H3 Max Insert Video: Replace One Scene, Keep the Rest

H3 Max Insert Video replaces one interval inside an existing clip, then returns to the original footage. The MiniMax H3 Max Insert workflow asks where the source should stop, where it should resume, and how long the newly generated scene should run. The result is a complete video: original opening, generated insert, original ending.

That makes Insert useful when most of a clip already works. You might add a product close-up, repair one weak beat, or create a reaction shot without regenerating the opening and ending you want to keep.

Insert is the separate upstream minimax/h3-max/insert-video endpoint. This site does not call it or accept a source-video upload. Its text and image modes use separate H3 Max Turbo generation endpoints, while reference mode is another workflow; none of them splices an existing video.

Insert, Extend, Recast, or a new shot?

These workflows can sit next to each other in an edit, but they solve different problems.

Insert is not an instruction to “improve this whole video.” It is a timed replacement. If the problem lasts two seconds but the generated insert runs five seconds, the finished clip becomes longer. If the source section and insert have equal lengths, the runtime stays roughly the same.

Start, Resume, and Duration are three different clocks

Suppose an 18-second narrative shot has a weak exchange from 6.0s to 10.0s.

• Start = 6.0 keeps the source up to that point.
• Resume = 10.0 tells H3 Max where to rejoin the source.
• Duration = 7.0 asks for seven seconds of newly generated footage between those two pieces.

For planning, the runtime is:

source length - (Resume - Start) + generated Duration

In this example, the four-second source interval is replaced by seven generated seconds, so the planned result is about 21 seconds long. Encoding and frame boundaries can introduce small differences, so treat the formula as an edit plan rather than a frame-accurate delivery promise. When the job finishes, use the returned start_time + injected_duration values—not only the requested markers—to locate the second join in the actual output.

The current official H3 Max Insert page sets Start at 1.625 seconds or later, requires Resume to come after Start, and keeps both markers within 60 seconds. Its upload interface labels the source limit as 60 seconds and 50 MB. Leave some usable footage after Resume; the model needs a destination to hand back to.

The current Insert contract

The live fal API schema, checked October 13, 2026, defines the following controls:

Color matching can reduce an obvious grade shift, but it cannot solve every continuity problem. A reference can help preserve a watch, outfit, room, or movement style; it cannot make contradictory camera angles meet cleanly. The two source frames on either side of the replacement still define the hardest part of the job.

Choose edit points before writing the prompt

A good insert has two seams. The first must leave the opening naturally; the second must land back in the original footage without making the viewer wonder what moved.

Look for edit points where:

• the camera is briefly steady or moving at an easy-to-read speed;
• the subject is not hidden by a hand, prop, or foreground object;
• lighting is stable rather than changing mid-flash;
• dialogue is not cut in the middle of a word;
• the returning frame has a clear subject position and screen direction.

Choose Start and Resume on frames that can survive a clean cut; avoid mid-word dialogue, occlusion, or a lighting change at either boundary. If the source contains music or speech that must stay perfectly timed, expect to rebuild that audio transition in an editor.

Write only the scene that belongs between the cuts

The prompt should describe the replacement, not retell the whole source and not ask the model to recreate the ending. Start and Resume already define where the original footage returns.

Imagine an 18-second scene in a rain-soaked bus shelter. The opening establishes two estranged friends facing opposite directions, while the ending already has one of them stepping onto a bus. What is missing is the reaction that makes that departure matter. An untested seven-second replacement prompt could be:

Seven-second medium two-shot in the same rain-soaked bus shelter at night.
The woman on frame left notices the folded letter in the man's hand, looks
up at him, and almost speaks; he closes his hand around the letter and looks
toward the arriving bus. Keep both characters on their established sides,
the locked shoulder-height camera, cool streetlight from frame right, wet
pavement reflections, and restrained natural acting. Audio: steady rain,
the approaching bus engine growing slightly louder, one quiet inhale. No
spoken words, embrace, camera move, new character, wardrobe change, or cut.

The prompt gives the insert one dramatic job: make the later departure feel earned. It also fixes screen position, camera height, light direction, ambience, and the approaching bus—the continuity details the two boundaries will expose. It does not ask the model to recreate the ending, because Resume already names the source moment that follows.

If likeness or wardrobe is critical, attach a clean authorized reference and name its job in the prompt. More references are not automatically safer. One clear, consistent view of each character is more useful than several compressed images shot under conflicting light.

fal's current Insert documentation does not specify how the source soundtrack is preserved across the replacement. Treat audio continuity as something to test and repair, not as a guarantee.

What the generated section costs upstream

fal currently lists Insert at $0.05 per generated second for 480P and $0.06 per generated second for 768P. Only the new scene is billed, not the source footage that surrounds it.

The first 4,096 reference tokens are included. Additional reference input is listed at $0.02 per 1,000 tokens, prorated. Those are fal's current upstream amounts, not this site's credit tariff or a checkout offer here. Recheck the endpoint before building cost assumptions into a client project.

Give the result a pass/fail review

The first output join sits at the returned start_time. The second sits at start_time + injected_duration, where footage from source time Resume begins again. Do not inspect the second join at the source's Resume timestamp unless the replaced interval and generated duration happen to match.

Use a small acceptance sheet instead of deciding from the prettiest frame:

Review the picture muted, then listen without looking. A compelling middle cannot rescue a broken hand-back. Keep the untouched source, save every generated version separately, and choose the take that passes both joins in normal playback.

If you only need a new angle, generating a separate shot here and making a visible cut may be simpler than Insert. The storyboard workflow helps plan that sequence, while the first-and-last-frame guide is better when you need a new clip to travel between two known still compositions.

Use Insert when the opening and ending already deserve to stay. The craft is not generating the loudest middle—it is making the new moment feel as though the edit always belonged there.