H3 Max Reframe Video: Change Aspect Ratio with AI Outpainting

H3 Max Reframe changes the canvas around an existing video without simply cropping the shot. It contain-fits the source inside a new aspect ratio, generates the empty areas around it, and keeps the source timeline and audio. That makes it useful when a finished horizontal clip needs a vertical or square delivery but the subject cannot be cut out of frame.

It is not the same as selecting 9:16 before generating a new clip. Reframe starts with a moving source video and builds the missing surroundings. The original picture is still processed, so the upstream documentation does not promise pixel-for-pixel preservation.

AI Live Generator does not currently accept a source-video upload or call the Reframe endpoint. The site's generator creates new H3 Max Turbo clips from text or still images. This guide explains the separate minimax/h3-max/reframe-video workflow and helps you decide whether it is the right edit.

Reframe, crop, stretch, and regenerate solve different problems

Suppose a cyclist crosses the whole width of a 16:9 shot. A center crop can make a 9:16 file immediately, but the rider may leave the crop. Stretching keeps every source pixel but distorts the scene. Regenerating from a prompt abandons the approved take. Reframe keeps the take on the canvas and asks the model to create believable space above, below, or beside it.

Use a normal crop first. It is deterministic, cheap, and often enough. Reframe earns its cost when cropping removes information the audience needs: a second person, a hand entering frame, a product at the edge, a camera move, or the environment that explains the action.

What the current endpoint keeps—and what it may invent

The official contract accepts one video_url from 0.25 to 300 seconds. It preserves the source timeline and audio while fitting that video onto a target canvas. The generated areas can change from frame to frame, and even the source region is not guaranteed to remain bit-identical after processing.

That leaves three practical invariants:

• Time stays owned by the source. Reframe is not an extension and does not add a new ending.
• The soundtrack stays attached. Dialogue, music, and effects keep their place on the existing timeline.
• The canvas may be newly interpreted. The model must infer scenery that the camera never captured.

If an exact logo, subtitle, legal line, or UI element must not move, add it after the reframe in an editor. Generative outpainting is a poor place to lock small text or brand geometry.

The route currently offers seven target shapes: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, and 9:21. Its result also reports placement and frame count, while the schema defaults to 24 fps. fal does not currently document the meaning of the placement map's keys, so treat that field as integration metadata rather than a public editing control.

Choose the target canvas before writing a prompt

The aspect ratio determines how much new image the model must invent. Converting 16:9 to 9:16 is not a small format change: the original wide shot becomes a band inside a tall canvas, leaving a large area to synthesize. Converting 4:3 to 1:1 asks for much less.

Start with the delivery slot, not a vague preference:

• use 9:16 for a standard vertical social placement;
• use 1:1 when the destination is square and the shot already has a centered subject;
• use 21:9 only when the extra width carries useful scene information rather than decorative emptiness;
• use 9:21 when the actual destination needs that taller format, not merely because it is available.

Then inspect the full source at normal speed. Note every moment when a person, prop, shadow, reflection, or camera move touches an edge. Those moments define the hard parts of the reframe. A clean first frame does not prove the tenth second will work.

Prompt the missing environment, not a second story

The prompt is optional. The route already captions the source as part of its expansion process, so the prompt should resolve ambiguity around the newly exposed space. It should not rewrite the action that is already on the source timeline.

For an untested 16:9-to-9:16 street shot, a restrained direction might be:

Continue the same rain-soaked street above and below the source frame.
Keep the storefront architecture, wet asphalt reflections, night lighting,
and perspective consistent. No new people, signs, vehicles, or camera move.

That prompt does four useful things: it names the space to continue, repeats the existing light and perspective, forbids new narrative objects, and avoids asking for a new shot. It is a request plan, not a claimed result.

Weak prompts often describe mood without defining continuity: “make it cinematic,” “add an epic city,” or “viral vertical video.” Those words give the model permission to decorate the border rather than finish the scene. When the source already tells the story, less prompt is usually the sharper brief.

Review edges over time, not only one frame

Outpainting can look convincing in a thumbnail and fail in motion. Review the result in four passes:

1. Mute the video. Watch each generated edge for crawling architecture, changing horizon lines, appearing objects, or texture that sticks to the frame instead of the camera move.
2. Track crossings. Pay special attention when hair, hands, wheels, shadows, reflections, or props move between the source region and the generated area.
3. Listen once without looking. Confirm the returned soundtrack begins and ends where the source does and that no required audio is missing.
4. Overlay safe areas. Check whether captions, platform controls, or a planned crop will cover the very subject you reframed to save.

One bad boundary does not always require a new generation. If the defect is brief and outside the safe area, a tighter crop or graphic layer may be the more reliable finishing move. If the scene repeatedly crosses the boundary, change the target ratio, shorten the source, or choose a source with more breathing room.

The 1080p field and the published price do not line up yet

As of October 21, 2026, fal's machine-readable route description publishes $0.05 per generated second at 480p and $0.08 at 768p. Its example prices a five-second 768p result at $0.40.

The live schema also lists 1080p and describes it as a learned latent upscale from a native 768p reframe. However, the route description and price block only document 480p and 768p. Until a 1080p rate is published consistently, do not estimate it by borrowing a multiplier from another H3 Max route.

Long footage needs a second budgeting caveat. fal says larger or longer sources are processed in overlapping, checkpointed windows, and the generated overlap seconds are billable. A 120-second output therefore may be billed for more than 120 generated seconds. Check fal's usage and billing records rather than multiplying final duration alone.

For an inexpensive proof, trim the source to the shortest continuous section that still contains the difficult edge crossing. Validate the canvas and continuity there before sending a long master.

Reframe is only for changing the canvas

The broader H3 Max video-to-video guide now covers seven specialist jobs. Pick another route when the requested change is not the frame shape:

• use Extend to continue beyond the source ending;
• use Recast to keep a performance but replace people;
• use Relight to keep the shot and change its illumination;
• use Insert when a weak interval between two useful source moments must be replaced;
• use Reference-to-Video when a clip should guide a new generation rather than retain its timeline.

Reframe is the right contract when the approved event already happened and the distribution canvas is the problem. It is the wrong contract when the performance, cast, light, or ending is the thing that needs to change.

The best reframe is not the one that invents the most scenery. It is the one where the new canvas stops calling attention to itself—and the original moment still reads exactly when it should.