The problem it solves
Consistency is the hardest requirement in AI image work. A prompt describing a character gets you a family resemblance, not the same person. Across eight shots that reads as eight actors. A seed holds one image, not one identity.
A LoRA moves the constraint into the model. You train on 20 or so images of the thing you want repeatable, and afterwards that subject is available as a concept the model knows, addressable by a trigger word and controllable by weight. From then on, identity is not something you fight for in every prompt.
That is why it matters well beyond hobby model sharing. A recurring protagonist, a product that has to look exactly like the real product, a brand illustration style across a hundred assets: those are consistency problems, and adapters are the practical answer.
How low-rank adaptation works, briefly
Fine-tuning a full model means updating billions of weights, which needs serious hardware and produces a file the size of the original. Low-rank adaptation exploits the fact that the adjustment needed to teach one concept is much simpler than the model itself. Instead of learning a full weight update, it learns two thin matrices whose product approximates it, and injects them into the attention layers.
The rank is how thin. Low rank (4 to 16) captures a style or a face efficiently. Higher rank (32 to 128) has capacity for more complex or multi-part concepts, at the cost of size and a greater tendency to overfit. This is also why the base model matters so much: an adapter is a delta, so it only makes sense on the checkpoint family it was trained against.
LoRA vs checkpoint, and the other options
| What it changes | Size | Best for | |
|---|---|---|---|
| Checkpoint | The whole model | 2 to 12 GB | General capability and aesthetic |
| LoRA | Attention weight deltas | 10 to 300 MB | One character, style or object |
| Textual embedding | One token's meaning | Under 1 MB | A narrow style cue |
| Reference image / IP adapter | Nothing, conditioning only | None | One-off likeness with no training |
The decision is mostly about reuse. If you need this subject in one image, a reference image is faster and costs no training time. If you need it in fifty images over three months, training is worth the afternoon.
Training in practice
Dataset first, hyperparameters second. Most failures are dataset failures. Aim for varied angles, distances, lighting and backgrounds, with the target the only constant. Twenty good images beat two hundred near-duplicates from the same shoot.
Caption for what should stay variable. Describe what changes between images (pose, clothing, setting) and use a single rare trigger word for what should not. Anything you leave uncaptioned tends to fuse into the concept, which is how a character LoRA ends up permanently wearing one jacket.
Watch the intermediate saves. Save every few hundred steps and compare. The best epoch is often not the last; it is the one just before flexibility collapses.
Then use it at the lowest weight that works. Full strength is a common reflex and usually wrong, because it drags composition and colour along with identity. Start at 0.7 and increase only if the likeness slips.
Where these adapters genuinely cannot help: motion quality, physics, and legible text. They teach appearance. If your problem is a warping camera move rather than an inconsistent face, this is the wrong tool.