LTX-2 19B · Text to Video With Audio

ABOUT THE WORKFLOW

Build a Clip With Sound From Text Write a prompt describing the scene, the action, and the sound. Gemma 3 12B enriches your prompt with visual and audio detail before the model generates video and audio latents together. A spatial upscaler doubles the resolution, and a tiled VAE decodes the result. The clip comes back with synced sound at about 1920 x 1088.

Model

• LTX-2 19B Distilled by Lightricks. Released January 2026 under Apache 2.0. A 19 billion parameter DiT-based model with a Gemma 3 12B text encoder. The first open-weight model to generate synchronized audio and video in one pass. Base render at 960 x 544, doubled to about 1920 x 1088 by a spatial upscaler.

• Gemma 3 12B (QAT Q4_0). The text encoder that also serves as a prompt enhancer, enriching short descriptions with visual and audio detail before generation.

• LTX-2 Spatial Upscaler x2. Doubles the base render resolution for the final output.

HOW IT WORKS

Step 1. Write your prompt Describe the scene, the action, and the sound. Short prompts are fine: the Gemma 3 enhancer expands them with visual and audio detail. Works great with: natural scenes · interiors · ambient soundscapes · cinematic shots

Step 2. Hit run and download The model generates video and audio together, upscales the result, decodes it through a tiled VAE, and saves the clip under video/LTX-2. Ready for: Premiere · DaVinci Resolve · CapCut · After Effects

First time? Leave every setting as-is. The defaults (121 frames · 24 fps · 960 x 544 base · random seed) are the right starting point for almost everyone.

RECOMMENDED SETTINGS

Quick-start guide. Find the goal that matches yours and copy the settings.

• Standard clip (most people) — 121 frames · 24 fps · 960 x 544 base · random seed. The right starting point for almost everyone. Output lands near 1920 x 1088 after upscaling.

• Want a longer clip — Raise the frame count above 121. Frame count must be divisible by 8 plus 1 (so 121, 129, 137). Longer clips cost more time and memory.

• Want a different frame rate — Change the Frame Rate primitive. 24 fps gives a cinematic look, 25 matches broadcast, 30 is standard web video.

• Sound is not matching the scene — Describe the sound in the prompt. "Keyboard tapping, soft cafe murmur, distant traffic" goes onto the audio track. The Gemma enhancer adds audio detail if you leave it vague, but naming the sounds is more precise.

• The prompt enhancement is changing your intent — The Gemma 3 enhancer rewrites your prompt before generation. If it drifts from what you wanted, write a longer, more specific prompt so the enhancer has less room to improvise.

• Want camera motion — Two camera LoRA slots sit muted in the Models subgraph. Enable one (Ctrl+M), load a camera LoRA from the LTX-2 collection, and set strength to 1.

• Repeat a take you liked — Both seeds ship on fixed (420 and 45). Change them to get a different result. Lock them to reproduce one.

Prompt: Describe the scene first, then the sound. "A young man with headphones works on a laptop in a crowded co-working space, large windows, exposed brick, soft natural light. Keyboard tapping, low ambient chatter, distant espresso machine." The Gemma enhancer fills gaps, but naming the sounds and the camera gives a tighter result.

LEARN

📹 Videos

• Intro to Floyo

• ComfyUI 101 Free Course ft. Sebastian Kamph

• Floyo 101 for Team Collaboration

✨ Quick links

• Help docs

• LoRA Training guide

• Discover more workflows

USE CASES

🎬 Cinematic B-Roll Generate natural-looking interior and exterior footage with ambient sound from a text description.

🎵 Scenes With Ambient Sound Build clips where the audio and the picture come from the same generation, keeping them in sync without a separate sound design pass.

🎨 Concept and Previz Test how a scene looks and sounds before committing to a camera day or a full production pass.

📱 Social Content Turn a one-line idea into a short clip with matching sound for Reels, TikTok, or Shorts.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Natural scenes with ambient sound (cafes, streets, rain, interiors)

• Prompts that describe both the picture and the audio

• 5 to 8 seconds at 960 x 544 base (1080p after upscale)

• Camera LoRAs for motion control

⚠️ May produce softer results

• Readable text in the video

• Fast action sequences where physics breaks down

• Frame counts that are not divisible by 8 plus 1

• Running without the prompt enhancer on a short prompt

FAQ

What is LTX-2 19B? LTX-2 is a 19 billion parameter DiT-based model from Lightricks, released January 2026 under Apache 2.0. It is the first open-weight model to generate synchronized audio and video in one pass. It supports text-to-video, image-to-video, video-to-video, and audio-to-video, and can output up to 4K at 50 fps. This workflow runs the distilled variant in text-to-video mode with a spatial upscaler for 1080p output.

Does LTX-2 generate audio with the video? Yes. Audio and video latents are generated together in the same diffusion pass, so dialogue, effects, and ambient sound land on the same timeline as the picture. Describe the sound in your prompt and it goes onto the track.

What does the Gemma 3 prompt enhancer do? Before generation starts, the Gemma 3 12B encoder rewrites your prompt with added visual and audio detail. A short prompt like "man working in a cafe" becomes a detailed scene description with lighting, textures, camera, and sound. This is why even short prompts produce rich results, but it also means the model may interpret your words differently from your intent on specific details.

How does this compare to LTX-2.3? LTX-2.3, released March 2026, is the newer iteration with a rebuilt VAE for sharper fine details, cleaner audio, better prompt understanding, and native portrait support. This workflow uses the original LTX-2 19B distilled model. Both are Apache 2.0. Pick this one if it is already loaded and working for you, and LTX-2.3 for the latest quality.

Is LTX-2 free for commercial use? Yes. The weights and code are released under Apache 2.0, which allows commercial use, modification, fine-tuning, and self-hosted deployment with no revenue threshold.

Why is this called a heavy workflow? It loads a 19 billion parameter video model, a Gemma 3 12B text encoder, a spatial upscaler, and two VAEs. The Generate Video subgraph runs two separate sampling passes with different sigma schedules. Expect longer generation times than lighter workflows.

How to run LTX-2 text to video online? You can run LTX-2 text to video online through Floyo. No installation, no setup, no model downloads. Open the workflow in your browser, write a prompt, and hit run. Free to try.

WHY FLOYO?

Floyo is the only platform with team collaboration for ComfyUI in the browser. You run workflows with no install. You share run history, assets, and models across your team. You pay only when you generate. Floyo supports open-source and closed-source models.

A designer runs an edit and likes the result. A teammate opens that exact run from shared history and keeps going. No file handoffs. No version confusion.

For studios and enterprise teams, Floyo adds private workspaces, pooled resources, and a team usage dashboard. Other ComfyUI cloud tools run for one person at a time. Floyo runs for the whole team, with transparent per-generation costs.

Ready to try it? Write a prompt and run it. The clip comes back with synced sound.

→ Launch Workflow, Free

Questions? Watch the free course or check the FAQ above.