Suhail Doshi: prediction

Transcript context

…On the 1K Prompt Benchmark, you are 2.5x prefer to Stable Diffusion XL. How do you get there? Is it better images in the training corpus? Can you maybe talk through the improvements in the model? I think they're still very early on in the recipe, but I think it's a lot of like little things and you know, every now and then there are some big important things like certainly your data quality is really, really important. So we spend a lot of time thinking about that. But I would say it's a lot of things that you kind of clean up along the way as you train your model. Everything from captions to the data that you align with after pre-train to how you're picking your data sets, how you filter your data sets. I feel like there's a lot of work in AI that doesn't really feel like AI. It just really feels like just data set filtering and systems engineering and just like, you know, and the recipe is all there, but it's like a lot of extra work to do that. I think we plan to do a Playground V 2. 1, maybe either by the end of the year or early next year. And we're just like watching what the community does with the model. And then we're just going to take a lot of the things that they're unhappy about and just like fix them. You know, so for example, like maybe the eyes of people in an image don't feel right. They feel like they're a little misshapen or they're kind of blurry feeling. That's something that we already know we want to fix. So I think in that case, it's going to be about data quality. Or maybe you want to improve the kind of the dynamic range of color. You know, we want to make sure that that's like got a good range in any image. So what technique can we use there? There's different things like offset noise, pyramid noise, terminal zero, SNR, like there are all these various interesting things that you can do. So I think it's like a lot of just like tricks. Some are tricks, some are data, and some is just like cleaning. Specifically for faces, it's very common to use a pipeline rather than just train the base model more. Do you have a strong belief either way on like, oh, they should be separated out to different stages for like improving the eyes, improving the face or enhance or whatever? Or do you think like it can all be done in one model?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.