Consistent Characters in AI Video: Four Models Tested
Updated · Live figures last changed · 4 min read · The GenMagic team
We gave four video models the same character sheet and the same scene. All four kept the character; the differences were in fidelity, style and price.
Every model in this test kept the character from a single reference image, through a five-second shot. The red curls, the round green glasses, the freckles, the yellow raincoat over a striped shirt: all four video models built to follow reference images held them from the first frame to the last. What separated them was how closely they matched the sheet, how much they restyled the face, how long they took and what they cost.
Keeping a character the same from shot to shot is the hardest everyday problem in AI video, and the reason studios and agencies choose one model over another. Our guide to consistent characters covers the techniques; this is a direct test of the models on October 29, 2026.
The test
- Reference: one character sheet, an original animated inventor girl, made with Nano Banana 2.1 for $0.05 in the test.
- Scene: The girl from the reference image walks into a cozy wooden workshop, picks up a small brass gear from the workbench and holds it up to the camera with a big smile. Animated film style, warm light.
- Settings: 5 seconds, 720p, 16:9, no sound, the sheet passed as a reference image (not as the first frame), through GenMagic's API.
We used an illustrated character on purpose: ByteDance's Seedance models refuse reference and opening images of realistic people, and an animated character lets all four compete on the same input. For photoreal presenters, see our UGC ad model test.
The results
| Model | Cost in the test (Oct 29) | Price today | Render time | Character | Notes |
|---|---|---|---|---|---|
| Seedance 2.0 Mini | $0.50 | $0.57 | 109 s | Closest to the sheet | Most cinematic camera, with a close-up of her hands on the gear |
| Wan 3.0 | $0.55 | $0.64 | 130 s | Faithful | Smooth, simple staging; also makes clips up to 30 s with sound |
| Grok Imagine Video | $0.46 | $0.53 | 59 s | Faithful, less detail | Fastest and cheapest; darker light, softer textures |
| HappyHorse 1.1 | $0.64 | $0.74 | 106 s | Recognisable, restyled | Most expressive; the face drifts toward a more exaggerated cartoon style |
The test's costs are what each clip was billed on GenMagic on October 29. Price today is the same 5-second clip at 720p without sound at today's prices, and changes when they do.
All four understood the action: she walks in, picks up the gear and holds it to the camera, smiling. None swapped her glasses, hair or coat colour, which older models routinely did within a few seconds.
The clips
Seedance 2.0 Mini: $0.50 in the test
Wan 3.0: $0.55 in the test
Grok Imagine Video: $0.46 in the test
HappyHorse 1.1: $0.64 in the test
Which model to use
- The closest match to your character sheet: Seedance 2.0 Mini, for animated and illustrated characters. It refuses realistic people.
- Long scenes and dialogue: Wan 3.0, which keeps the character as faithfully and makes up to 30 seconds in one shot with sound, so one voice carries a whole scene.
- Fast drafts and storyboards: Grok Imagine Video, the quickest and cheapest here, to try ideas before rendering the final shots.
- Expressive, stylised animation: HappyHorse 1.1, if a bolder cartoon look suits the project better than strict fidelity to the sheet. Its prices and routes: the HappyHorse API.
These were single takes of one scene. Over a sequence, every model drifts a little more with each new clip, so the habits in our consistent characters guide still matter: the same reference images on every shot, the character described the same way in every prompt, and the next shot started from the last frame of the one before (GenMagic's Continue does this while keeping the references).
What a sequence costs
At today's prices, a one-minute sequence of twelve five-second shots at 720p costs $6.30 with Grok Imagine Video, $6.80 with Seedance 2.0 Mini, $7.65 with Wan 3.0 and $8.89 with HappyHorse 1.1, before retakes. Plan on two or three takes for the shots that matter. Current prices for every model are on video API pricing, and studios producing for clients can see what a month costs on our agency page.
Run the same test yourself
curl https://genmagic.co/api/v1/videos \
-H "Authorization: Bearer $GENMAGIC_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "bytedance/seedance-2.0-mini",
"prompt": "The girl from the reference image walks into a cozy wooden workshop, picks up a small brass gear from the workbench and holds it up to the camera with a big smile. Animated film style, warm light.",
"references": ["https://example.com/your-character-sheet.png"],
"aspect_ratio": "16:9",
"resolution": "720p",
"duration": 5,
"generate_audio": false
}'
Up to nine reference images go in references; a Google Drive or Dropbox share link to the sheet works too. In the studio, attach the sheet as a reference on a video prompt.
Questions
Which AI video model keeps a character most consistent? In this test, Seedance 2.0 Mini matched the character sheet most closely, with Wan 3.0 close behind. All four models that follow reference images kept the character through the shot.
Can I use a real person as the reference? Some models refuse it: Seedance turns down references and opening images that may show a real person. Wan 3.0 and the Veo models accept a realistic presenter. Never generate a real, identifiable person without their written permission.
How do I keep the same voice across clips? Make one longer clip per dialogue (Wan 3.0 runs to 30 seconds with sound), or record the voice once with a speech model and lay it over silent clips in your editor.