LTX-2.5 Image-to-Video & Text-to-Video Workflow Guide

LTX-2.5 is an open-source AI model for image-to-video (I2V) and text-to-video (T2V) generation that runs locally on your machine. Built for production workflows, it generates synchronized video and audio with fast iteration times, modular control, and efficient VRAM usage.

This tutorial walks you through the official ComfyUI LTX-2.5 workflows, explains the multiscale rendering architecture, and shares optimization techniques for high-quality results.

Understanding LTX-2.5 Workflows

LTX-2.5 ships three official ComfyUI templates. Click Templates, search LTX-2.5, and press Download all to fetch the model files each one needs.

Image-to-Video (I2V). Animates a static image with motion, camera movement, and synchronized audio. Two-stage.

Text-to-Video (T2V). Generates complete videos from text prompts alone, no input image required. Two-stage.

First-Frame / Last-Frame (FLF). Generates the motion between a starting image and an ending image. Single-stage, so no upscaler and no 2x pass.

The two-stage templates share the same pipeline structure:

1. Load model components
2. Configure video parameters (resolution, frame count, frame rate)
3. Write and optionally enhance your prompt
4. Generate low-resolution base video
5. Upscale to final output resolution

Image-to-Video Workflow (Distilled Model)

The distilled I2V workflow is optimized for speed and lower VRAM consumption, ideal for rapid iteration and local development.

Step 1: Load Model Components

The template loads these components:

• LTX-2.5 distilled transformer, loaded via UNETLoader into models/diffusion_models/
• Gemma 4 text encoder, loaded via CLIPLoader with type ltxv into models/text_encoders/
• Gemma 4 prompt enhancer, also a CLIPLoader, optional
• Video VAE, which decodes the generated frames
• Audio VAE, which decodes the generated soundtrack
• Spatial upscaler, which handles the 2x pass, into models/latent_upscale_models/

LTX-2.5 uses separate VAEs for picture and sound, so both files are needed for any generation with audio. The FLF template is single-stage and does not need the upscaler.

Step 2: Configure Video Parameters

Define your output specifications:

• Frame count. Must be one more than a multiple of 8. The template default is 97 frames, roughly 4 seconds at 24fps. 121 frames gives about 5 seconds.
• Resolution. The template defaults to 768x512. Width and height must be divisible by 32, which rules out labels like 720 and 1080 as raw dimensions.
• Frame rate. Must match the frame rate in your Create Video node. 24fps is the default and the model supports up to 50fps.

Remember that the two-stage templates upscale 2x, so the final output is double the width and height you configure. A 768x512 base produces 1536x1024.

Step 3: Write Effective Prompts

Prompting quality directly impacts output quality. LTX-2.5 works best with detailed, structured prompts that specify visual style, character appearance and actions, camera motion, dialogue or voiceover, and background music or sound effects. Put spoken dialogue in quotation marks.

In image-to-video the prompt should describe what happens, not what the scene looks like. The model already has the visual context from your image.

A single prompt can also describe several connected shots joined by explicit cuts, which the model generates as one output while holding character, setting, and voice across each cut. Two to four shots per generation is the working range.

For detailed prompting strategies and examples, see the LTX-2.5 Prompting Guide.

Prompt Enhancement. By default, prompts pass through the Prompt Enhancer, which expands your input before it reaches the text encoder. Changing the enhancer's seed generates different variations. Turn Prompt Enhance off to use your exact wording, which is usually what you want once you're iterating on a specific shot.

How Multiscale Rendering Works

LTX-2.5 uses a multiscale rendering architecture that functions as a built-in upscaling workflow. Instead of rendering at full resolution immediately, it generates a low-resolution base video, then upscales 2x with the spatial upscaler and refines at full resolution.

In the I2V template, the source image is injected as conditioning at partial strength in Stage 1, establishing the starting point while leaving room for motion, then re-injected at full strength in Stage 2 to preserve detail through the upscale.

Why This Matters for Developers: faster iteration, lower VRAM usage, and easier experimentation across multiple variations without long render times.

VRAM-Friendly Development Tip

As long as you keep the random seed fixed:

1. Generate and save the low-resolution preview
2. Evaluate motion, pacing, and audio synchronization
3. Skip upscaling until you're satisfied with the result

This approach is essential for working with LTX-2 on machines with limited VRAM.

Audio-Video Generation Pipeline

LTX-2.5 handles audio and video through a split-merge process:

1. An empty audio latent is concatenated with the video latent
2. The two are sampled together, which is what keeps them in sync
3. Video and audio are split and decoded separately, video through tiled VAE decoding and audio through the audio VAE
4. The two are merged into the final file

Tiled decoding significantly reduces VRAM consumption during final rendering while maintaining quality. If you hit out-of-memory errors at the decode stage specifically, increase the tile count.

The output is a single video file with synchronized audio and video tracks.

Customizing with LoRAs

LoRAs let you modify LTX-2.5 behavior without retraining the base model. Add a LoRALoader node to apply style, motion, or character LoRAs.

Strength. Start effect LoRAs at 1.0. Roughly, 0.9 to 1.1 is subtle, 1.2 to 1.4 is balanced and suits most cases, and 1.5 to 1.6 is maximum style transfer. Keep total combined strength under 2.0 when stacking, and use 1.0 for IC-LoRA control adapters, which are designed for full strength.

Match the LoRA to the model. Using a LoRA trained on a different LTX generation is possible but can degrade output. Check which model an adapter was trained against before relying on it in production.

IC-LoRA. For structural control, restoration, and creative transforms, the repository bundles advanced workflows under example_workflows/2.5/, covering IC-LoRA control and in/outpainting.

Training custom LoRAs. The LTX Trainer is part of the LTX-2 repository on GitHub and supports LoRA, IC-LoRA, and full fine-tuning.

Full Model vs Distilled Workflow

LTX-2.5 offers a full model workflow designed for maximum fidelity and stronger prompt adherence.

Key Differences:

Distilled. The base model with the distilled LoRA applied, 8 steps. Fast, and the right default for iteration.

Full. The base model with the distilled LoRA at weaker strength, 15 steps in a second-order sampler for roughly 30 effective steps, with separate guiders for audio and video. Slower, higher fidelity.

Recommended workflow. Iterate with the distilled model for fast feedback, then render finals with the full model.

CFG configuration. This differs by path and it matters. The distilled model bakes guidance into distillation, so both stages run at CFG 1 and raising it adds overhead without improving output. Stay in the 1.0 to 1.5 range if you experiment. The full model responds to CFG normally, in the 2.0 to 5.0 range, where higher values improve prompt adherence and values that are too high introduce unwanted textures.

Text-to-Video Workflow

The T2V workflow follows the same pipeline as I2V but without an input image.

Setup Steps

1. Load model components (same as I2V)
2. Configure resolution and frame count
3. Write a detailed text prompt
4. Generate base video at low resolution
5. Upscale and refine to final output

Critical Difference:
Without a visual reference image, prompt detail becomes even more important, and the multi-shot guidance above applies most usefully here. Longer, more descriptive prompts consistently produce better T2V results. For comprehensive prompting techniques tailored to LTX-2.5.

Best Practices

General optimization:

• Write long, descriptive prompts, and include the audio
• Match frame rates across all nodes to avoid sync issues
• Preview at low resolution first to save time and VRAM
• Set CFG to 1 on the distilled path and around 4 on the full model
• Increase VAE decode tile count if you hit memory limits at decode

Development workflow:

• Use distilled workflows for iteration and experimentation
• Switch to the full model for final production renders
• Fix random seeds when comparing variations
• Describe camera movement explicitly in the prompt

Getting Started

LTX-2.5 combines open weights, local execution, synchronized audio-video generation, and first-party ComfyUI integration, making it one of the most practical open-source AI video systems for developers.

Next steps:

1. Open ComfyUI, click Templates, and search LTX-2.5
2. Pick Text-to-Video, Image-to-Video, or First-Frame / Last-Frame and press Download all
3. Experiment with the distilled model for fast iteration
4. Explore the advanced IC-LoRA workflows and LoRA training to customize behavior for your use case