The gradual or sudden change of a generated character's face, build or style between images. Drift is the main failure mode of prompt-only character workflows and the problem trained, identity-locked models exist to solve.
Drift shows up as a persona that is recognizably 'the same kind of person' but not the same person: the jawline narrows between posts, apparent age slides a few years, the build changes with the outfit. Audiences register it even when they cannot name it — a feed with drift reads as stock photos of lookalikes rather than one creator's account.
The root cause is almost always identity living in the prompt. Every text description is re-interpreted per generation, and small wording changes ('portrait' vs 'candid photo') land in different corners of face-space. Secondary causes include contradictory descriptors stacked in one prompt and heavy style transfer bleeding into facial features.
The durable fix is structural: train the identity so it no longer depends on the prompt, keep prompts situational, and reuse seeds when iterating on a specific look. The drift-fixing guide diagnoses the common patterns, and the consistency guide covers preventing drift by design.
Related terms
Fixing a persona's facial identity at training time so that every subsequent generation shares the same face and build. The alternative — re-describing a character in each prompt — produces visible identity drift between images.
A number that initializes the randomness of a generation. Reusing the same seed with the same prompt and settings reproduces the same image, letting you revisit and iterate on a look instead of re-rolling from scratch.
The one-time process of teaching an image-generation model a specific persona's identity from a set of reference photos. After training, the persona can be generated in unlimited new scenes without re-describing its appearance in each prompt.