Almost every AI video tool can produce a good clip. Almost none can produce the same character twice. That gap is the difference between a video that does well once and a channel people subscribe to.
Write an episode freeA video model has no memory between requests. Each clip is generated from scratch, and anything your prompt did not pin down is filled in fresh. That is not a bug you can prompt your way out of; it is what the request looks like from the model's side.
So consistency has to live outside the model. Something has to store what the character looks like, sounds like, and behaves like, and re-supply it on every single scene. That is a pipeline problem, not a prompting problem, which is why longer prompts usually make it worse rather than better.
From our own production runs
Six-second clips hold one written line and one action reliably. At eight seconds the model begins to improvise: characters say lines nobody wrote and faces drift inside the shot. Frutti therefore builds every episode from six-second scenes and adds length by adding scenes, from 5 on the shortest format to 12 on the longest, rather than by asking for longer clips.
Four fixed traits are enough: look, voice, flaw, and the kind of scene the character belongs in. Write them once and never rephrase them. Rewording a trait between episodes is the most common way a cast quietly becomes a different cast.
Text alone cannot pin a face. The same reference image has to travel from scene to scene, otherwise every scene starts from a description and lands somewhere slightly different.
Drift is a function of time. A model asked for six seconds delivers what you wrote; asked for twelve, it fills the gap with its own ideas. Build length out of scenes, not out of duration.
Consistency the audience notices is behavioural. If the lemon is the one who says the unkind true thing, viewers will forgive a shade of yellow. If it is only a colour, any change reads as a different character.
This is a comparison of two workflow shapes, not of named products. Plenty of excellent clip generators exist; they are simply solving the first step of a longer job.
| Step | Clip generator | Episode pipeline |
|---|---|---|
| Generates moving footage from a prompt | ||
| Keeps the same face across scenes | ||
| Writes the story before the shots | ||
| Assigns and holds a voice per character | ||
| Cuts the scenes into one episode | ||
| Burns in captions | ||
| Returns something publishable |
Describe a scene with two characters. Frutti writes the episode with a fixed cast, roles, and spoken lines, free and without an account.
Because most generators treat each clip as a fresh request with no memory of the last one. The model rebuilds the character from the prompt every time, and any detail you did not repeat is re-invented. Two clips generated from the same sentence will differ in face, colour, wardrobe, and camera distance unless something outside the model holds those fixed.
Four things: a written character sheet that never changes, the same reference image carried across scenes, short clips so the model has less room to improvise, and a story that gives the character a fixed role. Prompt length is not one of them. Longer prompts usually make drift worse, because they add details the model has to reconcile.
Shorter than feels natural. In our own production runs, six-second clips reliably hold one spoken line and one action. At eight seconds the model starts to improvise: characters ad-lib lines that were not written, and faces drift within the clip. Building a longer episode from more short scenes is more reliable than asking for fewer long ones.
Yes, and it is the point. A cast that returns turns individual videos into a channel people follow. Frutti holds the same look, voice, and behaviour for each character across every episode you generate, so viewers recognise them before reading the caption.