LTX 2.5 - Text to Video

LTX-2.5 Audio & Video Generation ComfyUI Workflow

A ComfyUI video generation workflow powered by LTX-2.5, featuring a 22B INT8-quantized diffusion transformer, Gemma 4 text conditioning, dedicated video and audio autoencoders, latent spatial upscaling, and optional AI-powered prompt enhancement.

Workflow Overview

This workflow utilizes the LTX-2.5 model architecture for AI-powered video generation within ComfyUI. It combines a distilled 22B diffusion transformer, Gemma 4 12B conditioning, separate BF16 video and audio autoencoders, and a dedicated latent spatial upscaling model.

The primary diffusion model uses INT8 quantization to help reduce memory requirements compared with higher-precision model weights. The selected model is identified as a distilled transformer configuration, providing the main generative component of the workflow.

The workflow includes configurable video duration, frame rate, image dimensions, and noise seed settings. The displayed configuration uses a duration of 5 seconds and a frame rate of 24 FPS. Width and height are exposed as external inputs, allowing connected nodes to determine the generation resolution.

An optional Gemma 4 prompt enhancement model is also included. This component can support AI-assisted prompt processing when enabled. Prompt enhancement is disabled in the current configuration, while the primary Gemma 4 12B conditioning model remains selected.

Generation Settings

Model Configuration

Video Generation Configuration

The LTX-2.5 workflow provides controls for video duration, frame rate, and output dimensions. These parameters define the temporal and spatial configuration of generated video. The displayed workflow is configured for a duration of 5 seconds at 24 frames per second.

The requested duration and frame rate provide the temporal settings for generation. The exact internal frame count may depend on the model's frame handling requirements. Since width and height are externally connected, the final resolution cannot be determined from the visible subgraph alone.

Prompt Enhancement Configuration

The workflow includes optional AI-powered prompt enhancement using an additional Gemma 4 model. This feature can be used to refine, restructure, or expand user instructions before video generation.

The prompt enhancement model is separate from the main Gemma 4 12B conditioning model. The enhancement component provides optional prompt processing, while the primary conditioning model supplies representations used by the diffusion transformer. Prompt enhancement is disabled in the current workflow configuration.

Latent Spatial Upscaling

The workflow includes a dedicated LTX-2.5 latent spatial upscaler configured as a 2× model. Latent upscaling operates on compressed representations rather than directly resizing decoded video frames.

The 2× designation indicates that the selected model is intended for spatial resolution enhancement. When integrated into the appropriate generation pipeline, latent upscaling can increase video resolution before the final decoding stage. The screenshot confirms the selected upscaling model but does not establish whether the upscaling stage is active.

Video and Audio Processing

The LTX-2.5 workflow includes separate video and audio autoencoders. These components are designed to process their respective latent representations within compatible generation pipelines.

The separate audio and video autoencoders provide components for handling both modalities in the generation pipeline. The exact audio generation, synchronization, and final output handling depend on the internal workflow connections.

Noise Seed Configuration

The workflow includes a configurable noise seed that controls random initialization during the video generation process. The displayed seed value is 558811532553686.

Using a fixed seed can help reproduce generation results when the model, prompt, resolution, and other settings remain unchanged. Changing the seed modifies random initialization and may produce different video variations.

Additional Parameters

Surrounding Nodes and Workflow Integration

The LTX-2.5 subgraph consolidates diffusion model selection, text conditioning, audio and video autoencoders, frame rate, duration, noise seed, and optional enhancement models into a reusable ComfyUI workflow component.

Width and height are exposed as external input connections. These inputs allow upstream nodes to determine the configured video dimensions. The duration and frame rate parameters control the requested temporal configuration of the generated video.

In a typical LTX-2.5 video generation workflow, text instructions are processed through compatible conditioning nodes. The selected Gemma 4 12B model provides conditioning representations that can guide the diffusion transformer.

When enabled, the optional Gemma 4 prompt enhancement model can provide an additional text-processing stage. In the displayed configuration, prompt enhancement is disabled. The main text conditioning model remains part of the selected model configuration.

The LTX-2.5 diffusion transformer performs the video generation process using the supplied conditioning and generation parameters. The noise seed controls random initialization, while the duration and frame rate settings define the requested video sequence.

The selected latent spatial upscaler provides an additional model for spatial resolution enhancement. When connected through compatible workflow nodes, it can process latent video data before final decoding. The screenshot does not show the internal upscaling connections or activation state.

The video VAE provides the component for decoding generated video latent representations into visible frames. The audio VAE provides corresponding capabilities for compatible audio decoding operations.

Generated video and audio data can be passed to downstream preview, processing, synchronization, or export nodes. Exact sampler settings, internal generation stages, output connections, and audio-video synchronization behavior are not visible in the supplied screenshot.

Workflow Highlights

• ✓ LTX-2.5: Uses the LTX-2.5 diffusion transformer for AI video generation.
• ✓ 22B Distilled Model: Includes a distilled 22B diffusion transformer as the primary generation component.
• ✓ INT8 Quantization: Helps reduce diffusion model memory requirements compared with higher-precision configurations.
• ✓ Gemma 4 12B: Provides text conditioning for the LTX-2.5 generation process.
• ✓ 5-Second Video: Uses a configurable duration currently set to five seconds.
• ✓ 24 FPS: Configures video playback at 24 frames per second.
• ✓ Video Autoencoder: Uses a dedicated BF16 video VAE for latent decoding.
• ✓ Audio Autoencoder: Includes a separate BF16 audio VAE for compatible audio processing.
• ✓ 2× Latent Upscaler: Includes a BF16 spatial upscaling model for compatible latent video processing.
• ✓ Optional Prompt Enhancement: Includes an additional Gemma 4 model for prompt processing, currently disabled.
• ✓ Flexible Resolution: Width and height are controlled through external node connections.
• ✓ Seed-Based Generation: Includes configurable noise initialization with the displayed seed set to 558811532553686.
• ✓ Reusable Subgraph: Organizes model configuration, video settings, autoencoders, and optional enhancement components for integration into larger ComfyUI workflows.