ComfyUI-LTXVideo is a collection of custom nodes for ComfyUI, designed to provide useful tools for working with the LTXV model. The model itself is supported in the core ComfyUI code. The main LTXVideo repository can be found here.
- LTXV Patcher VAE
The new node improves VAE decoding performance by reducing runtime and cutting memory consumption by up to 50%. This allows generation of higher-resolution outputs on consumer-grade GPUs with limited VRAM, without needing to load the VAE partially or decode in tiles.
⚠️ On Windows, you may need to add the paths to the MSVC compiler (cl.exe) and ninja.exe to your system environment PATH variable. - LTXV Preprocess Masks
Preprocesses masks for use with the LTXVideo model's latent masking. It validates mask dimensions based on VAE downscaling, supports optional inversion, handles the first frame mask separately, combines temporal masks via max pooling, applies morphological operations to grow or shrink masks, and clamps values to ensure correct opacity. The result is a set of masks ready for latent-space masking. - LTXV Q8 Lora Model Loader
Applying LoRA to an FP8-quantized model requires special handling to preserve output quality. It's crucial to apply LoRA weights using the correct precision, as the current LoRA implementation in ComfyUI does so in a non-optimal way. This node addresses that limitation by ensuring LoRA weights are applied properly, resulting in significantly better quality. If you're working with an FP8 LTXV model, using this node guarantees that LoRA behaves as expected and delivers the intended effect.
-
LTXV 13B Distilled 🥳 0.9.7
Delivers cinematic-quality videos at fraction of steps needed to run full model. Only 4 or 8 steps needed for single generation.
👉 Download here -
LTXV 13B Distilled Quantized 0.9.7
Offers reduced memory requirements and even faster inference speeds. Ideal for consumer-grade GPUs (e.g., NVIDIA 4090, 5090).
Important: In order to get the best performance with the quantized version please install q8_kernels package and use dedicated flow below.
👉 Download here
🧩 Example ComfyUI flow available in the Example Workflows section. -
Updated LTV 13B Quantized version
From now on all our 8bit quantized models are running natively in ComfyUI, still with our Q8 patcher node you will get the best inference speed.
👉 Download here
-
LTXV 13B 0.9.7 Delivers cinematic-quality videos at unprecedented speed.
👉 Download here -
LTXV 13B Quantized 0.9.7 Offers reduced memory requirements and even faster inference speeds. Ideal for consumer-grade GPUs (e.g., NVIDIA 4090, 5090). Delivers outstanding quality with improved performance.
Important: In order to run the quantized version please install LTXVideo-Q8-Kernels package and use dedicated flow below. Loading the model in Comfy with LoadCheckpoint node won't work.
👉 Download here
🧩 Example ComfyUI flow available in the Example Workflows section. -
Latent Upscaling Models Enables inference across multiple scales by upscaling latent tensors without decoding/encoding. Multiscale inference delivers high-quality results in a fraction of the time compared to similar models.
Important: Make sure you put the models below in models/upscale_models folder.
👉 Spatial upscaling: Download here.
👉 Temporal upscaling: Download here.
🧩 Example ComfyUI flow available in the Example Workflows section.
- New simplified flows and nodes
1.1. Simplified image to video: Download here.
1.2. Simplified image to video with extension: Download here.
1.3. Simplified image to video with keyframes: Download here.
- LTXV 0.9.6 – higher quality, faster, great for final output. Download from here.
- LTXV 0.9.6 Distilled – our fastest model yet (only 8 steps for generation), lighter, great for rapid iteration. Download from here.
We introduce the STGGuiderAdvanced node, which applies different CFG and STG parameters at various diffusion steps. All flows have been updated to use this node and are designed to provide optimal parameters for the best quality. See the Example Workflows section.
- Improved quality with reduced artifacts.
- Support for higher resolution and longer sequences.
- Frame and sequence conditioning (beyond the first frame).
- Enhanced prompt understanding.
- Commercial license availability.
Since LTXVideo is now fully supported in the ComfyUI core, we have removed the custom model implementation. Instead, we provide updated workflows to showcase the new features:
- Frame Conditioning – Enables interpolation between given frames.
- Sequence Conditioning – Allows motion interpolation from a given frame sequence, enabling video extension from the beginning, end, or middle of the original video.
- Prompt Enhancer – A new node that helps generate prompts optimized for the best model performance. See the Example Workflows section for more details.
The LTXTricks code has been integrated into this repository (in the /tricks folder) and will be maintained here. The original repo is no longer maintained, but all existing workflows should continue to function as expected.
Fixed a bug which caused the model to produce artifacts on short negative prompts when using a native CLIP Loader node.
- Improved model - removes "strobing texture" artifacts and generates better motion. Download from here.
- STG support
- Integrated image degradation system for improved motion generation.
- Additional initial latent optional input to chain latents for high res generation.
- Image captioning in image to video flow.
Installation via ComfyUI-Manager is preferred. Simply search for ComfyUI-LTXVideo in the list of nodes and follow installation instructions.
- Install ComfyUI
- Clone this repository to
custom-nodesfolder in your ComfyUI installation directory. - Install the required packages:
cd custom_nodes/ComfyUI-LTXVideo && pip install -r requirements.txtFor portable ComfyUI installations, run
.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-LTXVideo\requirements.txt
- Download ltx-video-2b-v0.9.1.safetensors from Hugging Face and place it under
models/checkpoints. - Install one of the t5 text encoders, for example google_t5-v1_1-xxl_encoderonly. You can install it using ComfyUI Model Manager.
Note that to run the example workflows, you need to have some additional custom nodes, like ComfyUI-VideoHelperSuite and others, installed. You can do it by pressing "Install Missing Custom Nodes" button in ComfyUI Manager.
🧩 Image to video mixed: mixed flow with full and distilled model for best quality and speed trade-off.
🧩 Image to video
🧩 Image to video with keyframes
🧩 Image to video with duration extension
🧩 Image to video 8b quantized
🧩 Image to video
🧩 Image to video with keyframes
🧩 Image to video with duration extension
🧩 Image to video 8b quantized
