Stock photos do not look like you. AI images from a general model do not look like you either. They produce a plausible human; not your face, not your product, not the precise object your brand has spent time making recognizable. The reason is simple: the base model has never seen you. A LoRA fixes that.
A LoRA (Low-Rank Adaptation) is a small set of weight deltas you train on top of a frozen base diffusion model. The base learns general image quality during its original trillion-image training run; the LoRA learns this specific subject from a much smaller set of photos you provide. Training takes 20-90 minutes on a decent GPU. After that, every generation prompt has access to your face, your product, or your brand's visual identity as a first-class concept.
What a trained LoRA actually gives you
Once the adapter exists, you can drop your subject into any scene the base model knows how to render:
- Founder headshots at scale. One set of training photos, a dozen professional environments — conference stage, studio white, outdoor city. The face stays consistent across every frame.
- Product shots without a studio. Physical products are expensive to photograph in every context: on a table, in a hand, on a shelf, lifestyle, flat-lay. With a product LoRA you render the same item in each context from a prompt.
- Marketing creative that matches. Ad sets, social headers, email hero images — all featuring the same recognizable subject instead of a different stock-photo face for each campaign.
- Social content at pace. A personal brand posting five times a week cannot commission five photoshoots a week. The LoRA closes that gap.
Privacy: your photos stay on our machine
When you train a LoRA via one of the large consumer platforms, your reference photos and the resulting weights are uploaded to their infrastructure. Whether they are used for further training, retained indefinitely, or accessible to their engineers is governed by a terms-of-service document that changes without notice.
We run the full pipeline on our own GPU server. Training happens in an isolated job. The reference photos you send us are deleted after the weights are generated; we do not retain them beyond the training window unless you explicitly ask us to. The adapter file itself is yours — you get the weights, you keep the weights, and you can run inference anywhere that supports SDXL or Flux adapters.
This matters more for some use cases than others. A product shot of a consumer gadget is low-stakes. Reference photos of a human face — especially a founder, a public figure, or anyone whose biometric data has personal or commercial value — is a different category. We treat it accordingly.
What training input actually looks like
The quality of the LoRA is bounded by the quality and variety of the reference set. A common mistake is submitting 20 near-identical selfies at the same focal length, same background, same expression. The model memorizes the setting, not the subject.
Good reference sets share a few properties:
- Varied backgrounds and lighting conditions — the model needs to isolate the subject from context.
- Multiple angles — for faces: front, three-quarter, profile. For products: all faces the camera will ever show.
- Clean focus on the subject — the subject should occupy most of the frame, without heavy compression artifacts or motion blur.
- 15-30 images is a practical range. More is not always better; duplicate information trains faster but generalizes worse.
We review the reference set before training starts and flag anything likely to degrade the output. It saves a training run.
What the pipeline looks like end-to-end
The rough sequence for a founder headshot batch:
- 1 Reference intake. You send us the photo set via a secure upload link. We check for quality and variety, request reshoots if anything is underrepresented.
- 2 Training run. We tag, caption, and fine-tune the adapter. Typical turnaround is same-day for a standard batch.
- 3 Prompt batch. We generate the agreed scene list — typically 30-60 candidates across the requested contexts.
- 4 Human QA pass. Every output is reviewed before delivery. Anatomy errors, identity drift, and prompt-miss failures are filtered out — not passed on to you to sort through.
- 5 Delivery. Final images plus the adapter weights, so you own the model going forward.
Being honest about the limits
LoRA output is not photography. It does not carry the resolution ceiling of a 50-megapixel camera sensor. Fine detail — fabric texture, hair strands at the edge, reflections in eyes — degrades at high magnification in ways that are obvious to anyone who looks closely. For billboard print or cover magazine work, you still need a photographer.
Identity consistency is strong but not absolute. The model learned a distribution of your face, not a lookup table. Unusual prompts — extreme lighting, heavy stylization, occlusion by hands or objects — can introduce drift. This is why the human QA step exists. We reject the drift cases; we do not send them to you to filter.
For use cases where the output will appear in public-facing media, we recommend a final review pass by someone who knows the subject well. The model will not catch the one-in-forty frame where the nose is slightly wrong. A person who knows the subject will.
Within those limits, the use cases that actually work well are numerous: LinkedIn profile headers, website about-page portraits, ad creative A/B sets, social media batches, product catalog images, investor deck visuals. For those, the turnaround is fast, the cost is a fraction of a studio day, and the consistency across frames is better than you will get from three separate shoots.
The practical question: when does this make sense?
It makes sense when you need volume, consistency, or flexibility that a fixed set of photographs cannot provide. If you have ten good headshots and will reuse them forever, there is nothing to solve. If you are launching a new brand, onboarding a new spokesperson, or producing creative across six channels simultaneously, a trained adapter pays for itself in the first batch.
It also makes sense when the subject is a physical product with a short production timeline. Getting the photography right before manufacturing is finalized means reshoots. A product LoRA trained on prototype photos can generate marketing creative while the physical product is still being refined.