Generate Your First Video

If you haven’t already, download the desktop app for your platform. No account required.

Linux

# Download the AppImage
wget https://thumper.run/download/linux/thumper-run.AppImage
chmod +x thumper-run.AppImage
./thumper-run.AppImage

macOS

Open the .dmg from the download page, drag to Applications, then right-click → Open on first launch.

Windows

Run the .exe installer from the download page. If Windows Defender flags it, click "More info" → "Run anyway".

LTX-Video is a text-to-video model that generates short video clips from natural language prompts. It runs inside ComfyUI with a dedicated workflow.

1. Open the catalog and search for LTX-Video
2. Click Install on the app card
3. Wait for ComfyUI and the LTX-Video model pack to download (~10 GB total)
4. The status changes to "Installed" when everything is ready

The download includes the LTX-Video model weights, a VAE decoder, and a text encoder. Total disk usage is approximately 10 GB.

With LTX-Video installed, let’s generate a video!

1. Click Launch on the LTX-Video card
2. Wait for the health check ("Starting..." → "Running")
3. The LTX-Video workflow loads automatically
4. Enter a prompt describing your scene, for example: "A golden retriever running through a sunlit meadow, slow motion"
5. Click Queue Prompt and wait for generation to finish

Resolution & Performance

Generation time scales with resolution and frame count. Here’s what to expect for a 5-second clip (120 frames at 24 fps):

Now that you’ve generated your first clip, let’s explore the settings.

Resolution

The default is 768×512. You can increase to 1024×576 for wider shots, but generation time increases proportionally. Stay within the model’s supported range.

Frame Count

More frames means a longer clip. To calculate duration: divide frames by fps. At 24 fps, 120 frames = 5 seconds, 240 frames = 10 seconds, and 360 frames = 15 seconds. Doubling frames roughly doubles both generation time and VRAM usage.

Seed

Like image generation, a fixed seed produces reproducible results. Set seed to a specific number to recreate the same clip, or leave it random for variety.

CFG Scale

Controls how closely the output follows your prompt. Values between 6–10 work well. Higher values can produce more vivid but sometimes less coherent motion.

Prompt Engineering Tips

Describe Motion, Not Just Appearance

Video prompts work best when you describe what is happening, not just what the scene looks like. "A cat walking across a sunlit windowsill" produces better motion than "a cat on a windowsill".

Specify Camera Movement

Include camera directions to control the perspective: "slow dolly forward", "aerial tracking shot", "static wide angle". Without camera cues, the model picks randomly.

Keep Prompts Focused

Short videos (5s) can only show one action. Avoid prompts with multiple events like "a bird lands on a branch then flies away" — pick one moment and describe it well.

Quality Checklist

Before exporting, review your clip for common issues: flickering (try a lower CFG scale), frozen frames (increase frame count), color banding (use higher resolution), and temporal incoherence (simplify the prompt to a single action). If motion looks jittery, reducing the CFG scale from 10 to 6–7 often helps smooth things out.