Shooting modest fashion with AI, and the defaults you have to override

Ask a generator for a fashion image and it produces something drawn from what fashion photography mostly looks like in its training data. For coverage-led clothing that default is actively wrong: sleeves shorten, necklines drop, layers thin out, a headscarf loosens or disappears between frames.

None of this is malicious and all of it is persistent. It reasserts itself across regenerations, which means it cannot be solved by re-rolling — it has to be constrained explicitly and checked every time.

The good news is that the constraints are learnable and stable. Once written properly they hold across a set, and the category becomes as reliable as any other.

Category names are interpreted loosely. Physical descriptions are not. Say where each edge of the garment sits relative to the body:

Sleeves reach the wrist, fully covering the forearm.
Neckline sits at the collarbone, no opening below it.
Hemline below the knee. Fabric is opaque, not sheer.
Headscarf covers the hair completely, wrapped at the neck,
no hair visible at the forehead or temples.

Every clause names a boundary. That specificity is what survives the model's pull towards its default, and it is what makes the frames consistent with each other rather than merely acceptable individually.

Generators tend towards fabric that is thin, glossy and close to the body, because that is what a great deal of fashion photography shows. For coverage-led garments this reads wrong even when the shape is technically correct.

State the material behaviour explicitly: opaque, matte, and a description of how it hangs — 'falls loosely from the shoulder', 'structured, holds its shape', 'heavy drape'. These are the terms that change how the cloth is rendered.

Also name what the fabric does at the edges: a cuff, a hem, a seam. Edges are where the model improvises, and describing them keeps the garment coherent between frames.

Certain shots repeatedly break coverage — arms raised, twisting movement, low angles, anything that stretches the garment. Rather than fighting these, design around them.

What works consistently: standing, walking, seated, three-quarter and frontal, arms relaxed or holding something. These are also the framings that show a garment's cut and drape most clearly, which is what the imagery is for.

Wider framings hold better than close ones. A full-length figure is less likely to drift than a tight upper-body crop, because there is proportionally less detail for the model to reinterpret.

Build a checklist from the constraint block and run it on every frame, in order, before assessing quality:

1. Sleeve length at both wrists.
2. Neckline position.
3. Hemline.
4. Headscarf coverage at forehead, temples, neck.
5. Fabric opacity, especially against light.

Check both sides. Asymmetric failures are common — one sleeve correct, one shortened — and they are easy to miss when the composition draws the eye to one side.

Backlit frames need particular attention, because that is where opacity fails and it is not visible in a thumbnail.

Coverage expectations vary considerably between brands and markets. What one client considers correct another will reject. This is not something to infer from their existing imagery.

Ask directly at the start, and write the answer into the brand file as a set of physical boundaries — the same clauses that go into the prompt. Then the standard is documented, checkable, and applied identically across every asset.

Clients in this category are used to explaining themselves repeatedly and to being misunderstood. Arriving with the right questions, and encoding the answers, does more for the relationship than any amount of craft.