---
title: LTX2.3_comfy
canonical_url: "https://www.modelscope.cn/models/Kijai/LTX2.3_comfy"
md_url: "https://www.modelscope.cn/models/Kijai/LTX2.3_comfy.md"
repository: Kijai/LTX2.3_comfy
last_updated: 2026-07-29
license: other
base_model:
  - Lightricks/LTX-2.3
base_model_relation: finetune
parameters: 322.5B
tensor_type:
  - U8
  - I8
  - F16
  - BF16
  - F32
  - F8_E4M3
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 44009
stars: 62
tags:
  - comfyui
  - diffusion-single-file
---

# LTX2.3_comfy

> LTX2.3_comfy - Kijai 在 ModelScope 开源的模型。Separated LTX2.3 checkpoint for alternative way to load the models in Comfy

Kijai/LTX2.3_comfy 是 ModelScope 魔搭社区上的 322.5B 参数机器学习模型，采用 other 许可，基于 Lightricks/LTX-2.3 构建。

- **Repository**: Kijai/LTX2.3_comfy
- **License**: other
- **Parameters**: 322.5B
- **Base model**: Lightricks/LTX-2.3
- **Tags**: comfyui, diffusion-single-file
- **Downloads**: 44009
- **Stars**: 62
- **Last updated**: 2026-07-29

Source: https://www.modelscope.cn/models/Kijai/LTX2.3_comfy

---

Separated LTX2.3 checkpoint for alternative way to load the models in Comfy


![image](https://cdn-uploads.huggingface.co/production/uploads/63297908f0b2fc94904a65b8/KnDASZbjsLio38bS9XdVQ.png)


The fp8 quantizations were done with the basic static weight scales and are set to not run with fp8 matmuls, the models marked `input_scaled` additionally have activation scaling, and are set to run with fp8 matmuls on supported hardware (roughly 40xx and later Nvidia GPUs).

As this is first time I'm attempting to do calibrate input scales, these are pretty experimental, but result wise seems to work, this is a test on a 4090, 8 steps with distill:

<video controls autoplay width=50% src=https://cdn-uploads.huggingface.co/production/uploads/63297908f0b2fc94904a65b8/ALNr3_j0klp29fHkI3pyt.mp4></video>

**update:** input_scaled_v3 follows the same pattern as the official one where blocks 0-1 and 46-47 (first two and last two) are kept in bf16, as well as better calibrated input scales, this does fix some of the issues in v2, especially when using input audio.

---

Tiny VAE by [madebyollin](https://github.com/madebyollin/taehv/)

Can be used like this currently:

![image](https://cdn-uploads.huggingface.co/production/uploads/63297908f0b2fc94904a65b8/E032h6eXmjaUdzvEsHU_b.png)