LTX Video Self-Hosted

LTX Video from Lightricks is a fast video generation model – roughly 5-10x faster than comparable models. Self-hosted on our dedicated GPU hosting it generates 5-second 720p clips in under a minute on a 5090.

Contents

• VRAM
• Deployment
• Performance
• Prompting tips

VRAM

~14 GB for LTX Video at FP16. With 5-second 720p generation and attention caching it fits a 24 GB+ card. For 30-frame clips at 512 resolution a 16 GB card can work.

Deployment

from diffusers import LTXPipeline
import torch

pipe = LTXPipeline.from_pretrained(
    "Lightricks/LTX-Video",
    torch_dtype=torch.bfloat16,
).to("cuda")

video = pipe(
    prompt="A golden retriever running on a beach at sunset, cinematic",
    num_frames=121,
    num_inference_steps=40,
    height=480,
    width=704,
).frames[0]

export_to_video(video, "output.mp4", fps=24)

Performance

Tips

• Detailed motion prompts work better than abstract descriptions
• Specify camera movement: “dolly in”, “static camera”, “tracking shot”
• Avoid requesting rapid multi-subject interactions – the model handles simple scenes best
• Image-to-video conditioning (starting from a single frame) often produces more coherent clips than pure text-to-video

See CogVideoX and Stable Video Diffusion.