Character consistency: the reading that separates companion apps from studios

Ask a generator for the same character twice and you get two strangers. That is the default behaviour of every diffusion model, and solving it is most of the difference between a picture and a character.

Why it happens

The model samples from noise. Without something pinning the identity, every generation starts fresh. Prompts describe a type, not a person.

How tools solve it

Reference images, identity adapters and character embeddings. Companion apps do this automatically: the character has a fixed identity and every image and video is conditioned on it. Studios expose the tools but leave the work to you. ourdream does both: a free-form prompt box and a character layer that pins the identity.

Why it is a separate reading

Realism asks whether an image looks real. Consistency asks whether four images and a video look like the same real person. A tool can score high on one and low on the other, and most do. ourdream and HeraHaven score highest on consistency because they are built around a character; Seduced scores highest on control because it is built around a prompt.

How to test it

Generate a character, then ask for them in four different scenes and one video. Line the outputs up. If you would not recognise them as the same person, the tool does not do consistency, whatever it claims.