Lightricks has released LTX-2.5, the latest generation of its open video generation model, and ComfyUI already has day-one support for the new model.
If you’re thinking about purchasing a new GPU, we’d greatly appreciate it if you used our Amazon Associate links. The price you pay will be exactly the same, but Amazon provides us with a small commission for each purchase. It’s a simple way to support our site and helps us keep creating useful content for you. Recommended GPUs: RTX 5090, RTX 5080, and RTX 5070. #ad
The latest release is more than a minor upgrade. LTX-2.5 introduces a substantially redesigned generation pipeline, including a new rendering approach, a new video decoder, a custom text encoder, an improved distilled model, and native multi-shot generation.
For ComfyUI users, this means LTX-2.5 is positioned as a powerful new option for generating high-quality video locally while retaining the flexibility of open weights and custom workflows.
LTX-2.5 Comes to ComfyUI on Day One
The ComfyUI team announced support for LTX-2.5 on August 12, 2026, the same day the new model became available.
LTX-2.5 continues Lightricks’ focus on making high-quality video generation accessible through local hardware. The model weights are available for download, and the model can be fine-tuned using users’ own data.
ComfyUI currently supports two open-weight versions:
- LTX-2.5 Dev — the primary full model.
- LTX-2.5 Distilled — a smaller and faster version designed for more economical inference.
The new distilled version has also been substantially reworked. According to the ComfyUI announcement, it retains more quality, prompt adherence, and motion than previous distilled releases, making it more practical when running the full model is too computationally expensive.
ComfyUI also provides hosted versions through Partner Nodes:
- LTX-2.5 Fast — supports 2–20 second generations, from 720p through 4K, in landscape or portrait formats.
- LTX-2.5 Pro — supports 2–10 second generations at 720p or 1080p.
The Fast version supports 24, 25, 48, and 50 fps, while longer-than-10-second generations are limited to 720p or 1080p at 24 or 25 fps.
Diffusion Fidelity Rendering Is the Biggest Change
One of the most significant changes in LTX-2.5 is Diffusion Fidelity Rendering.
Instead of distributing computational resources evenly throughout the video, LTX-2.5 attempts to allocate more computation to visually complex portions of a scene.
The model first generates the overall structure — including motion, composition, and framing — in an 8× temporally compressed latent space. It simultaneously generates high-fidelity keyframes, with more keyframes allocated to complex scenes and fewer to simpler sections.
A dedicated pixel-diffusion stage then uses this structural information and the keyframes to render the final video.
This approach is designed to improve fine details such as:
- Facial details
- Textures
- Materials
- Intricate objects
- Fine visual details in complex scenes
The model can also devote more rendering computation to visually complicated shots instead of wasting the same amount of compute on every part of the video.
This is an important architectural change because video generation has traditionally involved a difficult tradeoff between speed, resolution, temporal consistency, and fine detail.
A New Video Decoder Improves Fine Details
LTX-2.5 also replaces its previous standard VAE decoding approach with a Diffusion Video Decoder.
According to ComfyUI, the new decoder is designed to produce sharper faces, more legible text, and fewer smearing artifacts during fast motion.
This should be particularly useful for scenes involving:
- Fast-moving characters
- Detailed environments
- Facial close-ups
- Text and signage
- Complex materials
- Rapid camera movement
These are precisely the types of details that can expose weaknesses in AI-generated video.
Native Multi-Shot Generation
Perhaps one of the most interesting features for creators is native multi-shot generation.
Rather than generating individual clips and attempting to stitch them together afterward, LTX-2.5 can generate multiple connected shots in a single generation.
The model is designed to maintain consistency of:
- Characters
- Environment
- Lighting
- Voice
- Visual style
across the cuts.
This could make LTX-2.5 particularly interesting for short-form storytelling.
For example, instead of generating five separate three-second clips and trying to maintain the same character and environment across all five generations, a creator can describe a sequence containing multiple shots and allow the model to generate the sequence as a unified piece.
This doesn’t eliminate the need for editing, but it potentially reduces one of the biggest problems in AI video workflows: matching independently generated clips.
Custom Gemma 4 12B Text Encoder
LTX-2.5 also introduces a custom Gemma 4 12B text encoder designed specifically for the model.
The purpose is to improve the model’s ability to understand complicated prompts containing multiple subjects, actions, lighting conditions, and camera directions.
This is particularly relevant for ComfyUI users who create detailed video prompts.
Long prompts often contain many independent instructions. A model may correctly follow the first few instructions but gradually lose track of details later in the prompt.
LTX-2.5 is specifically designed to retain more of these details as prompt complexity increases.
Built-In Prompt Enhancement
LTX-2.5 also includes a dedicated prompt enhancer.
Instead of requiring users to manually write extremely detailed cinematic prompts, the lightweight enhancer can expand shorter prompts into more detailed instructions before generation.
According to ComfyUI, the additional computational cost is minimal.
This could make the model more approachable for users who don’t want to manually specify every camera movement, environmental detail, and action.
Advanced users may still prefer writing their own prompts for maximum control, but the prompt enhancer could be useful for quickly experimenting with ideas.
Automatic Video Duration
Another interesting addition is Auto Duration.
LTX-2.5 includes an experimental duration head that analyzes the action described in the prompt and predicts an appropriate video length before generation begins.
This means users may not always need to manually determine the ideal duration for a particular prompt.
Combined with the prompt enhancer, the system can take more responsibility for translating a natural-language description into generation parameters.
Native 4K and 50 FPS Remain
LTX-2.5 also carries forward several capabilities from the previous generation.
These include:
- Native 4K output
- Synchronized audio and video
- Up to 50 FPS generation
LTX-2 had already established Lightricks’ approach of combining video and audio generation in a single open model. The original LTX-2 release included open model weights and training code and was designed for local deployment and customization.
LTX-2.5 builds on that foundation rather than abandoning it.
ComfyUI Installation
Getting started with LTX-2.5 in ComfyUI is relatively straightforward.
First, update ComfyUI to version 0.32.0 or newer. The official ComfyUI announcement also notes that LTX-2.5 is available through Comfy Cloud.
Next, download the LTX-2.5 model weights and place them in the appropriate ComfyUI models directory.
ComfyUI provides templates for:
- LTX-2.5 T2V
- LTX-2.5 I2V
- LTX-2.5 FLF2V
These workflows can be loaded directly from the Templates panel or downloaded separately.
Download LTX-2.5 models from Hugging Face
LTX-2.5 vs. Previous LTX Models
The biggest difference is that LTX-2.5 isn’t simply focused on making the existing architecture faster or slightly higher quality.
It represents a broader redesign of the generation pipeline.
| Feature | LTX-2.5 |
|---|---|
| Open weights | Yes |
| Local generation | Yes |
| Native 4K | Yes |
| Audio + video | Yes |
| Up to 50 FPS | Yes |
| Native multi-shot | Yes |
| Diffusion Fidelity Rendering | Yes |
| Diffusion Video Decoder | Yes |
| Custom Gemma 4 12B text encoder | Yes |
| Prompt enhancer | Yes |
| Auto duration | Experimental |
| Distilled model | Yes |
| ComfyUI support | Day one |
Why LTX-2.5 Matters for ComfyUI Users
The combination of open weights, local inference, fine-tuning, native multi-shot generation, and improved rendering makes LTX-2.5 particularly interesting for the ComfyUI ecosystem.
Creators who already use LTX workflows can experiment with the new architecture without moving to a completely closed video-generation service.
More importantly, the native multi-shot capability could change how creators approach AI video production. Instead of treating each generation as an isolated clip, LTX-2.5 moves closer to generating an entire short sequence as a coherent visual unit.
The new rendering pipeline is equally interesting. By allocating compute according to scene complexity and adding a dedicated pixel-diffusion stage, LTX-2.5 attempts to improve visual detail without simply increasing computational cost across the entire video.
For users with capable local GPUs, this makes LTX-2.5 one of the most interesting open video models to test in ComfyUI right now.
Final Thoughts
LTX-2.5 is a substantial upgrade to the LTX family.
Its biggest improvements aren’t limited to resolution or generation speed. The new Diffusion Fidelity Rendering, Diffusion Video Decoder, native multi-shot generation, custom Gemma 4 12B text encoder, and improved distilled model address several of the fundamental challenges of AI video generation.
The fact that ComfyUI supports LTX-2.5 from day one is also significant. Users can immediately experiment with T2V, I2V, and FLF2V workflows rather than waiting for the ecosystem to catch up.
For anyone already using LTX-2 or other open video models in ComfyUI, LTX-2.5 is definitely worth testing.
Read the official ComfyUI announcement: LTX-2.5 Day-0 Support in ComfyUI
Leave a Reply