Reference images
Build first frames, character sheets, location sheets and product shots so characters and products stay the same from shot to shot.
If anything has to look the same in two shots (a face, an outfit, a room, a product), give the model an image of it. Text alone is not enough. Image to video beats text to video for hero shots, characters and products. Keep text to video for exploring ideas and for wide shots where nothing has to match.
Make all your images before any video. A bad still costs one image. A bad clip costs a full video render.
What to build
MindStudio's short film guide budgets roughly this many images before the first video:
| Image | How many | Used for |
|---|---|---|
| Concept images | 8 to 12 | Finding the look and the style line |
| Character references | 4 to 6 per character | Identity in every shot |
| Location references | 3 to 4 | Same room, same light |
| Shot frames | 3 to 5 variations per shot | First frames for image to video |
For a beginner's 30 s video with one character and one place, one character sheet, one location sheet and two or three first frames are usually enough.
Character sheets
A character sheet (or turnaround) shows the same character from several angles on one image. Attach it to every shot where the character appears.
- Four views: front full body, back full body, front close up, 90 degree profile close up.
- Neutral grey background, even studio light, no text, no other figures.
- Same scale and head height in every view, a relaxed A pose.
- Write 5 to 7 fixed traits and reuse the exact words in every prompt.
- Add a visual anchor that is easy to spot: a coloured hat, an earring, a jacket.
Character model sheet, 16:9, neutral mid grey seamless background, even soft
studio light from front left, no text, no watermark, no other figures.
Four views of the same character at identical scale and head height:
(1) front full body, relaxed A pose, (2) back full body, (3) front head and
shoulders close up, (4) 90 degree left profile close up.
Character: RIO, early 30s, lean build, short black curly hair, thin silver
hoop in left ear, rust orange work jacket over grey hoodie, dark olive cargo
pants, white high top sneakers, calm half smile.
Style: stylized 3D animated film look, soft subsurface skin, clean shapes.Check the sheet closely before you use it. A wrong shoe colour on the sheet repeats in every shot. Settle the design first: two versions of the same costume in your references give the model an unresolved design.
Location sheets
Do the same for places. One image per angle, with the same architecture, set dressing and light direction. Creators on X report that a character sheet plus a location sheet is what makes most shots work on the first try.
Same location, 3 images, identical architecture and set dressing:
(A) wide establishing shot at golden hour, (B) reverse angle medium shot,
(C) detail insert of the counter. Palette: warm wood, brass, deep green.
Lighting direction constant: low sun from camera left.Product shots
Always attach a photo of the real product. A logo or label described in words comes out as a made up version. Leave empty space for titles and add all text in the edit.
Packshot of the attached bottle exactly as shown (shape, cap, label layout),
centered on wet black slate, single hard rim light from behind right, soft
fill from the front, shallow depth of field, droplets on the glass, 9:16,
empty space in the top third for a title added in post. No added text.First frames and last frames
A first frame is the exact image your clip starts from. Make it from the character sheet, in the real location.
Using the attached sheet as the exact identity for RIO (face from view 3,
outfit from view 1), render a 16:9 keyframe: medium shot, 50mm, RIO leaning
on a rain wet railing at a night ferry terminal, sodium orange and teal
lighting, light drizzle, background ferry lights soft. Same style as the sheet.
The background must be the terminal, not grey.The last line matters. Close ups made from a sheet can inherit its grey backdrop, so restate the background.
Some models (Veo 3.1, Kling 3.0, Seedance) also take a last frame. Change as little as possible between the two frames. The more they differ, the more the model has to invent.
Frame A: kitchen table by a window, early morning cool light, a mug alone on the table.
Frame B: identical framing and objects, plus visible steam from the mug and a
warmer, brighter patch of sun on the table. Nothing else changes.Creators report one exception: with Seedance, a first frame does not always give the best motion. Giving references and letting the model compose the shot can work better. Try both on a cheap draft.
Style locking
- Write one style line and paste it word for word into every image and video prompt. Do not paraphrase it.
- Fix the aspect ratio, the model and the style line at the start of the project.
- If you cannot name a look, take a screenshot of it and ask an LLM to describe it in technical terms. Use that as your style line.
- A reference image's colours can override a lighting instruction written in the prompt (one tester's finding). If the light comes out wrong, fix the reference, not only the text.
- Stylized characters (3D animation, illustration, clay) hold identity better and pass content filters more often than photoreal ones.
Keeping identity across shots
Attach the same sheet every time
Every shot with the character gets the same character sheet, even if the previous shot already showed them.
Repeat the same words
Same name, same 5 to 7 traits, same style line, in every prompt. See Prompting.
Give each reference one job
Say what each image is for: "@Image 1 is the identity of RIO only, @Image 2 is the location and lighting only". This stops the model from copying the wrong thing.
Keep shots short and simple
Faces drift most in long clips, fast turns and when someone passes in front of the face. Use 3 to 8 s shots and few simultaneous actions.
On Seedance, one guide reports that a reference strength of 70 to 80% works best: 90 to 100% makes characters stiff, below 60% lets them drift. This is a single source.
Model limits
Limits checked on 15 Oct 2026. They change often, so check the model page before you build a big reference set.
| Model | Reference inputs | Notes |
|---|---|---|
| Seedance 2.5 | Up to 50 per generation: 30 images, 10 videos, 10 audio | Blocks reference images or videos with real human faces. The official guide recommends 1 to 8 subject images |
| Seedance 2.0 | Up to 15: 9 images, 3 videos, 3 audio | Same real face rule reported; also rejected a reference with blood in one test |
| Kling 3.0 | Up to 7 images or subjects, or 1 video (up to 15 s) plus 4 | Elements: build a character from 2 to 4 images or a video, with an optional voice |
| Veo 3.1 | Up to 3 reference images (person, character, product) | Only on 8 s clips, not on Veo 3.1 Lite; first and last frame supported |
| Gemini Omni | Image references (count not stated), up to 3 video reference clips of 3 s | Accepts a storyboard grid as input |
| Wan 3.0 | Up to 10 images, 5 videos, 5 audio, 20 references in total | First or last frame cannot be mixed with references |
| MiniMax H3 | Up to 9 images, 3 videos, 3 audio, 12 files mixed | Refer to them in plain text ("reference images 1 and 2") |
| Vidu Q4 | 1 to 15 images, 0 to 3 audio clips | No text to video at all, references are required |
| Grok Imagine 1.5 | Up to 14 images and 3 voice references | Reference to video renders up to 720p |
| Nano Banana Pro (image) | 6 to 14 inputs, depending on where you use it | Keeps several characters consistent across images |
Model guides: Seedance, Kling, Google, Wan, Hailuo and MiniMax, Vidu, Others.
Real faces and moderation
Seedance 2.5 does not accept references that show real human faces. Use a generated or stylized character instead. Moderation can also reject a reference it accepted before, so keep a stylized backup of every key character.
Sources
- Image to video for matching shots: invideo FAQ, Kompozy
- Image budget, style line: MindStudio
- "Spend your retries on stills": fal on X
- Character sheet layout, refusals of realistic portraits, describing a look: Kapwing, SelfieLab
- Wrong details repeating, grey backdrop: JXP
- Character and location sheets: OlatundeAI on X
- First and last frames: dev.to Veemo, Google Cloud Veo 3.1 guide
- Seedance first frames: vfxwaffu on X
- Reference colours override text: roo
- Reference strength, logos in text: heyuan110
- Limits: BytePlus Seedance 2.5, Seedance 2.5 prompt guide, Kling 3.0 guide, Kling blog, Veo docs, Omni docs, Wan 3.0 guide, MiniMax docs, Vidu, xAI video guide, Google Nano Banana Pro tips
- Blood rejection: Liromfilms