AI Video Editing with MiniMax H3 Max

Read the first column until one row describes your clip. If you need a new shot of your character rather than a change to a shot you have, this is the wrong guide: start from Animate an AI Character. For a bigger, sharper copy of a finished cut, see AI Upscaling.

• Trim the source. Reframe, Relight and Recast process every second of it, so cut it down to the part you need first.
• One take first. Apart from Turbo Extend, none of these tools is measured yet. Run one take, compare it with the source, then decide whether to run the rest.
• Keep the source next to the result. Most of the failures you are looking for only show up side by side.

Pictures for any of these tools should be a static PNG, JPEG or WebP, 256 to 8192 pixels a side, with an aspect ratio between 0.4 and 2.5. Pad a smaller picture rather than scaling it up.

Files can be up to 50 MB for a source video, 10 MB for a picture and 15 MB for audio.

Takes a video of up to 15 seconds and returns it in another shape, with the missing edges painted in. The aspect ratio is 16:9, 9:16 (the default), 1:1, 4:3, 3:4, 21:9 or 9:21; the output is 480p or 768p (the default). An optional prompt can describe what belongs in the new areas. The timing and the sound of the source stay as they are.

Look at: the new edges. Is anything stretched or cut off there?

Takes a video of 2.33 to 15 seconds and a lighting sphere: a picture of a ball lit the way the shot should be. There is no prompt. The output is 480P, 768P (the default), 1080P or 2K, the last two refined from 768P, and the aspect ratio is adaptive by default. The timing and the sound stay.

Look at: where the light comes from, and whether it stays there for the whole shot.

Takes a video of 5 to 30 seconds and one to four photos of the new people. It keeps the motion, the camera, the cuts and the sound and swaps the main people for the photos. An optional prompt runs up to 2,000 characters, and enhance_realism asks for a more photographic look. The output is 768P or 1080P (the default).

Use your own characters or people who agreed. Never use a real person's face without their consent.

Look at: whether these are the people of the photos in every shot, and whether the cuts fall where they did in the source.

Takes a video of 1.625 to 60 seconds and a prompt for the new scene. The source plays up to start_time, the new scene runs for 5 to 13 seconds, and the source resumes from resume_time. Both times sit inside the source, the start before the resume, and the start is at least 1.625 seconds in. The source's own aspect ratio has to be between 0.4 and 2.5. Color match is on by default, and the output is 480p or 768p (the default). Reference images or videos for the new scene are not offered, so the prompt has to carry it.

Look at: the two joins, against the source.

Both take a video of 1.625 to 60 seconds with an aspect ratio between 0.4 and 2.5, a prompt and up to three audio references of 2 to 15 seconds each, 15 seconds combined. They make 1 to 15 seconds of new footage at 480P, 768P, 1080P or 2K. Start with Turbo Extend. Move to the full H3 Max Extend when the continuation is not good enough.

Turbo Extend is older than the other tools here, and we have measured it on talking clips. The full H3 Max Extend arrived on 3 November 2026 and is not measured yet.

Send output: "continuation", the default of both. It returns only the new footage after a 1.625-second repeat of the end of the source: trim that repeat off, then join. "extended" returns the whole clip re-made and softer. The comparison, and talking clips in pieces, are in AI Talking Avatar.

Takes one picture as the first frame and either a preset or a keyframed path, for 1 to 15 seconds at 480P, 768P (the default) or 1080P. The output may run about 0.7 seconds longer than you asked. For a shot of your own clip, take a frame from it and use that as the picture.

Presets: dolly_in · dolly_out · truck_left · truck_right · pedestal_up · pedestal_down · pan_left · pan_right · tilt_up · tilt_down · roll_left · roll_right · s_curve_forward.

A path is camera_trajectory: 2 to 12 keyframes of time (0 to 1 across the clip), azimuth, elevation (−90 to 90) and distance (above 0). Send one or the other: with both, the preset wins and the path is ignored.

VHS, hand-drawn pencil, 1970s painted cartoon, 16-bit pixel art and retro low poly. Each takes a prompt and an optional first frame, runs 5 to 15 whole seconds and comes out at 768P. Without a first frame, aspect_ratio sets the shape (16:9 by default); with one, the frame does. VHS also has damage_level: light, medium (the default) or heavy.

These tools make new footage. They do not repaint an existing video. To put a shot of your clip into a look, take a frame from it and give that as the first frame.

Put the take next to the source and ask the question that fits the tool. Where the source had speech, compare a transcript of the take with the source as well.

H3 Max lip-sync, H3 Max 3D-to-video and the H3 Max director are outside this guide. For lips that follow a given voice, see Animate an AI Character.