4 ms·
This is a smart move and goes to show people move so quick on this. I wonder how long before these models overtake stock image websites, get used in games or vi
by Escapado 4y ago
This is a smart move and goes to show people move so quick on this. I wonder how long before these models overtake stock image websites, get used in games or video production, generate T-shirt prints and whatever other usecase people can come up with.
On that note: Iirc back when I learned about GANs, when they were still rather new there was a clear path towards improvements of them by making the model bigger and feed it _much_ more high quality training data. When I look at the outputs of stable diffusion or Dall-E there are still often visible artifacts and most prominently faces are often weird. Is there a clear path towards improvement here aswell or are we hitting a wall with the "just more" paradigm somewhen?
- dougabug 4y agoGANs have a key (learned) loss called a discriminator which tells it whether or not the generated output looks real or not. GANs took many years of research to get to the point where they could get to the point where they could generate realistic, full sized images. Progress in diffusion models has been much faster. But you can add additional guidance to diffusion models, for instance, (possibly pre-trained) discriminative models which detect if certain objects classes such as faces look abnormal or malformed. We’re still quite early in the evolution of this family of models.
- bee_rider 4y agoWhen my friends and I play D&D, usually the host will quickly sketch maps on the fly (just, like, a floorplan or something). This tool seems like 1-2 iterations off from being perfect for that sort of thing.
- in3d 4y agoSome Stable Diffusion repos are already using an additional specialized model that fixes faces.