You Can't Caption Your Way Out of a Bad Dataset #ai #tech

The standard advice is "caption what you want to be able to change." I did exactly that — every caption in my dataset described the background, word for word — and it still got memorised. Captioning doesn't create separability. Variance does. If a thing is in 100% of your examples, the model never sees a counter-example, so the words and the pixels just fuse together. 28 clips, one background, one camera, one framing. Full breakdown: https://youtu.be/w_7EpghF8fw #loracledesmiroirs #AITraining #StableDiffusion #MachineLearning #DatasetDesign