Evaluating Diffusion Models with PickScore
Evaluating Diffusion Models with PickScore
Setting the scene for some future videos where I’ll explore ways to improve diffusion models through various tricks. Here we learn why evaluating diffusion models is hard, that user preference is the gold standard, and that preference models like PickScore give us an approximation we can work with.Code on github - https://github.com/johnowhitaker/dm/funPickScore on GitHub - https://github.com/yuvalkirstain/PickScore and paper - https://arxiv.org/pdf/2305.01569.pdfDall-e-3 announcement (https://openai.com/dall-e-3) and paper (https://cdn.openai.com/papers/dall-e-3.pdf)EMU paper - https://scontent.fhio2-1.fna.fbcdn.net/v/t39.2365-6/10000000/210375442059452/321995130515097708/n.pdf/nc/cat=105&ccb=1-7&/nc/sid=3c67a6&/nc/ohc=i-w75ldGqcgAX-dER4g&/nc/ht=scontent.fhio2-1.fna&oh=00/AfBip5IVsto1QTZ1O7JFVUcR51I9a/XXuiU3hY9OdKOlJA&oe=653C28CDWurstchen - https://arxiv.org/pdf/2306.00637.pdfSDXL paper - https://arxiv.org/pdf/2307.01952.pdf