I’m working under a tight client deadline and can only use exported MP4 files. I tested the same 17-second prompt at a 9:16 setting, but the character’s face and clothing changed between shots more than I expected. Does the best AI video generator handle continuity reliably, or are others seeing this too?
A hidden cost of rendering the whole clip in one pass is that every additional second provides the model with an opportunity to reinvent the character. Facial features, hair style, clothing elements, and accessory details may transform if the camera views the character from different angles, the subject leaves or enters the frame, or the surroundings change.
The current state of the art does not allow for full continuity reliability from a text prompt. Picking the tool that has the best reference image controls, image-to-video generation, and first/last-frame editing would be a better choice than relying on a text-to-video demonstration. A fixed seed may work for the beginning section, but it will rarely preserve the initial appearance throughout the length of the video.
One should split the task into multiple shots, each rendered from a character image or the concluding moment of the previous scene. Keeping the clothing as simple as possible and avoiding occlusions of more than a second would be vital in this approach. Additional frames should be generated at the beginning and the end and later edited out to create a continuous take. It would be easier and faster to compile the final result from several such shots than attempt to re-roll an entire video timeline.
The vertical framing may exacerbate the issue since the person occupies most of the space in the image. Thus, their facial features, hairstyle, accessories, and clothing elements, such as collars, logos, or bracelets, would be more apparent in the generated video. One should plan on extra editing time to fix discontinuities between consecutive takes, mask irrelevant parts of the video, or utilize frame editing options if they become available. In general, video editing software would benefit from being able to maintain an entire take’s general mood and motion rather than rely on character consistency across multiple shots.
If “MP4 only” means you cannot re-encode or edit the output, the delivery format may be the bigger trap. A file ending in .mp4 can still use a codec or variable frame rate the client’s system rejects. Confirm they need H.264 and check the generator’s export specs first. For the character issue, prioritize reference-image or character-lock controls over maximum clip length. I agree with @cursorpro that a single long generation is risky, but splitting shots only helps if you’re allowed to combine and export them afterward.
The problem is not in the length. The real issue is that text to video is not a continuity tool but a continuity source for the same identifiable person.
I would go with record or animate a loose vertical performance before the deadline and attempt an edit with video to video, using a clean character reference and tight wardrobe spec. The source has to give enough data for the model to have the framing, timing, and camera moves, making it less demanding than text.
A few details stand out more than the name of the generator:
- Use a reference where the face is large, neutral, and evenly lit.
- Provide a second image showing the full outfit if the tool supports multiple references.
- Remove logos, jewelry, patterned fabric, and loose accessories unless they are essential.
- Keep the character facing within a limited range instead of turning fully away and returning.
- Avoid hands crossing the face, fast spins, hair covering facial features, and exits from the frame.
- Write the prompt as a description of appearance and motion, not a sequence of camera directions.
And so is the example that the assistant has given after. An input “close-up, then side view, then wide shot” will most likely not give you a single continuous clip. Rather it will demand the model to create the same character multiple times, each in a different viewfinder perspective. If you want an output of a continuous “take”, you have to ask for a single camera angle and movement that would allow for such a take.
I would partially disagree with the claim that shorter generations automatically address this concern. They do help with the issue of “drift”, but they could produce even more egregious identity shifts in the join between takes. If you have the option to edit the output prior to delivery, separate takes are not an issue. For a client that requires an unedited generator output, a continuous take is probably the safest option.
The service I would pick for this type of work would either allow me to use input video or images to define the animation. In an ideal scenario, the generator would explicitly allow for control over the reference strength, while the maximum length of such a clip would be somewhere near the lower end of the continuum. The closer the generated output follows the input motion capture, the less room there is for it to deviate and invent new movements or facial expressions on its own. Texture flickering is a bigger issue than identity drift in my opinion.
A model native clip and a clip extended to meet an unusual runtime are two different jobs, and the second is vastly more likely to drift because the extension point essentially gives the generator another chance to reimagine the person. Writing ‘17 seconds’ in the prompt will rarely control the exported timeline as well, and many generators work in preset clip lengths and then rely on the continuation or interpolation to reach the requested duration by either side.
Under a deadline, I would stop looking for the generator with the longest advertised output. Pick the one that lets you supply a character image, generate from that image, and continue from a chosen frame. Before committing to the full render, do a cheap motion test using the exact reference, vertical crop, and camera movement. If the face started wandering in that short test, more prompt detail will probably not rescue the longer version.
The safe setup is the deliberately boring shot. Keep the person in roughly the same screen area and avoid profile to front transitions and use a slow camera move instead of several implied cuts. A prompt containing ‘close up, walking shot, overhead view, then reveal’ is essentially four separate casting calls. For a continuous result, describe what stays the same if they are as clear as what moves: same lighting direction, same hairstyle silhouette, same jacket closure, uninterrupted camera movement, no scene transition.
I would mildly push back on making the wardrobe too generic. Plain clothes reduce artifacts, but a completely featureless outfit gives you fewer visual anchors for checking continuity. A solid jacket with one stable structural detail, such as a contrasting collar, can work better than either a complex patterned costume or an undefined shirt that changes shape every few seconds. The useful distinction is structural detail versus decorative clutter.
There is a simple deadline workflow here:
- Create and approve one clean portrait plus one full-body character frame.
- Crop those references for the vertical composition before uploading them.
- Generate the central action first, since that is usually the hardest section.
- Build the opening and ending around frames taken from that successful section.
- Assemble the result in an editor if the client merely requires the final delivery to be MP4.
- Export a short technical sample early and confirm that the client can actually play it.
That last check matters because @neonguru’s codec warning is real, but it should happen before the creative work is finished. Confirm the frame rate, audio requirement, dimensions, and H.264 compatibility with a disposable sample. “MP4 only” describes the container, not necessarily what their playback or upload system accepts.
If editing is disallowed, and they expect you to present a single untouched generator download, I would temper expectations rather than promise identity. In that case, use one reference-driven take, minimally turned, no exits from frame or occluded wardrobe pieces. If editing is allowed, the best generator is simply the one which offers the most repeatable short segments, to be edited on a timeline rather than relied on for the entire runtime via prompt.
Before you burn hours picking a generator, go back to the client and get a straight answer on whether editing is actually off the table. Nine times out of ten ‘MP4 only’ is someone in procurement describing the file they want to receive, not a rule that the footage must come raw out of the model. If they just need a playable H.264 file at the right dimensions, the whole single-take constraint disappears and you can stitch shots on a timeline like @zoetheproxy laid out. That one clarifying email changes which advice in this thread even applies to you.
The other thing nobody has pointed out: 17 seconds is a bizarre length to lock into. Most of these render in fixed native blocks and then pad the rest with continuation or interpolation, which is where the face starts to wander. If the client picked 17 arbitrarily, ask if 15 or 20 will do, since it’s landing on a length the model can cleanly generate that’s more important than the length itself.
I’m with @binaryspark2418lab that driving it from a source video is your safest bet under pressure. Text prompts leave too much open. Where I’d push back a little is the advice to keep everything static and boring, that reduces artifacts but it also makes any drift more obvious because the viewer has nothing else to look at. A slow deliberate move with one clear anchor detail hides small identity wobble better than a locked-off shot does. Run a short reference test first, and if the face won’t hold for a few seconds, it won’t hold for the full clip no matter how you word the prompt.