LTX-Video Model Explained: Settings, Steps and Seeds

LTX-Video, usually shortened to LTXV, is the video model behind LTX Studio and is released by Lightricks as open weights. That means you can run it yourself in ComfyUI or through hosted services, with full access to the knobs that a polished web interface hides. This guide explains what each setting does, how they trade off against each other, and where to begin so your first clips are usable rather than noisy.

What LTX-Video is

LTXV is a diffusion-based video model that generates short clips from text, from an image (image-to-video, where your still becomes the first frame) and, in some workflows, from video for extension or restyling. It is known for being fast relative to its quality, which makes it a good model for iterating on motion ideas before you spend credits elsewhere. Lightricks publishes several variants: sizes differ (a smaller, faster model and a larger, higher-quality one) and some are distilled, meaning they are tuned to produce a result in very few steps. The exact version list changes often, so check which checkpoint your workflow loads.

Because the weights are open, there is no single interface. The names below follow ComfyUI conventions; hosted playgrounds expose the same values as sliders and may label them differently.

The settings and what they do

• Prompt. LTXV responds best to long, concrete, present-tense descriptions. Lightricks’ own guidance recommends describing the scene as a sequence: what the subject looks like, what it does, how the camera moves, what the light and environment are like. Short tag-style prompts tend to produce bland or static results.
• Negative prompt. Supported by most workflows. List artefacts you want to avoid: blur, flicker, jitter, distorted faces, text, watermark.
• Resolution. Width and height of the output. LTXV expects dimensions that are multiples of 32. Lower resolutions render faster and often move more coherently; higher resolutions show more detail but can drift. Start modest and upscale afterwards.
• Frames. How many frames are generated. At the time of writing, LTXV workflows expect a frame count of the form 8n+1 (for example 25, 57, 97 or 121); if a workflow rejects your value, that is why. More frames mean a longer clip and more memory.
• Steps. How many denoising iterations the sampler runs. More steps usually mean cleaner frames up to a point, then diminishing returns. Distilled variants are built to work in very few steps; giving them many steps wastes time without improving the result.
• Guidance (CFG). How strongly the sampler follows the prompt. Low values give the model freedom and smoother motion; high values follow the text more literally but can oversaturate and strobe. Distilled variants often expect guidance near 1 and ignore or break with higher values, so read the notes shipped with the checkpoint.
• Seed. The random starting point. A fixed seed with identical settings reproduces the same clip, which is what makes controlled comparisons possible. Randomise the seed when you want variations; fix it when you are tuning another setting.
• Image conditioning. In image-to-video, the strength with which the first frame is held. Higher keeps your still intact; lower lets the model reinterpret it.

How the settings interact

Resolution, frames and steps all cost time and memory, and they compete. If you run out of memory, drop resolution first, then frames, then steps. If motion looks jittery, lower guidance before raising steps. If the clip ignores the prompt, raise guidance a little and make the prompt more specific rather than longer.

Seed is the control that makes the others legible. Fix the seed, change one setting, generate, compare. Without a fixed seed you cannot tell whether a difference came from your change or from random variation. Write down the seed with every clip you keep; it is the one value you cannot recover by looking at the output.

Settings at a glance

Hosted services charge per second or per clip and plans vary; check the current pricing page of whichever host you use.

Running it in ComfyUI

1. Install the LTX-Video nodes or use ComfyUI’s native support, and download the checkpoint and text encoder the workflow asks for.
2. Load an official example workflow rather than building from scratch. The example encodes the frame and resolution rules so you do not hit them by accident.
3. Replace only the prompt and seed on your first run. Confirm it works before touching anything else.
4. Save the workflow with your clip. ComfyUI embeds workflow data in saved outputs by default, which is the easiest way to keep prompt, seed and settings together.

If you are new to the node interface, the ComfyUI beginners guide covers the basics, and LTX Studio storyboard guide shows the same model inside Lightricks’ own tool.

Try it

Prompt: A woman in a yellow raincoat stands on a wet pier at dusk, holding a paper cup of coffee with both hands. She turns slowly to look out at the grey sea as wind lifts strands of her hair. The camera pushes in gently from a medium shot to a close-up. Overcast soft light, cool blue tones, fine rain visible against the dark water. Realistic, cinematic, shallow depth of field.
Negative: blur, flicker, jitter, distorted face, extra fingers, text, watermark.
Resolution: workflow default. Frames: 97. Steps: workflow default. Guidance: 3. Seed: 42.

Keep the seed at 42 and change only guidance (try 2 and 4) to see how adherence and motion smoothness trade off; then randomise the seed to explore variations of the version you prefer.