Read this before you train a LoRA: how far one reference image gets you
"Train a LoRA if you want the same character every time" was correct until recently — back when attaching a reference simply did not produce a likeness. That premise has changed, so the decision is worth redoing.
What changed
Measured, GPT Image 2 preserved a person's features from a single reference image — including mole placement — across changes of angle, expression, and outfit. No training; just attach the reference on each run.
| Attach a reference | Train a LoRA | |
|---|---|---|
| Setup | Choose one image | Collect and prepare a training set |
| Upfront cost | None | Training spend plus your time |
| Per run | Generation cost only (~$0.019 up) | Generation cost plus amortised training |
| Changing subject | Swap the reference | Retrain |
| Switching models | Carries over | Retrain |
I assumed a LoRA was simply required for a recurring character.
Some models now hold a face from one reference. Try that before paying for training.
The decision turns on how many subjects, for how long
One character, produced heavily over a long period, amortises training cost. Where the subject changes often, training restarts each time.
In practice the switching cost matters more. Image models turn over quickly, and today's best is not guaranteed to be best in six months. A trained asset is discarded when you move; a reference image is not.
Switch models and the training is gone. A reference image carries over.
Where a reference is not enough
- Subjects with no existing image — a character you are still designing
- Fixing a style that does not exist anywhere yet
- High-volume batch work at fixed quality with no room to attach references
The first is the common one. Anything that already exists as a photo or illustration can be referenced; an appearance you have not decided yet cannot be. There, settling one image first and using it as the reference from then on works well.
Getting more out of a reference
- Choose an image with the face clearly visible and low contrast between light and shadow
- One good reference beats several mediocre ones
- Describe only what changes; never describe the face
- Pick a model that can hold identity — on one that cannot, nothing else helps
The last point decides the rest. Seedream 5.0 Lite returned a different person even with a reference attached. That is the model, not the prompt, so model selection comes first when identity matters.
Frequently asked
- Do I need to train a LoRA?
- Not for identity alone — some models hold a person's features from one reference. Check whether a reference suffices before paying training cost.
- Which is cheaper, references or a LoRA?
- References carry no upfront cost, only generation. LoRA adds training cost and has to be redone when you change models. Training can win for one subject produced heavily over a long period.
- What about a character I haven't designed yet?
- There is no reference to attach, so settle one image first and reuse it as the reference from then on.
