Complete LTX-2.5 ComfyUI Workflow Guide: T2V, I2V, FLF2V & Multi-Shot Setup

LTX-2.5 from Lightricks represents a major breakthrough in open-weights video generation. Designed as a DiT-based world model, LTX-2.5 delivers synchronized high-fidelity video and 32 kHz stereo audio in a single pass. With novel features like Diffusion Fidelity Rendering (DFR), native multi-shot scene continuity, a fine-tuned Gemma 4 12B text backbone, and a built-in Prompt Enhancer, LTX-2.5 offers studio-grade production capabilities directly on local hardware.

With official native support available in ComfyUI, setting up LTX-2.5 is straightforward. In this step-by-step guide, we walk you through everything: resolving Hugging Face gated repository access, organizing split model checkpoints, mastering the three official workflows (T2V, I2V, FLF2V), writing effective multi-shot prompts, and tuning performance for 16GB+ VRAM GPUs.

1. Introduction & Key Capabilities

Before setting up your canvas, let me introduce what makes LTX-2.5 unique for local ComfyUI creators:

Key Upgrades in LTX-2.5

• Diffusion Video Decoder (DFR): Replaces static VAE decoding with a generative diffusion reconstruction pass that sharpens faces, skin pores, clothing weave, and text legibility.
• Native Multi-Shot Generation: Maintains character identity, lighting, background geometry, and voice timbre across explicit camera cuts without needing external editing software.
• Gemma 4 12B Vision-Language Encoder: Replaces legacy T5/CLIP text encoders to deliver high prompt adherence for complex multi-subject interactions.
• Integrated Prompt Enhancer: Automatically expands short user prompts into rich cinematic scenes, camera movements, and audio cues.
• 8-Step Distilled Inference: Offers ultra-fast generation at fixed CFG=1.0 while matching the visual quality of 50-step diffusion runs.

2. Getting Started – Loading Official Templates & HF Gated Access

ComfyUI provides official native workflow templates for LTX-2.5. However, there is one critical step you must complete first regarding model licensing.

Loading Templates in ComfyUI

3. Model Files & Directory Structure

If you prefer to download checkpoints manually or set up low-VRAM quantized versions, place files according to the following directory layout:

Component Roles & Usage Table

4. Step-by-Step Workflow Walkthroughs

Workflow A: Text-to-Video (T2V) Mode

In T2V mode, LTX-2.5 generates complete cinematic scenes with synchronized audio purely from a text prompt.

1. Load T2V Template: Open Templates -> Video -> LTX-2.5 -> Text-to-Video.
2. Set Prompt: Enter your prompt into the text box. Check Enable Prompt Enhancer to automatically expand camera instructions.
3. Configure Sampler:
   • Steps: Set to 8.
   • CFG: Keep locked at 1.0 for the distilled model.
4. Resolution & Duration: Set initial resolution to 1216 x 704 (or 16:9 equivalent). Select target clip duration (e.g. 97 frames ≈ 4 seconds at 24 FPS).
5. Run Queue: Execute prompt generation.

Workflow B: Image-to-Video (I2V) Mode

I2V mode uses an input image as the initial keyframe while generating forward motion, camera movement, and audio.

1. Load I2V Template: Open Templates -> Video -> LTX-2.5 -> Image-to-Video.
2. Upload Reference Image: Load your image into the Load Image node.
3. Write Motion Prompt: Describe what happens next in the scene rather than re-describing what is already visible in the static image:
   • Example Prompt: "The woman turns her head toward the window as a thunderstorm breaks out. Camera slowly zooms in. AUDIO: Heavy rain, thunder rumble."
4. Run Queue: Execute prompt generation.

Workflow C: First & Last Frame (FLF2V) Keyframe Interpolation

FLF2V interpolates fluid motion, camera transition, and audio between two static keyframe images.

1. Load FLF2V Template: Open Templates -> Video -> LTX-2.5 -> First & Last Frame.
2. Match Aspect Ratios: Ensure Image 1 and Image 2 share identical dimensions and aspect ratios (e.g. both 1280x720).
3. Describe Transition: Write a prompt explaining the physical movement connecting the two images:
   • Example Prompt: "A samurai pulls his sword from its sheath in a swift arc. Camera pans right. AUDIO: Metallic blade unsheathing sound, wind whistling."
4. Run Queue: Execute keyframe interpolation.

5. Prompting Best Practices & Multi-Shot Syntax

LTX-2.5 understands detailed cinematic scene descriptions and native multi-shot camera cuts.

Good vs. Weak Prompt Comparison

6. Resolution, Duration & VRAM Optimization Tips

Recommended Starting Resolutions

VRAM Tuning Guide

• 16 GB VRAM GPUs: Use the distilled ltx-2.5-transformer-distilled-int8-convrot.safetensors model. Enable CPU offloading for the Gemma 4 12B text encoder inside your loader node.
• 24 GB VRAM GPUs: Run full FP8 distilled models with the Diffusion Video Decoder (diff-vae) enabled for maximum face and micro-texture detail.

7. Common Issues & Troubleshooting

8. Conclusion & Next Steps

LTX-2.5 provides local video creators with a versatile open-weights foundation model. By understanding how to authenticate Hugging Face gated access, utilize distilled INT8 models, and structure multi-shot prompts, you can generate high-fidelity videos with synchronized audio right inside ComfyUI.

Next Steps to Explore

• Experiment with multi-cut narrative prompting (SHOT 1 + CUT TO + SHOT 2).
• Try enabling the 2x Spatial Upscaler for two-stage 1440p renders.
• Explore related guides on our ComfyHub Tutorials Directory.