LTX Video 2.3 on RTX 4090: From 300s to 45s with ComfyUI

What You'll Build

Generate videos locally with LTX Video 2.3 — the fastest open-source video model. With INT8 optimization on Ampere GPUs, generation drops from 300 seconds to 45 seconds per clip.

Performance: Stock: ~300s per clip · With INT8: ~45s per clip · Peak VRAM: 16GB · See all data

LTX Video vs Wan 2.1: LTX is 5–7× faster but lower quality. Use LTX for rapid iteration and Wan for final renders. Compare →

Requirements

Component	Minimum	Tested
GPU	RTX 4060 Ti 8GB	RTX 4090 (24GB)
VRAM	8GB	16–24GB
RAM	16GB	32GB
Storage	15GB	15GB

8GB users: LTX 2.3 is specifically designed to run on 8GB GPUs — it's one of the few video models that fits.

Installation

1. Install ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt

2. Install LTX Video Nodes

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo
cd ComfyUI-LTXVideo
pip install -r requirements.txt

3. Download Model Weights

huggingface-cli download Lightricks/LTX-Video \
  ltx-video-2b-v0.9.5.safetensors \
  --local-dir ./models/ltxvideo/

Place in ComfyUI/models/ltxvideo/

4. (Optional) INT8 for Ampere GPUs — 6× Speedup

If you have an Ampere GPU (RTX 30-series or 40-series), install the INT8 loader:

cd ComfyUI/custom_nodes
git clone https://github.com/overpresentme/ComfyUI-ltx-int8-loader

# Download INT8 weights
huggingface-cli download ovpresent/ltx-2.3-distilled-1.1-INT8 \
  --local-dir ./models/ltxvideo-int8/

Note: INT8 acceleration is effective for Ampere GPUs. RTX 50-series users (Blackwell) already have native FP8 support and may not see additional benefit.

Running

Start ComfyUI:

python main.py --listen

Load a workflow from ComfyUI-LTXVideo/examples/

Standard Workflow Settings

Parameter	Value
Steps	25–40
Resolution	480×832 (9:16) or 832×480 (16:9)
Duration	2–5 seconds
CFG	3.0
Sampler	DPM++ 2M

Performance Comparison

Config	Time per Clip	VRAM
Stock FP16	~300s	16GB
INT8 (Ampere)	~45s	~12GB
Distilled 1.1	~60-90s	14GB

Source: community benchmarks. Full LTX data →

Using the Distilled Version

LTX 2.3 Distilled generates in fewer steps with similar quality:

huggingface-cli download Lightricks/LTX-Video \
  ltx-video-2b-v0.9.5-distilled.safetensors \
  --local-dir ./models/ltxvideo/

Use 8–12 steps instead of 25–40 for 3× faster generation.

LTX Audio Mode (New in 2.3)

LTX 2.3 includes experimental audio-driven video generation:

Install audio processing nodes (included in ComfyUI-LTXVideo)
Use the ltx_audio_workflow.json example
Provide an audio file and let LTX synchronize motion to it

Text-to-Video vs Image-to-Video

Mode	Use When
T2V	You want full creative control
I2V	You have a starting frame/image

LTX 2.3 supports both — select the appropriate workflow from the examples folder.

Troubleshooting

OOM on 8GB GPU: Reduce resolution to 320×512, use 2s duration, and enable --lowvram

Slow generation without INT8 speedup: INT8 only helps Ampere (RTX 30xx/40xx) — check your GPU generation

Flickering video: Reduce motion guidance, lower CFG to 2.0–2.5

Poor motion quality: Increase steps to 40, add motion keywords to prompt