28 Clips. One Room. That Was the Whole Problem. #ai #tech

The dataset explains every failure in this experiment. 28 clips. Two minutes and forty-eight seconds of source material. All of it cut from a single video - one background, one camera, one framing. At 8000 steps that's 286 passes over the same few minutes when about 63 would have done. If a thing appears in 100% of your examples, the model never sees a counter-example. The words and the pixels fuse, and no amount of captioning separates them again. Full breakdown: https://youtu.be/w_7EpghF8fw #LoRA #DatasetDesign #AITraining #MachineLearning #StableDiffusion