Newly Improved LTX Video v0.9.6 Video Generation Model - Workflow & Installer
$4.00
✓ In Stock🧡 Get all 62 installers for one monthly price
Join The Local Lab on Patreon and get instant access to every installer we've ever made — plus every new one we release. No buying individually, no waiting. One membership, everything included.
View Patreon membership →Opens in our store — secure checkout via Wix
// About this product
What’s New in v0.9.6?
-
Ultra-fast inference: Up to 15x faster than previous versions, generating high-quality videos in seconds with just 8 diffusion steps recommended.
-
Improved quality: Enhanced prompt adherence, smoother motion, and finer details for more realistic and coherent video outputs.
-
New default resolution & FPS: 1216 × 704 pixels at 30 FPS for crisp, fluid videos.
-
Distilled model available: Smaller, faster, and still delivers stunning results without needing classifier-free or spatio-temporal guidance.
-
Versatile generation modes: Text-to-video, image-to-video, and video-to-video support with updated, intuitive workflows.
Why Upgrade?
The models support sampling with 8 (recommended), 4, 2, or 1 diffusion steps.
On my RTX 4050 6GB GPU, the distilled model generates videos in under a minute with surprisingly coherent outputs. Higher VRAM cards will likely achieve even faster speeds.
Get Started
-
Download the models on Hugging Face:
-
Full model: ltxv-2b-0.9.6-dev-04-25.safetensors
-
Distilled model: ltxv-2b-0.9.6-distilled-04-25.safetensors
-
https://huggingface.co/Lightricks/LTX-Video/tree/main
-
-
Explore the GitHub repo with updated code and workflows:
https://github.com/Lightricks/LTX-Video -
Try it with ComfyUI or your preferred pipeline for seamless integration and best performance.
As always, I created a one click windows installer for my Patreon/Youtube members along with custom workflow which includes a range of slightly different settings compared to the original. Notably, I removed the prompt enhancer from the original workflow because it requires the unquantized LLaMA 3.2 model, which can be too demanding for lower VRAM devices. Instead, I added an Ollama node that can be prompted to analyze images with Ollama vision models and enhance prompts for generation. Additionally, I replaced the VAE decode node with the VAE decode tiled node to improve generation output times, remove the STG guider and replaced it with a CFG guider and finally added custom stigmas float values(which i pulled from the amazing workflow - https://civitai.com/models/1482620/private-modified-workflow-for-ltxv-096-distilled) to improve video generation quality and stability.
You can find the Patreon post here - https://www.patreon.com/posts/newly-improved-9-126980363