LTX-2.3 INT8 Quantized Models
I have quantized various LTX-2.3 models and text encoders into INT8 to be used in ComfyUI.
Make sure you have INT8 nodes installed, as ComfyUI does not natively support INT8 as of Q1 2026.
These INT8 models are designed to speed up workflows on Ampere (RTX 30XX) GPUs, which are relatively inefficient with FP8 or FP4 quantization.
Demo Video
Usage
Example loading with Transformers:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("Winnougan/LTX-2.3-INT8")
model = AutoModelForSeq2SeqLM.from_pretrained("Winnougan/LTX-2.3-INT8")
These models are primarily intended for ComfyUI workflows using INT8-compatible nodes.
Upate Your ComfyUI to use the INT8 diffusion model and text encoder natively
Text Encoders (INT8)
In addition to the diffusion models, the following text encoders used by LTX-2.3 have also been quantized to INT8:
Gemma 3 12B (it) — INT8 quantized, embedding/lm_head/norm layers kept in high precision
Load the regular text projector
Both load via the standard Load CLIP node alongside the INT8 diffusion model.
Notes
- Quantization performed for performance on Ampere GPUs (RTX 30XX series).
- INT8 offers a good balance between speed, VRAM usage, and quality.
- Ensure your ComfyUI installation supports INT8 tensor loading.
Disclaimer
These are quantized versions of the original LTX-2.3 models.
All credit for the original models goes to their respective authors.
Quantization may introduce minor differences in output quality compared to FP16 models.