Recast the actor in a short video with AI: H3 Max Recast, TikTok rules

To recast an actor in a short video with AI, send h3-max-recast a source video plus one photo per person to replace (1 to 4 people, 5 to 30 seconds, 768p or 1080p). Use only photos of people who agreed, because TikTok's Shop pages bar AI personas that impersonate celebrities or public figures and ask for disclosure of AI-made avatars.

This is the model for the case where you shot a performance and want a different on-screen person, not for generating a scene from text.

What Sume documents

The Video Router docs describe h3-max-recast as replacing the people in a source video_url with the people from 1-4 reference_image_urls, one photo per person, at 768p or 1080p. The prompt is optional and the duration is the source length. The Video generation docs list the same limits: one source video, 1-4 person photos, 5-30 seconds.

Sume bills the provider list times 1.25, as for every video model. Read the live price from GET /v1/videos/models instead of copying a number from a blog.

What TikTok says about whose face you may use

The UK Shop page says not to use AI-generated avatars, voices, or personas that impersonate celebrities, political or public figures, medical professionals, or official authorities. The US Shop page lists unauthorized use of likeness, voice, name, or trademarks as a prohibited use.

Both pages describe two ways to disclose AI content: TikTok's "AI-generated content" toggle or in-content text, a watermark, a sticker, or a note in the description. A recast video is significantly AI-edited, so plan on disclosing it.

A safe way to use it

Use photos of cast members who signed off, or of your own avatar. If you want a repeatable on-screen person, Create new avatar gives you a handle you control. If the face has to be a real person's, get the consent in writing before you upload the photo.

Log the recast job id next to the post date and note that the toggle was on. For an avatar face on a clip, the older Face swap (Beta) page describes a narrower tool, with sources around 4-15 seconds that have audio.