Lore

Split placement from appearance: a stand-in sets geometry, the image model sets looks

# Split placement from appearance

Claim: realism is no longer the hard part of AI image generation. Control is. Typical control failures are showing a specific part of a product label, or deciding where two products touch, how close they sit and which one overlaps the other. Prose prompts are a poor way to state these spatial facts.

The split

Divide the job between two sources instead of asking the image model to do both:

| Question | Answered by | |---|---| | Where is the product, how is it oriented, where is the camera? | A deliberately crude stand-in, see garbage-3d-product-mannequin | | What does it look like (texture, label, material, light)? | The image model, fed separate detail references, see nine-panel-context-sheet |

The stand-in only has to be right about geometry. It does not have to look good, so it can be ugly.

Why it works

Limits

The split fixes pose, angle and position. It does not fix the quality of the surrounding scene. Output still depends on how much context the model gets (extra angles, close-ups) and on how much effort goes into the scene description. The logic of angle and placement can hold even when the scene itself is illogical, for example an implausible Antarctica setting.

Related