The decision, with the numbers

Stop treating it as a preference

People adopt one method and defend it. It is a decision with two inputs: how many images, and how closely will it be looked at.

The comparison

Image-prompt adapter (IP-Adapter and relatives)

• Setup: seconds. Load a reference, set a weight.
• Cost: nothing beyond generation.
• Result: a family resemblance. Bone structure and general appearance carry; small features drift between images.
• Flexibility: works with any prompt, though high weights start to compete with it.
• Requires: one image.

Trained LoRA

• Setup: about an hour end to end, once you have done it before.
• Cost: free on Kaggle or Colab; a few pennies of electricity locally.
• Result: identity holds tightly, including small features, across dozens of images.
• Flexibility: composes cleanly with prompts and controls at sensible strengths.
• Requires: 15 to 30 varied images, which for a fictional character means building the sheet first.

The decision rule

One to three images, seen separately: adapter. The hour of training buys nothing a viewer will register.

Four to twelve images, a campaign, side by side: LoRA. Adapter drift becomes visible when frames are compared, and a set is always compared.

A recurring character across months: LoRA, without question. It is an asset you keep, and the alternative is re-deriving the character every time.

Close scrutiny at large size: LoRA, plus face compositing for the frames that matter most.

You do not yet know if the character works: adapter first. Test the concept cheaply, then train once it is approved. Training a LoRA for a character the client then rejects is an hour spent on nothing.

Using both

The strongest arrangement for a demanding set is both together: the LoRA at 0.7 to 0.8 carrying the identity, and a face adapter at 0.3 to 0.4 from the single best reference nudging it toward one specific rendering.

The caution is the stacking budget from the previous block. A LoRA and an adapter are both constraints, and adding a pose control and a depth control on top leaves nothing free. If four seeds look identical, drop something.

What neither solves

Both fix who the person is. Neither fixes how they are lit, framed or graded, and those are what make a set look like a set. A perfectly consistent face across twelve differently-lit, differently-framed images still reads as twelve unrelated pictures.

That is the next lesson, and it is the part people skip. Identity consistency is the technically interesting problem, so it absorbs all the attention, while the visual system (one light, one lens, one palette, one grain) is what a viewer actually perceives as coherence and costs almost nothing to enforce.

Put differently: a set with a slightly drifting face and a rigorously consistent look reads as consistent. A set with a perfect face and twelve different lighting setups does not. If you have limited time, spend it on the system.

Testing which one you need, cheaply

You do not have to guess. The test takes fifteen minutes and settles the question with evidence rather than opinion.

Generate four frames with an adapter at 0.75: the four most different situations the brief contains, so different angles, different distances, different lighting. Lay them out side by side at the size they will be delivered.

Then ask one question: would a stranger say this is the same person? Not “can I see the differences”, because you can, having stared at them. A stranger, at delivery size, in the order the audience will see them.

If yes, ship the adapter and keep the hour. If no, train. If you are unsure, show it to somebody who has not been in the project: the answer is usually immediate and it is usually right.

Strength values worth starting from

For an adapter: 0.6 to 0.8 for a face. Below 0.5 the identity barely registers; above 0.9 it starts overriding the prompt and the pose.

For a LoRA: 0.6 to 0.85, depending on how heavily trained it is. A LoRA that needs 1.0 to show up is undertrained; one that dominates at 0.5 is overtrained, and the middle epoch you did not pick is probably the right one.

Write the working value next to the LoRA’s filename, because you will not remember it in a month and rediscovering it costs a dozen renders.