Pieter Levels

So I had this photo guy and a lot of people using it. There was like a million or more photos a month being generated. And I discovered I was testing parameters, like increase the step count of generating a photo or changing the sampler, like a scheduler.

So I had this photo guy and a lot of people using it. There was like a million or more photos a month being generated. And I discovered I was testing parameters, like increase the step count of generating a photo or changing the sampler, like a scheduler.

So I had this photo guy and a lot of people using it. There was like a million or more photos a month being generated. And I discovered I was testing parameters, like increase the step count of generating a photo or changing the sampler, like a scheduler.

Like you have DPM, two caras, all these things I don't know anything about, but I know that you can choose them when you generate an image and they have different resulting images. But I didn't know which ones were better. So I would do it myself, test it. But then I was like, why don't I test on these users? Because I have a million photos generated anyway.

Like you have DPM, two caras, all these things I don't know anything about, but I know that you can choose them when you generate an image and they have different resulting images. But I didn't know which ones were better. So I would do it myself, test it. But then I was like, why don't I test on these users? Because I have a million photos generated anyway.

Like you have DPM, two caras, all these things I don't know anything about, but I know that you can choose them when you generate an image and they have different resulting images. But I didn't know which ones were better. So I would do it myself, test it. But then I was like, why don't I test on these users? Because I have a million photos generated anyway.

So on like 10% of the users, I would randomly test parameters. And then I would see if they would, because you can favor the photo or you can download it. I would measure if they favor it or like the photo. And then with A-B tests and you test for significance and stuff, which parameters were better and which were worse.

So on like 10% of the users, I would randomly test parameters. And then I would see if they would, because you can favor the photo or you can download it. I would measure if they favor it or like the photo. And then with A-B tests and you test for significance and stuff, which parameters were better and which were worse.

So on like 10% of the users, I would randomly test parameters. And then I would see if they would, because you can favor the photo or you can download it. I would measure if they favor it or like the photo. And then with A-B tests and you test for significance and stuff, which parameters were better and which were worse.

And then if it's significant enough data, you switch to that for all the users. And so that was like the breakthrough to make it better. Just use the users to improve themselves. And I tell them when they sign up, we do sampling, we do testing. on your photos with random parameters. And that worked really well.

And then if it's significant enough data, you switch to that for all the users. And so that was like the breakthrough to make it better. Just use the users to improve themselves. And I tell them when they sign up, we do sampling, we do testing. on your photos with random parameters. And that worked really well.

And then if it's significant enough data, you switch to that for all the users. And so that was like the breakthrough to make it better. Just use the users to improve themselves. And I tell them when they sign up, we do sampling, we do testing. on your photos with random parameters. And that worked really well.

I don't do a lot of testing anymore because it's like, I kind of reached like a diminishing point where it's like, it's kind of good. But there was a breakthrough, yeah.

I don't do a lot of testing anymore because it's like, I kind of reached like a diminishing point where it's like, it's kind of good. But there was a breakthrough, yeah.

I don't do a lot of testing anymore because it's like, I kind of reached like a diminishing point where it's like, it's kind of good. But there was a breakthrough, yeah.

Yeah. But actually, so like Stable Diffusion, I used 1.5, 2.0 came out, Stable Diffusion XL came out, all these new versions, and they were all worse. And so the core scene of people are still using 1.5 because it's also not what you call neutered. They neutered to make it super with safety features and stuff. So most of the people are still on Stable Diffusion 1.5.

Yeah. But actually, so like Stable Diffusion, I used 1.5, 2.0 came out, Stable Diffusion XL came out, all these new versions, and they were all worse. And so the core scene of people are still using 1.5 because it's also not what you call neutered. They neutered to make it super with safety features and stuff. So most of the people are still on Stable Diffusion 1.5.

Yeah. But actually, so like Stable Diffusion, I used 1.5, 2.0 came out, Stable Diffusion XL came out, all these new versions, and they were all worse. And so the core scene of people are still using 1.5 because it's also not what you call neutered. They neutered to make it super with safety features and stuff. So most of the people are still on Stable Diffusion 1.5.

And meanwhile, Stable Diffusion, the company went like... the CEO left, a lot of drama happened because they couldn't make money. And yeah, so they gave us, it's very interesting. They gave us this, this open source model that everybody uses. They raised like hundreds of millions of dollars. It all, they didn't make any money with it or not a lot. And they did an amazing job.

And meanwhile, Stable Diffusion, the company went like... the CEO left, a lot of drama happened because they couldn't make money. And yeah, so they gave us, it's very interesting. They gave us this, this open source model that everybody uses. They raised like hundreds of millions of dollars. It all, they didn't make any money with it or not a lot. And they did an amazing job.