LTX-2 generates videos up to 20 seconds at 4K resolution, handles audio and lip sync in the same generation pass, and runs fast. This guide covers the full setup on RunPod from nothing to finished video.
What LTX-2 does differently
Most open-source video models generate silent clips. You layer audio separately. LTX-2 produces synced audio and lip movements in the first render. Combined with smooth camera movement and high resolution output, it's one of the more capable open-source video models available right now.
Step 1: Find the RunPod template
Log into RunPod, go to Pod Templates, and search for LTX2. The template page shows everything it supports: text-to-video, image-to-video, edge-guided, depth-guided workflows. It also lists included models and prompting tips. Read those tips before generating — they explain the constraints that matter.
Step 2: Configure the pod
Select your GPU. For 4K resolution and longer videos, you need serious VRAM. The RTX 5090 works well. For HD tests, a 4090 is fine.
Click Edit on the template configuration:
- Container disk: 200 GB default is enough for the included models. Increase this if you plan to add larger models later.
- Environment variables: expand this section. The only thing you need to change is
CIVITAI_TOKEN.
Step 3: Get your CivitAI token
Go to civitai.com, click your avatar, scroll to Account Settings, find API Keys. Create a new key (name it anything). Copy it.
Back in RunPod, paste it into the CIVITAI_TOKEN field. Everything else stays default unless you have low VRAM and need to enable LIGHTWEIGHT_FP8.
Click Deploy on Demand. Wait for the container to build and the models to download. Check the logs — you'll see model downloads happening. Once you see "ComfyUI is ready," connect.
Step 4: Open ComfyUI
Click Connect, then the HTTP port for 8188. ComfyUI loads with all required nodes and models already installed. No manager setup, no missing model hunts. The template handles all of that.
Go to Workflows. You'll see four options: text-to-video, image-to-video, canny-to-video, depth-to-video. Select text-to-video.
Step 5: Understand the workflow controls
The text-to-video workflow shows a simplified front end. Key inputs:
- Prompt: describe the scene, camera movement, lighting, and action
- Frame count: controls video length. Must be divisible by 8, plus 1 (so 121 for 5 seconds, 241 for 10 seconds, 481 for 20 seconds)
- Width / Height: default is HD (1280x720). For Full HD, use 1920x1080. For 4K, use 3840x2160 with a powerful GPU
- Noise seed: randomizes the generation. Set to "Randomize" for variation
Technical constraint: both width and height must be divisible by 32, plus 1. Invalid values don't error — they round to the nearest valid resolution silently.
Generating your first video (HD)
Leave defaults at HD. Paste a descriptive prompt. LTX-2 responds well to detailed scene descriptions:
A person stands in a sunlit forest clearing, raises their hands slowly, camera pans left to follow the movement. Cinematic lighting, shallow depth of field.
Click Queue Prompt. HD generation at default frames takes roughly 2 minutes on an RTX 5090.
When done, a download link appears at the output node. The video includes audio and lip sync from the first pass — no extra steps.
Increasing to Full HD (10 seconds)
Change width to 1920, height to 1080, frame count to 241. Same prompt. Run again.
Expect roughly double the generation time (4-5 minutes). The video quality jump is visible — sharper detail, better camera movement rendering.
Extending to 20 seconds
Frame count 481. Everything else the same. Generation time scales roughly with frame count. On an RTX 5090, 20-second Full HD videos take about 10-11 minutes.
Getting your output files
If you forget to download before the workflow resets, go to Assets in the ComfyUI left panel. All generated videos are listed there.
If videos aren't appearing in Assets, open RunPod's JupyterLab for your pod. Navigate to comfyui/output/video/ in the file tree to find them directly.
Important: terminate the pod when done
RunPod charges per uptime, not per generation. A pod left running overnight costs real money. When you're done generating, go back to RunPod, find your pod, click More Actions, and click Terminate Pod.
If you need AI video generation integrated into a production workflow, get in touch. We build AI UGC ad pipelines for content and product teams.