Lore
Garbage 3D product mannequin
# Garbage 3D product mannequin
Turn a single product image into a crude, textured 3D model and use it only as a mannequin: it fixes the product's pose, camera angle and position, and contributes nothing to realism. The model will look terrible, and that is acceptable. The principle is in Split placement from appearance: a stand-in sets geometry, the image model sets looks.
Procedure (Weavy)
- Plug the product image into a 3D node. Rodin v2 or Meshy both work. Any generator that returns a texture is fine, because the texture is what lets you see which part of the label is facing the camera.
- In the 3D viewer, rotate the model, orbit the camera and position it until the part of the label or the angle you want is showing.
- Export that view. It becomes reference image A, see compositor-annotation-layers for labelling it and master-prompt-with-reference-image-roles for how the image model reads it.
- The final image model (Nano Banana) renders the realistic result from separate appearance references.
What it gets you
- Deterministic control of label side, tilt and camera angle, which prose prompts handle poorly.
- Several mannequins in one frame, see multi-mannequin-composition.
Known limits
- When the product and the target image have different aspect ratios, the product size is a guess.
- The mannequin carries no fine detail, so detail must come from the product image and nine-panel-context-sheet.
- Some outputs will be wrong, and the placement logic should be judged separately from scene plausibility.
Related
Texture requirement
Texture requirement
The mannequin's geometry quality and appearance do not matter, but the 3D asset must carry some texture, even a bad one. Without a texture the image model estimates the product's position in space poorly, and placement drifts.
- Author's wording: "It really doesn't matter that it's a garbage texture, but it needs to be there."
- Rodin v2 and Meshy are described as interchangeable. Either works as long as it outputs a textured model. The author says the generator "needs to be the one that can give you also a texture", so a geometry-only generator is not enough.
- The texture is for position and orientation cues only. Real surface detail comes from the separate detail reference (see master-prompt-with-reference-image-roles), and the split of jobs is described in Split placement from appearance: a stand-in sets geometry, the image model sets looks.
Aspect-ratio mismatch with the output
The layout image can be 1:1 while the output is 9:16. In that case the model has to guess the product's size and fill the extra pixels above and below. In the author's run the rotation still came out in line with the mannequin. Expect size and framing to be approximate, and expect rotation and pose to be reliable.