Generating videos

Want a moving picture? Ask for one. In an agentic chat, describing the clip you want is enough — Catalyst renders a short video and drops it straight into the reply, with an inline player: play, pause, and full-screen, right there in the conversation.

There are two ways to get a clip:

• Describe it — “make a video of waves rolling onto a beach at sunset” — and Catalyst generates it from scratch (text-to-video).
• Animate a still — point at an image you uploaded, or one Catalyst just generated, and say “animate this” (image-to-video).

Note

Video generation runs as a tool, so it only happens in agentic mode. If you ask for a video and just get a written description back, check that agentic mode is on.

From a description

Write what you want to see happen — the subject, the motion, and how the camera moves:

Make a video of a paper boat drifting down a rain-soaked street, camera following slowly from behind.

The assistant tells you it’s starting the render and roughly how long it will take, then the finished clip appears inline in its reply. Text-to-video runs on LTX-Video, the fast default (more on the models below).

Tip

Lead with motion. A prompt that reads like a photo caption comes back looking like a frozen frame. Say what moves — “drifts,” “sways,” “ripples” — and how the camera behaves. See Prompting tips.

Animating an image

The image-to-video flow pairs naturally with image generation: make the still first, then set it in motion.

1. Get the still into the chat — attach an image to your message, or ask Catalyst to generate one (“generate an image of a lighthouse on a cliff at dusk”).

2. Ask for motion, describing what should move:

   Animate this — waves crashing below, the lighthouse beam sweeping across the frame.

3. Catalyst animates it and returns the clip inline, right below your still.

This two-step pattern — generate a still, then animate it — is also how you make a moving version of something that doesn’t exist yet. Want an animated logo of a fox? Ask for the image first, get it looking right, then say “animate it.” Catalyst handles both steps for you in the same conversation.

Give it a voice

This is what sets Catalyst’s video apart. The default model, LTX-Video, generates a synchronized audio track — including spoken dialogue. Put the actual lines in your prompt, in quotes, and the characters speak them, roughly lip-synced.

Close-up of a barista sliding a coffee across the counter. She looks up, smiles, and says, “Careful — it’s hot.”

The catch is timing: spoken audio runs at roughly two to three words per second, so the dialogue has to fit inside the clip’s length. Keep it short.

• A 5-second clip (the default) fits about one short line or a quick exchange.
• A 9-second clip fits about two or three short lines.

If a line is too long to land inside the clip, trim it rather than let it overflow. And note that audio is LTX-only — the WAN model below is silent, so for anything with talking, narration, or sound, stay on the default.

Tip

Attribute each line to who says it (“she says…,” “the robot replies…”) and keep the wording tight. Short, punchy lines lip-sync far better than a paragraph crammed into five seconds.

Choosing a model

You don’t normally pick a model — the assistant chooses the right default for what you asked. Workspace admins curate which video models are available under Admin → Video Models. Two models cover the range:

LTX-Video — fast, with audio (default)

The default for both text-to-video and image-to-video. Fast, and the only model that produces synchronized audio and spoken dialogue. Use it for almost everything — and always when there’s talking or sound.

WAN 2.2 — higher quality, silent

A higher-quality, image-to-video only alternative. Noticeably slower (around three minutes) and silent — no audio. Reach for it when you want the best-looking animation of a still and don’t need sound: just say “use WAN” or “higher quality.”

Note

WAN only animates an existing image — it can’t generate a clip from a text description, and it never carries audio. Text-to-video and anything with dialogue always use LTX-Video.

Length and how long it takes

Video is far heavier than images, and Catalyst renders it synchronously while you wait.

• Clips are about 5 seconds by default. You can ask for longer — LTX handles up to around nine seconds well. You can also ask for an orientation in plain language (“make it vertical” / “9:16”) and the assistant sets it up.
• Expect roughly one to three minutes per clip (WAN is at the slower end). The assistant says so before it starts, so a quiet pause mid-reply is normal — the clip appears as soon as the render finishes.
• One clip renders at a time. If the video engine is busy with another render, the assistant tells you instead of silently queueing — just ask again in a minute or two.

Prompting tips

Video prompts reward different habits than image prompts:

• Describe motion, not just a scene. “Waves rolling,” “leaves drifting down,” “steam rising from the cup” — a video prompt without a verb comes out looking like a still.
• Use camera language. “Slow push-in on her face,” “camera pans across the skyline,” “gentle orbit around the statue” — models understand cinematography words surprisingly well.
• One moment per clip. A few seconds of one continuous motion beats a compressed mini-movie. Want a sequence? Generate several clips.
• Quote your dialogue, and keep it short. Put spoken lines in quotes, attribute them, and fit them to the clip length (about two to three words per second — see Give it a voice).
• For image-to-video, nail the still first. Iterate on the image until the composition is right, then animate — motion won’t fix a frame you don’t like.

Working with the result

A generated clip renders as a native video player inline in the reply — with standard controls — and it travels with the conversation. Expand it full-screen from the lightbox, and anyone viewing a shared chat can play it too.

Under the hood, plain language is all you need: a video-author skill loads automatically and teaches the assistant how to pick the model, budget dialogue to the clip length, and phrase prompts for motion — so you can just ask for the clip you want.