Ask how any platform keeps an AI character's face the same across a hundred images and you'll find one of two architectures underneath: either the platform trained a model on that face, or it shows the model reference images of that face every time it generates. LensGo does the second, and this post explains what that means in practice — including where the trade-offs genuinely cut the other way.
The two architectures
Training a personal model. Some platforms fine-tune a model — often a LoRA, a small adapter bolted onto a large image model — on photos of one subject. You typically assemble a dataset (commonly twenty or more images of the face), launch a training job, wait for it to run, and get back a custom model that has learned the identity. Every future generation runs through that personal model.
Reference-based identity. The alternative skips learning entirely. The character exists as a small set of curated reference images, and at generation time those references are handed to the model alongside the prompt. The model doesn't know the face; it looks at it, every single time. Identity comes from the references, scene comes from the prompt.
Neither approach is a gimmick — both are used seriously across the industry. But they behave very differently, and which one you're standing on shapes everything about the product experience.
What a saved character actually is in LensGo
In LensGo, a saved character (a Lens ID) is three reference images of one face, generated and curated at creation time.
Note what it isn't: there's no photo upload. You don't bring pictures of a real person. You either pick one of the five presets or design a face in the Builder — gender, age, look, hair, eyes — and shuffle until it's right. Once you approve the face, LensGo generates it from additional angles, and those shots become the reference set. From that point the character is fixed: a compact, curated identity anchor stored with your account.
Creation is measured in moments, not sessions, because nothing is being trained. There is no dataset to assemble, no job to queue, no progress bar to babysit.
What happens at generation time
When you shoot the character into a new scene, LensGo conditions FLUX.2 — an image model with multi-reference support — on the character's reference images plus your prompt. The model composes the scene the prompt describes while holding the face the references show.
This is the architectural difference in one sentence: a trained model carries the identity in its weights; a reference-conditioned model carries the identity in its inputs. Three practical consequences follow:
- The character is portable across everything the model can do. The same three references anchor a café scene, an editorial portrait, and a product shot — and the Cast a creator flow extends it to ads by adding a product image alongside the identity references, so the still shows your character holding that product.
- Improving the character doesn't mean retraining. With a trained model, a bad dataset means training again. With references, better curation of the reference set directly improves every future generation. The anchor is editable in a way baked weights never are.
- Nothing about the identity is trapped in a custom model. There's no per-character model artifact sitting on someone's GPU cluster that your persona depends on.
The honest trade-offs
Reference-based identity is not universally better, and it's worth being precise about where each approach wins.
Where training holds an edge. A model that has genuinely learned a face can hold identity more tightly at the extremes — very wide angles, heavy occlusion, hard profiles, or aggressive style transfer where the subject is rendered far from photorealism. A LoRA that has internalized a face can keep it recognizable as a watercolor sketch viewed from behind-and-above; a reference-conditioned generation drifts more readily out there. If your entire use case lives at those extremes, training is a rational choice.
Where references win.
- Speed to a usable character. Minutes instead of a dataset-plus-training-run. The persona can post today.
- Cost structure. Photos with a saved character cost 3 credits — and with a preset character, free daily credits cover a shoot a day (free-tier output watermarked). There's no separate training fee amortized into every character you create or revise. Custom characters, video, and the Builder use paid credits.
- Photo and video from one identity. The video path is image-to-video: you approve a still of your character, then animate that frame into a 5- or 10-second clip at 480p or 720p. The clip inherits the face from the image, so the same reference set powers both formats. (Clips are silent B-roll — no lip-sync or speech.)
- It improves with curation, not re-training. The ceiling on consistency rises as the reference set and your prompting discipline get better — an editorial skill you can exercise, rather than a compute bill you re-pay.
Where both approaches are equal. Neither exempts you from the craft: extreme shots are the hardest case for any system, curation matters everywhere, and a character only reads as one person if their world is coherent too. The fundamentals in what is character consistency apply regardless of architecture.
Why video doesn't need training either
A common assumption is that consistent video requires a trained model. The image-to-video path routes around it: identity is settled before motion begins, because the video starts from a frame you approved. The i2v model's job is motion — a push-in, a head turn, ambient life — not identity, which is already in the pixels.
That's also why the casting flow can afford a human checkpoint. Cast a creator generates the ad still first (a 4-credit step: identity references plus your product image), you approve exactly the frame that will move, and only then is the clip generated at normal UGC video pricing. Nothing animates that you haven't signed off on.
Choosing for your use case
- You want a persona posting this week, across photos, clips, and product ads, with the freedom to iterate on the character cheaply: reference-based is built for this, and it's what LensGo ships.
- Your work lives at stylistic and geometric extremes — heavy stylization, severe angles, the character rendered far from photoreal — and you're prepared to assemble datasets and wait on (and re-run) training: a trained personal model earns its overhead there.
For creators and brands running an actual content calendar, the first profile is overwhelmingly the common one — which is why LensGo chose references. Create a character and test the claim on your own feed: same face, new scene, no training run in sight.
