Keep the same character across clips: the reference sheet prompt
A character description block you reuse verbatim in every prompt, plus a generated reference sheet, so a face, outfit and build stay the same from shot to shot.
CHARACTER BLOCK (paste unchanged in every prompt): {Name}: {age range}, {ethnicity / skin tone}, {hair: colour, length, style}, {face: 2 distinctive features, e.g. 'strong jaw, small scar on left eyebrow'}, {build}, wearing {exact outfit: colours, fabric, one accessory}. Always the same outfit and hairstyle. REFERENCE SHEET (image model, once): Character reference sheet of {Name} — {paste CHARACTER BLOCK} — three views: front, three-quarter, profile; neutral expression; plain grey background; even studio light; full body and a close-up of the face; no text. SHOT (video model, each clip): {paste CHARACTER BLOCK}. {action}. {setting}. {camera}. {light}. {style}. Use the attached reference image for the face and outfit.
Paste into the video model as one paragraph; remove the labels once you are happy with the content.
Steps
- Write the character block once and never paraphrase it. Models key on exact repeated phrasing.
- Generate the reference sheet with your image model, pick the cleanest, and attach it as a reference / first frame where the video model supports it.
- Change only the action, setting and camera between clips. If you change the outfit, you have a new character.
Tips
- Two distinctive features beat ten generic ones: 'red beanie, round glasses' survives generation; 'kind eyes, warm smile' does not.
- Side-by-side the first frame of every clip in a contact sheet before editing; drift is obvious at thumbnail size and invisible at full screen.
| works with | Image models: Midjourney, Flux, Imagen: keyframes and style boards for image-to-video. · Veo: Google text-to-video / image-to-video prompts. · Kling: Kling text/image-to-video prompts. · Runway: Runway Gen prompts and camera controls. · Any LLM: Model-agnostic: Gemini, Mistral, a local model. |
| cost | Free. The video or image model you paste it into has its own pricing. |
| status | Unverified: written from the technique, not yet run end to end by us. |
| license | CC BY 4.0 for the text here; skills are MIT. |
| updated | 2026-11-01 |
| for agents | character-consistency-reference-sheet.md |
Related
Camera movement vocabulary that video models understand
The 14 camera moves that generation models respond to reliably, with the phrasing that works and the phrasing that produces a wobbling mess.
Image-to-video: generate the first frame, then animate it
Make a still with an image model first, fix the composition there, and only then ask the video model to animate it. Cheaper iterations and the look you chose, not the one the video model guessed.
Product hero shot: a 5-second generated clip for an ad or a listing
Prompt for a clean product shot with one motion and one light change, the kind that sits at the top of a listing or opens a 15-second ad.
Style lock: one paragraph that makes every clip look like the same film
A reusable style paragraph (stock, palette, grain, light rules) appended to every generation prompt so a series cuts together instead of looking like ten different ads.
The 7-part prompt structure for text-to-video
A fill-in template that gives any video model the seven things it needs: subject, action, setting, camera, light, style and sound. Fewer vague words, fewer wasted generations.
Anime style: 2D cel look, limited animation, clean line weight
A filled 7-part prompt for a hand-drawn anime shot: flat cel shading, animation on twos, a held painted background, and the three things the style breaks in current video models.