H3 Max Video to Video: Choose by What Must Stay

H3 Max video to video is not one universal mode. The current fal catalog has several H3 Max routes that accept a video, but they make very different promises. Insert keeps the opening and ending. Extend carries the ending forward. Reframe keeps the timeline but changes the canvas. Recast keeps the performance but changes the people. Relight keeps the shot but changes its illumination. 3D-to-Video treats a blockout as structure. Reference-to-Video uses media as guidance for a new generation.

Choose the route by answering one question: what part of the source are you unwilling to lose? The answer matters more than the shared “video input” label.

The seven jobs on one page

The last row is the important exception. Reference-to-Video can accept video references, but it is not a source-preservation contract. It creates a new clip. If an exact cut, gesture, or camera path has to survive, “use this as a reference” is too loose a brief.

This guide covers seven current H3 Max workflows that accept source or reference video: Insert, Extend, Reframe, Recast, Relight, 3D-to-Video, and Reference-to-Video. fal also publishes a faster H3 Max Turbo variant of Extend. What the catalog still does not have is one generic button that safely means “edit my video.”

Start with the invariant, not the effect

Imagine a finished product film with an approved opening, a weak middle beat, and an approved pack shot at the end.

• If the bad moment is between two useful source frames, use Insert and name the interval that can be replaced.
• If the film is good but stops before the payoff, use Extend and describe only what happens next.
• If the take works but the delivery needs a different canvas, use Reframe and describe only the missing surroundings.
• If the cut and acting already work but the approved presenter is missing, use Recast with a consented portrait.
• If the watch movement looks right but the cool, flat light does not, use Relight with a clear lighting-sphere image.
• If the “source” is really a Blender greybox that defines a drone move, use 3D-to-Video.
• If a reference clip only contributes the rhythm of the camera while the product, location, and action should be new, use Reference-to-Video.

Those are seven different production decisions. A longer prompt cannot rescue the wrong one. Asking Reference-to-Video to “keep every cut exactly” does not turn a reference into a locked timeline. Asking Recast for a new performance wastes the very source performance that route is designed to retain.

Insert and Extend disagree about the ending

Insert returns to the source. Extend leaves it behind.

The current minimax/h3-max/insert-video contract requires a source video plus Start and Resume times. H3 Max generates a five-to-thirteen-second replacement, keeps the source before Start, then hands back to the source at Resume. This is the right shape for a repair in the middle of a clip. The current documentation does not promise how the source soundtrack crosses that generated interval, so audio continuity belongs in the review plan rather than the marketing claim.

The minimax/h3-max/extend-video route and its faster minimax/h3-max-turbo/extend-video variant instead read the source ending and add 0.71 to fifteen seconds. Their combined extended output returns the original and continuation together. The separate continuation output begins with a 1.625-second repeat of the source ending, giving an editor material for the seam. That repeated handle is not extra story time; align it with the original ending and keep one copy.

A simple test settles the choice:

Must the final video return to an existing frame later in the source?
Yes -> Insert
No, it should continue past the last frame -> Extend

Recast and Reference disagree about the performance

Recast is for a performance worth keeping. Reference-to-Video is for a performance worth borrowing from.

The current Recast schema accepts a five-to-thirty-second source and one to four replacement-person photos. It is designed to retain the source motion, camera work, cuts, and audio while changing who appears. “Designed to retain” is not a frame-by-frame identity guarantee: profiles, occlusion, motion blur, crossings, and fast cuts still need review.

Reference-to-Video does not return the source performance with a different surface. It can use image, video, and audio references to condition a new shot. That is useful when a clip's camera rhythm should inspire a different scene, or when several appearance references need to meet in one new composition. It is the wrong choice when an approved actor's timing or an exact edit must survive.

Use this distinction in the brief:

Recast: Keep this exact performance and replace the main person with Image 1.
Reference: Use Video 1 only for the handheld rhythm; create a new scene with Image 1.

The words are similar. The ownership of the timeline is not.

Relight and 3D-to-Video disagree about what the source is

Relight treats the source as finished photography. 3D-to-Video treats it as a plan.

The Relight endpoint requires a source clip and a lighting-sphere image. It has no prompt field. The sphere communicates direction, color, softness, and contrast; the route is designed to preserve the subjects, motion, camera, and audio while changing illumination. Use it after the shot works. It cannot repair blocking, framing, or a weak performance.

The 3D-to-Video endpoint accepts a short Blender or previs render as structural guidance. The blockout carries layout, camera movement, object motion, timing, and object count into a more finished generation. Optional images guide appearance. It is not expected to preserve the grey geometry pixel for pixel, so it fits a source built to be replaced rather than photographed footage built to be kept.

If the video is the deliverable, refine it with Relight. If the video is the blueprint, translate it with 3D-to-Video.

Check the contract before preparing the upload

The route names hide different input and output limits. As of October 21, 2026, the official schemas describe these boundaries:

These are upstream API capabilities, not a promise that every website labelled “H3 Max” exposes them. The live Reference-to-Video schema now allows fractional durations from 0.92 to 15 seconds, while older fal guides still describe five to fifteen seconds. This site's product deliberately keeps whole five-to-fifteen-second requests.

Resolution is not a universal H3 Max property either. Insert stops at 768p, while some finishing routes expose 1080P or 2K. Do not buy a plan for a resolution simply because another route from the same model family lists it.

Write a one-line preservation brief before the prompt

Before uploading anything, complete this sentence:

The result must keep __ from the source; it may change __.

Then add four operational lines:

Timeline rule: return at 12.4s / continue after the ending / keep every cut.
Audio rule: preserve, replace later, or treat only as reference.
Allowed change: one scene, one person, lighting, rendered appearance, or everything.
Acceptance test: the return seam is invisible / the bicycle keeps moving left to right / every cut retains the mapped presenter.

That short document is more useful than a page of mood adjectives. It tells you which endpoint to call and gives the result a pass/fail test. Keep the untouched original, trim irrelevant footage before upload, and review the complete output at normal speed. For footage of real people, confirm consent and disclosure requirements before a Recast or reference-led generation; a model endpoint does not grant likeness, voice, music, or source-video rights.

AI Live Generator is not a general source-video editor. The main generator creates new H3 Max Turbo clips from text or still images at 480P or 768P. Separate themed tools expose selected still-image Reference-to-Video workflows. This site does not accept a source video, lighting sphere, 3D blockout, motion-reference video, or reference audio, and it does not call Insert, Extend, Recast, Relight, or 3D-to-Video.

“Video to video” tells you the file type going in. It does not tell you what the model is allowed to forget. Choose that first, and the correct H3 Max workflow becomes much easier to see.