Prompting for Better Video Results

English, not tags

Image sprites use comma-separated tags like 1girl, silver hair, blue eyes. Video does not.

The video model reads present-tense prose: who moves, how they move, and what the camera does.

Do not re-describe hair, eyes, clothes, or body. Those details are already in the image you picked.

For character and outfit tags on stills, use the image prompting guide.

What to type

One beat is enough: the action plus the camera. Example:

Her hair blows in the wind while she breathes slowly. The camera holds a static shot.

You pick one still. Play once is the default: the clip starts on that image and the motion continues forward. Choose Loop under the prompt when you want that same image as the end frame too. We wrap your sentence into a timeline with no extra audio. You do not write those fields yourself.

Clip length is 4 to 12 seconds. The default is 5. The model snaps that length onto its frame grid, so the finished clip is a fraction of a second off the number you pick.

Keep it to one shot. Extra scenes, cuts, or a second camera setup in the same prompt usually fight the short clip.

Timeline

The video model follows time, not a tag list. A short clip is one shot that begins on your still. Play once continues forward. Loop returns to that same still.

Play once sends only Picture 1. Loop sends that same image again as Picture 2 at the end. You never pick a second frame.

You type:

Her hair blows in the wind while she breathes slowly. The camera holds a static shot.

With Play once selected, the model receives something like this (clip length here is 5 seconds):

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video. There is no end-frame picture.

integrated_multimodal_description: [Shot 1] 2D-animated, the subjects begin in the pose and framing of Picture 1. Her hair blows in the wind while she breathes slowly. The camera holds a static shot. The motion continues forward and does not return to the starting pose.

overall_soundscape: N/A

non_diegetic_music: N/A

integrated_multimodal_description is the motion. The two N/A lines turn audio off. Realism clips use live-action, cinematic instead of 2D-animated.

Choose Loop when the beat should repeat. You type:

She rides in a slow, repeating rhythm and returns to the starting pose. The camera holds a static shot.

Wrapped:

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 5.00-second mark of the target video.

integrated_multimodal_description: [Shot 1] 2D-animated, the subjects begin in the pose and framing of Picture 1. She rides in a slow, repeating rhythm and returns to the starting pose. The camera holds a static shot. The action is cyclic and returns to the starting pose so the clip loops seamlessly, ending in the pose and framing of Picture 2.

overall_soundscape: N/A

non_diegetic_music: N/A

Recipe: action + camera

Use verbs, not tag lists. Name the subject only if it helps: she, he, they, or the camera.

Camera language that works well:

• The camera holds a static shot
• The camera slowly zooms in
• The camera slowly zooms out
• The camera pans gently to the left
• The camera pans gently to the right
• The camera tilts slowly upward
• The camera shakes slightly

Prefer slow, continuous motion. Fast whip-pans and stacked camera tricks usually look worse on a short clip.

Do not use Hailuo-style brackets such as [Zoom in] or [Pan left]. Write the camera in a normal sentence.

Loops

Play once is selected by default. The clip starts on your still and does not have to end there. Choose Loop when you want a seamless repeat: we send the same still as the end frame, and the motion should return to that pose.

Breathing, blinking, a small sway, hair in the wind, or a repeating hip motion all loop cleanly when Loop is selected. Walking across a room fits Play once. It does not fit Loop: that clip has to end on the same pose as the still you picked.

When Loop is selected, do not describe a one-way change such as standing up, sitting down, or leaving the frame. The start and the end are the same image.

Do and don't

Do:

• Stay in short, present-tense English.
• Match the pose already in the chosen image.
• Pick one action and one camera move.
• Use the Motion and Action chips on the create page when you want a starting phrase.

Don't:

• Paste a Danbooru tag soup.
• Re-list hair color, outfit, or body tags unless something actually changes on camera.
• Write scene cuts, shot numbers, or a second location.
• Ask for music, voiceover, or sound effects. Audio is left off for these clips.

Examples

Weak prompts fight the image. Better prompts describe only the beat you want to see.

1girl, silver hair, school uniform, looking at viewer, wind, cinematic, masterpiece

She blinks and smiles. Her hair blows in the wind. The camera holds a static shot.

She walks across the room, changes outfit, then the camera cuts to a close-up in another location.

She sways her hips slowly and breathes. The camera slowly zooms in.

cowgirl, 1boy, motion lines, ahegao, looping gif

She rides in a slow, repeating rhythm and returns to the starting pose. The camera holds a static shot.

Advanced

Most people should stay with the short sentence. If you paste a block that already starts with integrated_multimodal_description, we send that through as-is and only fill missing audio lines.

When Loop is selected, a custom timeline still has to return to the same still. Do not write a second location, a costume change, or a last pose that the image does not show.

Optional video LoRAs are a separate field, not part of the text prompt. Paste a Civitai version ID, then an optional strength (default 1.0).

On Civitai, open the LoRA and copy the version ID from the URL: civitai.com/models/123456?modelVersionId=789012. Use 789012. Do not use the model ID (123456) and do not use a file ID.