Not prettier shots. The same person, recognisable in episode one and episode twenty, wearing what they wore, sounding how they sound, remembering what they did.
Why AI characters drift
A generation request produces a plausible person matching a description. Ask twice and you get two plausible people. Nothing in that process is trying to return the same individual, because nothing in it holds a definition of who that individual is.
Describing them in more detail narrows the range but never closes it, and past a certain point extra detail makes results less stable rather than more, because every added constraint is another thing the model can weight differently. Consistency requires a stored reference that every request is anchored to.
- The character exists only as a prompt.
- Every generation re-imagines them from scratch.
- More detail narrows variance but never eliminates it.
- No way to evaluate whether a result is correct.
- A good result cannot be reproduced deliberately.
- The character exists as approved references you accepted.
- Generations are anchored to those references.
- Results are evaluated against them rather than against a description.
- Deviation is detectable, which is what makes it correctable.
- A good result becomes the reference for what follows.
What gets locked, property by property
Consistency is not one switch. It is a set of properties, each tracked separately, because each one fails on its own:
- Face and body. Approved references travel with the character into each shot, and results are evaluated against them.
- Expressions. The emotional register the character has been approved to play, so a restrained lead does not suddenly overplay.
- Costume state. Not just the outfit but its current condition, including damage and changes that happened on screen.
- Voice. The approved voice, local phrasing, names, and pronunciation. A new voice requires creator review.
- Significant props. Objects the character carries and the plot depends on, tracked as deliberately as the character.
- Relationships. How this character stands with each other character at this point in the season.
Locations, lighting, and visual rules are treated the same way, so a consistent character appears in a world that is also consistent. A stable face in a room that rearranges itself is still a continuity failure.
Consistency beyond appearance
Looking the same is only half of it, and it is the half people think of. The other half is behavioural: a character who saw something in episode two should not behave in episode six as though they did not.
Tosheo tracks story state alongside the visual references: what happened, who witnessed it, what changed, and what remains unresolved. This is what makes a character continuous rather than merely identical, and it is the failure mode audiences describe as a show not respecting itself.
One character, six episodes
The recurring lead of the demonstration season, whose approved references and story state carry from episode to episode.
- Locked at approval. Face and body references. Ivory kurta, rust dupatta, silver earrings. A quiet, precise voice with emotion held beneath restraint.
- Changes on screen. The dupatta is torn in episode three. From that point the torn state is the current costume state, not the original reference.
- Accumulated knowledge. She has read the name on the ticket. She has chosen not to tell Kabir. Both facts constrain how later scenes can be written.
- What a drift would look like. An intact dupatta in episode four, or a scene where she speaks freely about the ticket. Both are caught by the continuity check.
Note that two of these four rows are not about appearance at all. A character is a visual reference plus a history, and both have to be enforced.
How to define a character so it holds
The way a character is specified affects how stably it renders. What works, in practice:
- Describe silhouette before detail. Hair shape and build survive generation better than freckle placement.
- Choose one or two distinctive, simple identifiers rather than many subtle ones.
- Define costume as a small number of named states, not an open wardrobe.
- Describe behaviour, not adjectives. "Slow to answer" renders; "mysterious" does not.
- Give the voice a register and a rhythm, not just an accent.
- Approve fewer characters than you think you need. Every recurring character is a continuity obligation.
What this does not promise
Tosheo tracks these properties, evaluates generated results against approved references, and surfaces contradictions for review before generation. It does not promise perfect consistency, because no current pipeline can, and a promise like that would be the least trustworthy claim on this site. You retain final approval over every episode, and correction happens at the shot level so a deviation costs one shot rather than an episode. Tosheo is in closed validation and is recruiting a focused founding creator cohort. Applications are open.
Questions creators ask
How does Tosheo keep characters consistent?
By storing approved references for each character, covering faces, expressions, body references, costume states, and voices, and anchoring later episodes to those references rather than re-describing the character each time.
Is character consistency guaranteed?
No. Continuity is actively managed and evaluated rather than guaranteed. Contradictions are surfaced for review before generation, and you retain final approval over every episode.
Does consistency cover voice as well as appearance?
Yes. Voices, local phrasing, names, and pronunciation guidance are approved and carried through the season. A new voice requires creator review.
What if I want to change a character mid-season?
Costume states and character development are part of the tracked state rather than fixed forever, so deliberate changes are supported. You approve what becomes canon, so an intended change is distinguishable from drift.
Why does adding more description not fix drift?
Because every added constraint is another thing a model can weight differently, so past a point more detail makes results less stable. Consistency needs a stored reference to anchor to, not a longer prompt.
How many recurring characters should a season have?
Fewer than you think. Every recurring character is a continuity obligation that has to be honoured in every episode, so a small cast is easier to keep coherent and cheaper to produce.