license: other
license_name: ltx-2-community-license-agreement
license_link: https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE
base_model: Lightricks/LTX-2.3
library_name: diffusers
pipeline_tag: text-to-video
tags:
  - ltx-video
  - ltx-2.3
  - int8
  - convrot
  - sdnq
  - audio-video
---

# LTX-Video 2.3 distilled — INT8 ConvRot

This repository contains a transformer-only INT8 ConvRot quantization of the official
`Lightricks/LTX-2.3` distilled checkpoint. It is derived directly from
`ltx-2.3-22b-distilled.safetensors`; no fine-tune, task adapter, or third-party model
weights were merged.

## Artifact

- File: `ltx-2.3-convrot-int8.safetensors`
- Size: 19,485,049,192 bytes
- SHA-256: `722e2498f59306aa007f25560d6069f621ae57c90705cdaa0e2de4a1db996d05`
- Source SHA-256: `14409a4d1337a8ded02fa87fb895b17a91ab2c6588f7cc3352e624ff18a689bf`

The file contains the main audiovisual transformer only. Load the text encoder,
connectors, VAE, audio VAE, and vocoder from the official LTX 2.3 Diffusers pipeline.
Weights use Comfy-style per-layer `int8_tensorwise` sidecars and are consumed by an
SDNQ Hadamard/ConvRot loader.

## Quantization

The checkpoint was quantized on an NVIDIA H200 with
`Comfy-Org/comfy-model-tools` commit
`1fe341bb8a4e46f161a978b5faa2412d8c39c768`:

```bash
python quant_int8_convrot.py   ltx-2.3-22b-distilled.safetensors   ltx-2.3-22b-distilled-int8-convrot.safetensors   --mseclip   --verify-report quantization-report.tsv
```

The packaging step retained `model.diffusion_model.*` tensors and omitted the bundled
connector copies, since the target Diffusers pipeline supplies its connector components
separately.

- ConvRot group size: 256
- Quantized transformer linears: 1,344
- Mean relative weight reconstruction error: 0.859%
- Maximum relative weight reconstruction error: 1.446%
- Worst per-layer cosine similarity: above 0.99990

The full per-layer measurements are in `quantization-report.tsv`.

## Validation

The artifact was loaded with SDNQ 0.2.3 into
`LTX2VideoTransformer3DModel` with all 1,344 quantized parameters wrapped and no
remaining meta tensors. A real audiovisual transformer forward at 768×512, 121 frames,
and 1,024 text tokens produced finite video and audio outputs. On the calibration H200,
three warmed steps averaged 0.5493 seconds with 19.139 GiB peak allocated memory.
These are transformer-only measurements, not end-to-end generation timings.

## License

This is a derivative of LTX-2.3 and remains subject to the
[LTX-2 Community License Agreement](LICENSE). Review that agreement, including its
commercial-use and acceptable-use restrictions, before using or redistributing the
checkpoint.