Serving trays so tied up girls can deliver snacks. After a lot of fussing around with controlnets and inpainting to get a training set of 86 images with reasonably good perspective on every single tray, V3 is done. It often generates trays that make physical sense on the first try now. I also tagged the data more carefully, so now you can prompt for 'tray at waist', 'tray under breasts', 'small tray', 'large tray', 'straps to tray' and 'chains to tray' for better control of exactly how and where the tray is attached. It's also tagged for a few different kinds of restraints/gags (armbinder, rope, cuffs, ball gag, ring gag, bit gag). Finally, there's some data/tagging for various poses and viewing angles, so you can get standing/walking/kneeling and viewed from side/behind/above. Most of the training data was 1280 resolution, 2:3 semi tall aspect ratio.

Serving Tray

The trays, straps and chains follow perspective much better now.

Files with SHA256: 24fe8fffe20bd91fb0c2e1c580469eff6f1dcc16d7e9ddd130d1a59eacbd8501