The same face twice

Why this is genuinely hard

Every generation starts from noise and is steered by conditioning. There is no persistent object anywhere that means "this person". A face that appeared in one image exists only as pixels in that image.

So consistency is always reconstructed, never retrieved, and every method is an approximation. Anyone who tells you a tool has solved it is describing an approximation they have not stressed.

The demand is real, though. Comics, children's books, explainer videos, brand mascots, storyboards, product ranges — all of them need the same thing to appear repeatedly and be recognisably itself.

The methods, honestly rated

Fixed seed and identical prompt. Free, instant, and only works while nothing else changes. Change the pose or the setting and the face changes with it. Useful for variations on one shot, useless for a sequence.

Describe the character in detail and reuse the description. A very specific description — age, face shape, hair, eye colour, a distinctive feature — narrows the distribution enough that the outputs are cousins. They are not the same person. Adequate for background characters, not for a protagonist a reader will look at forty times.

A trained LoRA. The strongest general method. Twenty to thirty images of the character in varied settings, trained as in the previous lesson. The problem is circular for an invented character: you need images of someone who does not exist. The standard route out is to generate a set, fix the best ones by hand until they genuinely match, train on those, then regenerate a better set with the LoRA and retrain. Two rounds usually suffice.

Face-swapping and identity adapters. A method that extracts a face embedding from one image and injects it during generation. Fast, no training, and it holds the face while leaving everything else free. The limitations are specific: it carries the face and not the body, hair or clothing; quality drops at unusual angles; and it makes generating a convincing image of a real person from one photograph very easy, which is why several services restrict it.

Reference-image conditioning in hosted tools. Convenient, closed, and version-dependent — it can change under you between releases, which is the reproducibility problem again.

Draw or photograph the character once and use structural conditioning. For sequential work this is often the fastest honest route. A puppet built in Blender, or a real photographed model, gives you exact pose and identity, and the generative pass supplies the rendering.

What professionals actually do

The workflow used in production is a combination, and it is worth stating plainly because it removes the expectation of a single button:

1. Establish the character with a LoRA or a strong reference set.
2. Generate each shot with the character's conditioning applied.
3. Fix the face in every shot with a masked detail pass.
4. Colour-match all the shots to each other afterwards in a free editor, which does more for perceived consistency than most people expect.

Step 4 matters more than it sounds. A large part of what reads as "the same character" is consistent colour and light, not exact geometry. Two images with matched grade and different face proportions read as more consistent than the reverse.

What still does not work

Clothing and props remain the weakest link. Faces are now reasonably solvable; a specific jacket with a specific number of buttons, across twelve shots, is not. The reliable answer is to make the costume simple and distinctive — one strong silhouette, one colour, one accessory — which is exactly what character designers have always done for animation, and for the same reason: consistency is cheaper when there is less to be consistent about.

Hands holding specific objects, characters interacting, and any shot where two of your consistent characters appear together are the remaining hard cases. Budget for hand fixing, and expect to composite two single-character generations for the two-shot rather than to generate it.

The honest summary: consistency is achievable to a standard that survives a reader's attention, with effort, using several methods together. It is not achievable by prompting, and it is not yet a solved feature in any tool regardless of what the marketing page says.