How to keep a recurring AI model while changing pose and context

Define what should remain and what should change

For a collection campaign, you may want the same face and hairstyle while changing the item being held and the background. Write those as separate decisions. A request to “use the same model” is less clear when the pose, lighting, hair and camera angle all change drastically at once.

Identity should not be inferred from a matching outfit alone. Equally, an altered pose does not necessarily mean the reference was ignored; inspect the face and other stable features.

Start with a modest change

1. Save and select one approved person reference.
2. Create a simple product scene with clear framing.
3. Compare the person and product with their separate references.
4. Change one pose or scene decision for the next image.
5. Record which changes caused identity or anatomy problems.

If the result becomes inconsistent after a dramatic profile angle, use a less extreme view or provide a clearer reference. If hair or styling obscures the face, simplify the requirement. More descriptive adjectives will not necessarily fix weak visual evidence.

Check product interaction, not just the face

Hands, straps, clothing contact and product scale can determine whether a scene is usable. A recurring face with an impossible grip is still a rejected image. Make the action simple enough that the item remains identifiable and anatomically plausible.

When people are unnecessary for a module, disable model use instead of carrying a reference through every request. In an A+ sequence, a person may serve the use scene while a product detail module works better without them. This keeps the narrative coherent without making every image a portrait.

Treat video consistency as a separate review

For the current video workflow, approve a product-and-model still image and use it as the first frame. Then inspect the whole clip for changing facial features, hands and product components. A good starting frame does not establish stable identity throughout motion.

Keep approved results and notes in your normal asset workflow. The saved reference is a reusable input, while each generated scene is a separate output requiring approval. This distinction prevents a successful single example from being mistaken for a guarantee across a complete campaign.