Orbit, Learned
Twelve drawings of an original robot taught to an image model: the same prompts before and after training.
A subject LoRA teaches an image model one specific character or object so it can be placed in new scenes. This method freezes and checks the dataset, trains with a selected native trainer, keeps every checkpoint, and judges them against the untrained base model on the same prompts and seeds rather than on the loss curve. Orbit is an original robot drawn from simple shapes: teal rounded head and body, one amber eye, a short antenna and an upward coral triangle.
To make your own, supply 10 to 20 consistent images of your subject with short captions and pick a trigger word. The 300- and 1,000-step checkpoints still drew generic robots; only the 3,000-step checkpoint held every defining feature. Details can still drift in new scenes: in a later reuse with a watering can, the triangle turned into a circle and a correction gave the eye a triangular pupil. The download contains the full dataset, all six base and trained comparisons, the training and evaluation scripts, the watering-can reuse images and the trial write-up.
