Getting great results
Three things drive output quality more than anything else: the input image, the prompt, and matching the settings to the job.
1. Start from a sharp, large image
Your image is the video's first frame — the model animates it, it doesn't re-imagine it. Whatever softness or noise is in the input will be in every frame of the output.
- Use an image at least as large as the output resolution: for a 720p video, a short side of 720 px or more; for 1080p, 1080 px or more.
- Small images get upscaled and render soft or grainy. The submit form warns you when this would happen — take the warning seriously, or drop to a lower output resolution.
- Pick a subject with an obvious way to move — water, hair, fabric, clouds, crowds, a face.
2. Describe the motion, not the picture
The model can already see the image. The prompt's job is to say what should happen:
- Weak: “an old man with a beard” — restates the image, gives no direction.
- Strong: “the old man breaks into a warm laugh, camera slowly pushes in” — subject motion plus camera motion.
Name one or two motions, not ten. Camera direction (“slow push-in”, “gentle pan left”, “static camera”) is the single most reliable lever. On LTX-2.3, you can also mention the soundscape — “ambient beach sounds, gentle waves”.
3. Match settings to the destination
- Aspect ratio: 9:16 for Reels/Shorts/TikTok, 16:9 for YouTube and web pages, 1:1 for feeds.
- Resolution: feeds compress heavily — 720p is usually plenty there. Save 1080p for hero placements.
- Duration: shorter clips loop better and cost less. Start at 5s and go longer only when the motion needs room.
4. Iterate cheap, finish high
Run drafts at Faster quality and a low resolution until the prompt does what you want, then re-run once at High or Highest for the final. Run again on any job re-fills the form with its settings. Since billing is per second and only on success, a draft costs pennies — see how pricing works.