Why Does My AI Actor's Face Keep Changing?

An AI actor's face changes between generations because most video and image models re-imagine that face from scratch every time you generate — a text prompt describes a type of person, not a specific one, so an unanchored generation is a fresh independent guess each time. On Flovaly, that risk is mostly designed out once a character exists: the character's actual saved photo is reused as the literal start frame for Talking Actors, base video models, and Face Swap, rather than being re-described in words. There is one honest exception, and one internal step that can still introduce drift — both covered below.

Why AI faces drift in the first place

Image and video generation models don't hold a persistent memory of “your character” between calls. Each generation samples from the model's own distribution given whatever it's conditioned on. If that conditioning is only a text description — “a woman in her 30s with brown hair and a warm smile” — the model has to invent the specific face every time, and there are countless faces that match that description equally well. Nothing in the prompt pins down bone structure, exact eye shape, or the small asymmetries that make a face read as the same person. That's the core reason character consistency is a known, actively-worked-on problem across the industry, not a bug unique to any one tool.

Two different consistency problems

It helps to separate “picking a character's first look” from “keeping an established character consistent.” They have different causes. When you create a Flovaly character from a text description instead of a photo, the three portrait options you pick from are genuinely independent samples — Nano Banana Pro, the image model behind that step, is stochastic per call, and Flovaly nudges the three options apart on purpose with a different expression hint in each prompt so they don't look identical. There's no seed or reference-image locking wired into that specific step, so don't expect the three options to look like the same person wearing different expressions — expect three different plausible people matching your description, and pick the one you want as your character's anchor going forward.

Once you've picked (or uploaded) that anchor image, the second problem — keeping later generations consistent with it — is a different mechanism entirely, and it's the one this article is mainly about.

How Flovaly keeps a saved character's face steady

A Flovaly character's primary reference image isn't just a thumbnail — it's the actual start frame most generations animate from. For any video model that supports image-to-video (Seedance 2.0 Pro, Veo 3.1 Fast, and Kling 3.0 in the current picker), the worker passes your character's stored photo straight through as the starting image, and the model animates those real pixels rather than reconstructing a face from your text prompt. Talking Actors works the same way — OmniHuman 1.5 animates directly from the single still image plus your audio track, so the face in the output is the face in that photo, not a reinterpretation of it. Face Swap is the most literal case of all: it takes that stored image and swaps it directly into existing footage, rather than generating a new face at all.

There's one honest exception: Sora 2 Pro is registered as text-to-video only, with no start-image input at all. Pick it and you're trading the fixed-photo anchor for its cinematic camera work — a fair trade for a stylised establishing shot, a weaker one if the same recognisable face needs to carry across multiple clips. See the full model comparison for how the picker's other options differ.

Where drift can still creep back in

Not every Flovaly pipeline reuses the stored photo completely untouched. Product UGC ads add one extra step: before animating, Nano Banana Pro composes a new start frame from two references — your character's photo and your product photo — so the character appears to be genuinely holding the product rather than a hallucinated stand-in. That compose step re-renders the frame, which means it's the one place in Flovaly's own pipeline where a slight variation in the face can appear, even though the same source photo drives every attempt. It's a deliberate trade for a real product appearing in shot — see how Nano Banana Pro composes reference images — and in practice the effect is subtle, not a different-looking person, but it's the honest reason a UGC product clip can drift very slightly more than a plain Talking Actor clip from the same character.

Practical steps to keep a character consistent

• Upload a real photo when the exact face matters. If you're building a character around yourself or a specific presenter, upload their photo directly rather than generating from a text description alone — a real photo is a fixed anchor from the first generation, with nothing left for the model to guess.
• Reuse the same saved character, don't rebuild it per clip. Every generation from one character pulls the same primary image. Recreating a “similar” character from scratch for each new video reintroduces the exact randomness you're trying to avoid.
• Pick an image-to-video model over Sora 2 Pro when identity carries the clip. Seedance 2.0 Pro, Veo 3.1 Fast, and Kling 3.0 all animate from your character's actual still; Sora 2 Pro doesn't take one at all.
• Draft at 720p before committing to 1080p. Confirm the likeness holds up on a cheap generation first — see the cost breakdown for how the 1080p multiplier applies — rather than discovering a weak likeness after paying the surcharge.
• Use a clean, well-lit, front-facing reference photo for product ads. Because the UGC pipeline recomposes the frame around your product, a sharp, plainly-lit starting photo gives that step more accurate material to work from than a dim or heavily angled one.

None of this is a Flovaly-only quirk — every AI video and image platform faces the same underlying problem, because none of them remember a face between calls unless something anchors it. The difference is whether a product's default workflow anchors to a real photo automatically or leaves you to work around drift yourself. On Flovaly, saving a character once and reusing it does most of that work — a $1 trial is enough credits to generate a few clips from the same character and see how closely the face holds.

FAQ