Made with Wavemaker on September 18, 2026, during an internal subject-training trial. Dataset: twelve original procedural 512px drawings of Orbit (examples/training/subject-trial/make_dataset.py), captions with trigger 'wmorbit'. Native Z-Image Turbo LoRA training (rank 8, bf16, AdamW) on one owned H100: 300 steps, a deliberately interrupted run, recovery to 1,000 steps, then +2,000 steps at a higher learning rate; the 3,000-step checkpoint was selected (sha256 649dbf3e...). Images: Z-Image Turbo at 512x512 with and without the selected LoRA, same prompts and seeds. The sheet is a layout of those unaltered images with text labels added. Review notes are in the source download.

Wavemaker-authored concept example. Source is study material; model weights and third-party tools are not included. Retain their original licenses. Importing source never starts generation.
