CoolFace
Modelpublic

elismasilva/ltx2.3-image-distilled-1.1-sdnq-int8

sourceHugging Faceotherupdated 1mo agoView on Hugging Face
0likes9downloads
Model Card

LTX 2.3 Image Distilled 1.1 (SDNQ int8)

This repository contains an SDNQ int8 distilled image-only checkpoint derived from the distilled LTX 2.3 model by Lightricks.

It is a pruned image-only execution package: the model is organized for the LTX 2.3 image generation path and does not expose the original video/audio generation flows. The underlying model family and weights remain derived from LTX 2.3, and this package continues to inherit the original LTX Video 2 Open Source License.

What Changed

  • —Packaged for Diffusers-based image-only inference.
  • —Pruned to remove components that are not needed by the image execution path.
  • —Intended for the LTX2ImagePipeline and LTX2ImageImg2ImgPipeline implementations used by the DEVAIEXP image runners.
  • —Not a new training run and not a dataset modification.

Variant

This variant uses SDNQ int8 quantization for the image transformer.

  • —Higher-fidelity quantized distilled variant compared with uint4/int4 in local tests.
  • —Uses the tested mid/late preservation policy, which reduced latency versus the fully automatic int8 configuration while keeping output close to bf16 in the tested scenes.
  • —Recommended as the balanced quantized option when quality matters and lower VRAM use is still required.

Usage

For installation, environment setup, model path expectations, and runnable text-to-image / image-to-image examples, see:

https://github.com/DEVAIEXP/ltx2-image-diffusers

That repository uses a temporary Diffusers branch with the LTX 2.3 image pipelines until equivalent support is available upstream.

License

These weights are derived from Lightricks/LTX-2.3 and are released under the LTX Video 2 Open Source License:

https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE

Please review the original LTX 2.3 model card and license before redistribution or commercial use.