The interesting question is not whether a model can render a beautiful shot. It is whether anything remembers what the last shot established.
Two categories that get confused
Almost everything marketed as an AI movie tool falls into one of two categories, and they are not competitors. One of them needs the other underneath it.
- Takes a description, returns footage.
- Measured on how good a single shot looks.
- Fast, cheap, and genuinely impressive.
- Holds no memory of anything it produced before.
- Correct tool for a mood test, a pitch visual, or a thumbnail.
- Takes decisions, enforces them, and then generates.
- Measured on whether a hundred outputs agree with each other.
- Slower, because approval gates are the point.
- Holds references and story state between episodes.
- Correct tool for a series with a returning audience.
Why generated video drifts
Every generation is an independent attempt. Without something external holding the references, each request produces a plausible result rather than the same result. The specific things that slip, roughly in order of how quickly an audience notices:
- Faces. Noticed almost immediately, especially in the close framing short-form depends on.
- Voices. Noticed as fast as faces, and often faster in a dark or tight shot.
- Costume detail. Collars, jewellery, damage, and layering shift between takes.
- Room geometry. Doors and windows move, which quietly destroys spatial sense across a scene.
- Props. The object the plot depends on changes shape, size, or colour.
- Story facts. The slowest to appear and the most damaging: characters acting on information they should not have.
Longer prompts do not solve this. The information that has to persist is state, and state has to be stored and enforced rather than described again every time.
What a production system adds
None of the following is a model capability. It is production discipline expressed as software, and it is the entire difference between a clip and a series:
- Approved references for cast, costume, voice, location, and props.
- A structured record of what has happened and what is unresolved.
- Contradiction checks that run before generation, not after.
- Cost estimation and approval before a render is authorized.
- Correction at the level of a single shot rather than a whole episode.
- An arc, so an episode ending withholds something the season will pay off.
The three things people try before this sinks in
Almost everyone attempting a generated series arrives at the continuity problem by the same route, and tries the same three fixes in the same order. They are all reasonable and none of them work, which is worth knowing before spending a month on them:
- 01 · Write a longer character description The instinct is that the model lacks detail. Past a point extra detail makes results less stable, because every added constraint is another thing that can be weighted differently.
- 02 · Reuse a seed or a reference image This helps within a shot and degrades across scenes. It holds appearance loosely and holds nothing about costume state, location geometry, or story facts.
- 03 · Keep a continuity document by hand This is the correct instinct, and it works until roughly episode four. Then the record and the footage diverge, because nothing enforces the document at generation time.
The third attempt is the interesting one. It fails not because the idea is wrong but because a document a human consults is not the same as a constraint a pipeline applies. Making that record authoritative, and checking against it before rendering, is the entire difference between the two categories above.
Choosing for the project you actually have
The honest recommendation depends entirely on what you are making, and for a lot of projects the answer is not Tosheo:
- One striking piece of footage. Use a clip generator. It will be faster, cheaper, and better at exactly this.
- A short film. A clip generator plus an editor and a lot of manual continuity work is a reasonable path.
- A narrated story video. Use a dedicated tool for that format. It is a different production entirely.
- A series with returning characters. You need continuity enforced outside the model, which is the category Tosheo sits in.
Tosheo is built for the last case specifically, for regional serialized microdrama. Tosheo is in closed validation and is recruiting a focused founding creator cohort. Applications are open.
Questions creators ask
Are AI movies real yet?
Generated short films and generated episodes are real and being made now. Feature-length work that stays visually and narratively consistent throughout is not a solved problem, which is why serialized short episodes are the practical format.
Why do my AI videos look different every time?
Each generation is an independent attempt. Unless approved references and story state are stored outside the model and enforced on every request, faces, costumes, locations, and voices are re-invented each time.
What should I use to make a series?
Something that treats continuity as a first-class concern: approved cast and world references, tracked story state, contradiction checks before generating, and correction at the shot level. That is the category Tosheo sits in.
Is a clip generator a competitor to Tosheo?
No. A production system needs generation underneath it. What it adds is memory, structure, and approval, which is a different layer of the same stack.
What drifts first when consistency fails?
Faces and voices, because an audience recognises people fastest. Costume detail and room geometry follow, and story facts are slowest to appear but most damaging.