Character consistency across shots, when the model has no memory between them

Generate the same described person five times and you get five siblings. Close enough that you notice the resemblance, different enough that cutting between them looks like a continuity error — because it is one.

This is not a limitation you prompt your way out of. No description is specific enough to reconstruct a face. Add more detail and you narrow the range slightly while making the prompt brittle; the fundamental problem is that identity is not something language encodes precisely.

Consistency therefore comes from carrying something visual between generations, and the whole craft is in choosing what to carry and how much of the frame it should control.

When you attach a reference, you are implicitly asking for one of these:

• Identity — this specific face and body, in a new situation.
• Style — this treatment, light and palette, applied to different content.
• Composition — this framing and arrangement, with different subjects.

Using one image for all three is the most common mistake and it produces the characteristic failure: a shot that is nearly right in every dimension and correct in none. The face drifts because the model was also trying to match the lighting; the composition changes because it was also trying to match the face.

Where the tool allows separate reference slots, use them separately. Where it does not, decide which single job matters most for this shot and let the other two be handled by the prompt — which is much better at style and composition than it is at identity.

Produce one image of the character that you are genuinely happy with. Frontal or three-quarter, evenly lit, neutral expression, nothing dramatic. This is your master.

Every subsequent shot is an edit or an image-to-video starting from that master, not a fresh generation. You are changing the situation around a fixed person rather than re-describing the person each time.

The limitation is real: you are constrained to angles and expressions reachable from the master. A profile shot from a frontal master will invent the far side of the face. Plan the sequence around what the master supports, or produce two or three masters at different angles and treat each as the source for the shots it covers.

If you need a character across dozens of shots, over weeks, in situations a single master cannot cover, training an identity is the answer. You supply a set of images and get back something you can call up repeatedly.

The dataset decides everything, and the usual instinct — supply the best photographs — is wrong. What you want is variety:

• Multiple angles, including profile and three-quarter from both sides.
• Multiple lighting conditions, including unflattering ones.
• Neutral expressions predominantly, with a few others.
• Consistent identity, varying everything else.

Twenty varied images beat sixty similar ones. If every training image is the same flattering three-quarter under the same soft key, you have trained a pose rather than a person, and it will refuse to turn its head.

Training also has a cost the single-master method does not: time before you can start, and a commitment to a look you may want to change. Use it when the sequence is long. For a six-shot job it is rarely worth it.

Whatever method you use, write the character's description once and paste it identically into every prompt. Not a paraphrase. The same words.

People rewrite the description each shot without noticing — 'dark hair' becomes 'black hair' becomes 'short dark hair' — and each variation moves the result slightly. Across ten shots those small movements compound into a visibly different person.

# character block, pasted verbatim into every prompt
A woman in her early thirties, shoulder-length dark brown hair
worn loose, matte black jacket, no jewellery, neutral expression.

Keep this block in a file next to the master image. Together they are the character, and treating them as a single unit is what makes a sequence hold together.

The most effective technique is not technical. It is choosing shots that do not stress the weakness.

Wide shots hold better than close-ups, because there is less face to get wrong. Shots where the subject is turned away, backlit, or partially obscured hold almost perfectly. A sequence that opens wide, moves to a three-quarter medium, and cuts away before the close-up is both a normal edit pattern and a technically robust one.

This is the same discipline as shooting with a real constraint — a location you cannot light, an actor available for one day. You design the sequence around it, and the audience never knows the constraint existed.

Fighting the limitation head-on with ten close-ups of the same face is where the work goes wrong. Designing around it is where it starts looking professional.