Best AI video model for a consistent character: reference rows
For one character across clips, use a Sume row with reference images: Wan 3.0 (10), Omni (10), H3 and H3 Max (9). Kling 3 takes none. Read 2026-10-13.
To keep one character looking the same across clips, send the same reference images to a row that accepts them. On Sume those rows are wan-3.0 (up to 10 images), gemini-omni-flash-1.1 (up to 10 images), minimax-h3 and minimax-h3-max (up to 9 images each), plus the Seedance family. kling-3 takes no reference images on Sume, so for a character it can only start from a first frame. No row guarantees the same face every time, and the docs make no such claim.
What are the reference limits?
Reference images go in input_references on /v1/videos. Each row lists what it accepts, and anything else is a 400.
| Row | Reference images | Reference video | Reference audio |
|---|---|---|---|
wan-3.0 | up to 10 | up to 5 (15 s total, 16 fps or more) | up to 5 (15 s total) |
minimax-h3 | up to 9 | up to 3 (2-15 s each, 15 s combined) | up to 3 (2-15 s each, 15 s combined) |
minimax-h3-max | up to 9 | up to 3 (same limits) | up to 3 (same limits) |
gemini-omni-flash-1.1 | up to 10 | up to 3, each 3 s or less | not accepted |
kling-3 | not accepted | not accepted | not accepted |
How do I keep the look steady between clips?
Use the same reference set in every request, keep the character description in the prompt word for word, and change only the action. For H3 and H3 Max the images, videos and audios together are capped at 12, and audio cannot be the only reference. On H3 the first 5 reference images are free and each extra one adds to the charge, so a 9-image set costs more than a 5-image set.
Omni addresses references in the prompt by position: <IMAGE_REF_0> is the first image, counted from zero. A prompt that says "the woman in <IMAGE_REF_0> opens the door" ties the character to a file.
Sources
Related posts
More in Use cases
- Best AI video model for lip-synced dialogue clips on Sume
For speech that matches the mouth, Sume offers Omni (native synced audio), H3 Max Lip Sync (audio 5-14.8 s, 10 MB max), or an avatar route. Read 2026-10-13.
- Best AI video model for product videos: pick the row by input
For a product video, pick the Sume row by the photos you have: first frame, end frame or reference images. Limits for seven rows, read 2026-10-13.
- Best AI video model for seamless loops: the end-frame rows
For a looping clip, use a Sume row that takes both a first and an end frame: Seedance, Wan 3.0, Kling 3, MiniMax H3 or Omni. Rows without one, read 2026-10-13.
- Best AI video model for UGC-style ads: a model or an avatar?
For UGC-style ads, Sume offers two routes: a script-driven Avatar video of 4-60 s, or a video model with native speech such as Kling 3. Read 2026-10-13.
Written by Sume