H3 Max character consistency needs more than putting the same name in two prompts. A name is not a saved actor. To carry a person from one clip into another, give the next generation a suitable starting image and check the returned person against the original—not just against the previous take.
On AI Live Generator, that means working with first frames in Image to Video. We use H3 Max Turbo, and our interface has no character library or identity-lock control. You can plan a coherent sequence, but matching faces across separate generations is not guaranteed.
The distinction matters most when a good first clip persuades you to generate five more. A small change in the second person's face can become the starting point for the third. By the end, the sequence may be internally smooth and still feature someone different.
Decide what the second clip is supposed to inherit
There are three different jobs hiding inside “use the same character.”
Another version of the same shot: reuse the original source image and change the action brief. Each result starts from that composition. This is useful for alternate gestures or expressions, not for cutting the results together as one continuous action.
The next moment in the same scene: extract an acceptable frame from the first result and use it as the second clip's opening image. This is a manual handoff between two jobs, not a video-extension feature.
A new location or camera angle: prepare and approve a new opening image outside this generator. It needs to show the person in the new composition already. Reusing a workshop portrait does not turn it into an identity-only reference for a beach scene; the uploaded image also supplies the workshop, pose and framing.
The image-to-video guide covers source-image preparation. For now, choose the second job: two adjacent moments, with no location change to disguise a continuity problem.
A two-clip brief: the ceramicist and the blue bowl
Imagine a short workshop film. A ceramicist turns a bowl on a bench, then leans forward to examine its rim. The interest is in her attention to the object, not in a dramatic camera move.
Prepare an opening image you own or have permission to use. It should show one adult ceramicist, with her face and both hands visible, a blue bowl on the bench, and enough space for the small movement. Use a medium framing rather than a distant full-body view.
Keep a small note beside the original image:
Person: compare every take with the approved original face and hair.
Clothes: plain charcoal shirt, pale ochre apron, no added accessories.
Object: one blue bowl, already on the bench; same rim and surface markings.
Handoff: bowl has stopped turning; right hand rests beside it, not on it.
Light and camera: window light from screen left; fixed medium framing.Change these details to match your actual image. The note is for your decisions; it is not a character profile stored by the app.
Here is an original, untested prompt for the first clip:
Continue from the supplied workshop image. The ceramicist gently turns the blue bowl a quarter turn on the bench, releases it, and rests her right hand beside it. Keep her recognizable face, tied-back hair, charcoal shirt and pale ochre apron unchanged. The bowl stays on the bench. Fixed medium camera, steady window light from screen left, one continuous shot. Audio: a soft ceramic scrape and quiet workshop ambience, no dialogue or music.Select Image to Video, 5 seconds and Balanced expansion. Use the same resolution for both jobs. Our current rate makes two five-second generations cost 30 credits at 480P or 50 credits at 768P, before any extra takes. That buys two requests—not a guarantee of ten usable seconds. Check the displayed cost before submitting.
Choose the handoff frame—not simply the last frame
Download the first clip and inspect it in an external video editor. Look at the opening, the moment when the face turns furthest, and the intended ending. A matching first frame is not enough if the face changes halfway through the action.
Find a clean frame after the bowl stops, with the face, apron and hands still acceptable. If the final moment has a distorted mouth or an extra finger, do not pass that frame forward. Choose an earlier acceptable moment and trim the first clip to the corresponding cut point. If no frame supports the intended handoff, revise or reject the take instead of building on it.
Export that frame as a still image at the clip's dimensions, without player controls or screenshot borders. Keep the untouched original source image too. The extracted frame tells the next job where to begin; the original remains your reference for who the person is.
Upload the extracted still as the second job's first image. Leave the optional last image empty for this example. It describes an ending composition, not an extra face reference; the first-and-last-frame guide explains that separate use.
For a handoff frame that actually shows the state above, the second prompt could be:
Continue from the supplied frame. With the blue bowl stationary on the bench and her right hand resting beside it, the ceramicist leans slightly forward to inspect the rim, then returns upright. Preserve the person, tied-back hair, charcoal shirt and pale ochre apron visible in the image. Do not pick up or turn the bowl. Keep the fixed medium framing and steady light from screen left. One continuous shot. Audio: quiet workshop ambience and a small cloth rustle, no dialogue or music.If the first result leaves her holding the bowl, this prompt no longer fits. Either choose another cut point or rewrite the second action around the actual state. A prompt cannot make an inconsistent handoff disappear retroactively.
Watch the person, then watch the cut
Place the two clips next to each other in your external editor. First, compare stills from each clip with the original image. Did the hairline change? Did the apron acquire a pocket? Are the bowl's markings still the same shapes, without new ones appearing? Check details at a similar face angle: a different shadow alone does not establish that the identity changed.
The bowl's orientation is a different check. It was supposed to turn in the first clip, so its markings should move with it. Compare the second clip's opening with the approved handoff frame, not the original bowl angle. Its position and orientation should carry forward rather than reset.
Then play across the cut at normal speed. The person can match while the action does not. A hand that jumps from the bench to the bowl, a reversed gaze, or a sudden change in light direction can make two individually attractive clips unusable together.
A duplicated opening frame or a brief hold may need trimming. A different face needs a different decision: use another take or return to a clean source. A dissolve can blend images; it cannot turn two different people into one convincing performance.
Do not judge only the seam. Review the second clip's full length against the original before using any of its frames for a third job. That check prevents accumulated changes from becoming your new definition of the character.
When the next scene really is somewhere else
A new scene should earn a new approved opening image. Prepare it with an external image workflow, check the face and intended wardrobe against the original, and only then animate it here. If the still already looks like a different person, spending video credits will not establish a dependable identity match.
This is where tutorials for other MiniMax workflows can be misleading. A dedicated subject-reference input and an opening-frame input have different jobs. The official Turbo image-to-video input schema documents opening and ending images; our generator does not expose a separate subject-reference mode. Nor does it offer a seed control that can serve as a saved actor.
Sound needs its own continuity decision. Two clips showing the same face are not guaranteed to produce the same voice. For recurring dialogue, plan the voice track separately and finish it in an external editor; see the lip-sync troubleshooting guide before assuming a soundtrack replacement will match the mouth movements.
Start with two clips you can actually cut together. Keep the original image, approve the handoff, and reject drift before extending the sequence. When that brief is ready, open the H3 Max Generator and make the first shot—not the whole series at once.