Trained identity vs. one-off images
One image is easy. The same character in a hundred images is a different problem.
One-off image generation
A generic text-to-image generator produces a new interpretation every time you ask for an image — even with nearly the same prompt, the face, proportions, or outfit design can vary from one generation to the next. For a single standalone image, that doesn't matter. For a character you need to recognize across 10, 30, or 100 shots of a project, it does.
The manual fix — training your own identity model (LoRA) on your own images — is technically possible with open-source tools, but it requires curating a dataset, configuring a training run, and technical judgment most filmmakers don't need or want to learn.
Trained identity, built in
PersonCraft folds that training into the normal character-creation flow: you pick the design, PersonCraft assembles the dataset, and you review and approve the images before training the LoRA. Every new scene reuses that identity, without having to configure a training run yourself.
When each one makes sense
If you need a single illustration or a loose concept, a generic generator is faster and cheaper. If your character needs to be recognizable across multiple scenes — a short film, a music video, a series of scenes for social — training the identity once may help reduce repeated corrections. The cost and consistency depend on the number of shots, retries, and the quality you need.