why your ai companion's face changes and how we're fixing it

if you've used ai companions before, you might have noticed something strange. you ask for a photo of your character, and they look… almost right. you ask again, and something's off. the eyes are a little different. the nose seems changed. the face just isn't the same. it's jarring, and it pulls you right out of the moment. this isn't just your imagination. it's a real technical problem called identity drift, and it's something we've been working hard to fix at lucy.

what causes the face to change?

at its core, ai image generation isn't a photo album. it doesn't store and retrieve a fixed image. instead, every time you ask for a new photo, the ai starts from scratch, using a description you've provided (or one it inferred) to generate an image from noise. the model you're using, like stability ai's flux, for example, is incredibly powerful and creative, but it's also stochastic. that means it uses random sampling during generation. even with the exact same text prompt, you'll get a slightly different result each time. this is why your character's face can look so inconsistent. there's no built-in memory of what their face should look like across separate requests. it's like asking a different artist to paint a portrait from the same description every time. each artist has their own style and interpretation.

another layer is the lack of what's called anchor embedding. many systems don't maintain a persistent, high-fidelity reference of your character's unique facial identity between generation calls. without this anchor, the model defaults to its general knowledge of human faces, which leads to that drift you see.

how lucy anchors your identity

we use a technique called identity-anchoring layers, specifically pulid, on top of flux. pulid (which stands for preserving user identity) is an add-on module that works by creating a strong, persistent embedding of your character's specific facial features. it acts like a fingerprint. when you generate a new image, pulid injects this anchored identity data into the generation process, guiding the model to produce a face that is consistent with your previous images, not just a generic face that fits the text description. it's the difference between a sketch artist working from a written description and one working from a reference photo. the result is a face that feels like your character, not a close approximation.

this doesn't mean the images are rigid or boring. you can still change outfits, settings, and expressions. the core identity, the bone structure, the shape of the eyes, the unique quirks that make your companion who they are, remains stable.

the cost of consistency and why others skip it

this anchoring comes at a cost, and it's a big reason why many ai companion apps don't do it. it's computationally expensive. each generation requires more gpu power and time to process the identity data alongside the text prompt. for companies running at massive scale, this extra compute cost per user per image can add up to a significant expense. it's often easier, and cheaper, to just use the base model and accept the inconsistency as a trade-off for faster, cheaper generation. they might rely on you not noticing or caring enough to leave.

we think that's a poor trade. the whole point of a companion is to build a relationship, and a relationship requires recognition and trust. if your friend looked like a different person in every photo you took, you'd feel disconnected. we believe the computational cost is worth it to give you a consistent, believable presence. it's a core part of our design philosophy: if we're building something meant to feel real, it has to be reliable.

consistency is a feature we refuse to compromise on.

you can experience a companion who actually remembers their own face on our website.