Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai

>> Okay. So, >> Okay. So, I am Sangwha. I'm from Korea. I'm going I am Sangwha. I'm from Korea. I'm going I am Sangwha. I'm from Korea. I'm going to be talking a little bit about how we to be talking a little bit about how we to be talking a little bit about how we recently trained our like image recently trained our like image recently trained our like image foundation model Create 2, as well as foundation model Create 2, as well as foundation model Create 2, as well as and we also recently like open-sourced and we also recently like open-sourced and we also recently like open-sourced the medium version of our model. So, the medium version of our model. So, the medium version of our model. So, I'll be like talking about like how we I'll be like talking about like how we I'll be like talking about like how we like trained that mostly from research like trained that mostly from research like trained that mostly from research perspective. Later in the day, uh perspective. Later in the day, uh perspective. Later in the day, uh like my colleague will also give a like my colleague will also give a like my colleague will also give a little bit of little bit of little bit of details on like infrastructure and details on like infrastructure and details on like infrastructure and training infrastructure and like training infrastructure and like training infrastructure and like everything that needed to be like set up everything that needed to be like set up everything that needed to be like set up for that. But, today I'll mostly be for that. But, today I'll mostly be for that. But, today I'll mostly be talking about like research. Also, talking about like research. Also, talking about like research. Also, public speaking is not one of my best public speaking is not one of my best public speaking is not one of my best abilities. So, take that in mind. But, abilities. So, take that in mind. But, abilities. So, take that in mind. But, anyways, I'll go ahead and start. anyways, I'll go ahead and start. anyways, I'll go ahead and start. So, we recent As I said, we recently So, we recent As I said, we recently So, we recent As I said, we recently open-sourced our open-sourced our open-sourced our Create 2, Create 2 medium variant of our Create 2, Create 2 medium variant of our Create 2, Create 2 medium variant of our like closed-source model. Now, it's like like closed-source model. Now, it's like like closed-source model. Now, it's like open-source and I think people are open-source and I think people are open-source and I think people are enjoying quite a bit. One thing we enjoying quite a bit. One thing we enjoying quite a bit. One thing we focused quite a bit was like stylistic focused quite a bit was like stylistic focused quite a bit was like stylistic diversity.

diversity. diversity. So, how we are going to like go about So, how we are going to like go about So, how we are going to like go about this is that I'm going to like describe this is that I'm going to like describe this is that I'm going to like describe a little bit of the training like a little bit of the training like a little bit of the training like pipeline pipeline pipeline that we have like used to like curate that we have like used to like curate that we have like used to like curate data, curate data, actually like train data, curate data, actually like train data, curate data, actually like train the model, train the model, and then the model, train the model, and then the model, train the model, and then I'll like talk a little bit about like I'll like talk a little bit about like I'll like talk a little bit about like what were like the most effective levers what were like the most effective levers what were like the most effective levers for actually improving model for actually improving model for actually improving model performance. And then, performance. And then, performance. And then, the last thing would be, you know, like the last thing would be, you know, like the last thing would be, you know, like what I think would be our like kind of what I think would be our like kind of what I think would be our like kind of exciting directions for the next exciting directions for the next exciting directions for the next generation image models. So, I'll go generation image models. So, I'll go generation image models. So, I'll go ahead and then start talking about that. ahead and then start talking about that. ahead and then start talking about that. So, one of the trend that we've been So, one of the trend that we've been So, one of the trend that we've been like seeing recently is that, you know, like seeing recently is that, you know, like seeing recently is that, you know, existing models like ChatGPT, existing models like ChatGPT, existing models like ChatGPT, ChatGPT and Nano Banana like the ChatGPT and Nano Banana like the ChatGPT and Nano Banana like the like the production grade uh models that like the production grade uh models that like the production grade uh models that are like from the big labs. One thing is are like from the big labs. One thing is are like from the big labs. One thing is that they kind of like focus on like that they kind of like focus on like that they kind of like focus on like slower generation, but like very slower generation, but like very slower generation, but like very reliable output. So, I know Chat GPT-2 reliable output. So, I know Chat GPT-2 reliable output. So, I know Chat GPT-2 or like Nano Banana Pro, I think or like Nano Banana Pro, I think or like Nano Banana Pro, I think they take up to like a minute or like they take up to like a minute or like they take up to like a minute or like two to like give an output. Typically, two to like give an output. Typically, two to like give an output. Typically, it's like something very acceptable. it's like something very acceptable. it's like something very acceptable. There's like hardly barely any like There's like hardly barely any like There's like hardly barely any like flaws, but flaws, but flaws, but one problem is that in order to get like one problem is that in order to get like one problem is that in order to get like good consistency, they have like kind of good consistency, they have like kind of good consistency, they have like kind of like significantly mode collapse their like significantly mode collapse their like significantly mode collapse their models because I don't know, like if models because I don't know, like if models because I don't know, like if you're trying to render a person, the you're trying to render a person, the you're trying to render a person, the easiest and most reliable way to like easiest and most reliable way to like easiest and most reliable way to like render a person is render the most render a person is render the most render a person is render the most boring average person that exists and boring average person that exists and boring average person that exists and then like put it in a center frame. Uh then like put it in a center frame. Uh then like put it in a center frame. Uh And like this for example, like this is And like this for example, like this is And like this for example, like this is one of the examples that I like to use.

one of the examples that I like to use. one of the examples that I like to use. So, like if you type like burning skull, So, like if you type like burning skull, So, like if you type like burning skull, Chat GPT-2 like is very consistent. All Chat GPT-2 like is very consistent. All Chat GPT-2 like is very consistent. All the outputs are like fine, but you know, the outputs are like fine, but you know, the outputs are like fine, but you know, there's barely any diversity. there's barely any diversity. there's barely any diversity. On the other end, like kind of the thing On the other end, like kind of the thing On the other end, like kind of the thing that we kind of like focus on is like that we kind of like focus on is like that we kind of like focus on is like faster generation so that like people faster generation so that like people faster generation so that like people can like iterate with like different can like iterate with like different can like iterate with like different visual ideas and then also have like a visual ideas and then also have like a visual ideas and then also have like a little bit of knobs on kind of because, little bit of knobs on kind of because, little bit of knobs on kind of because, you know, when people like comment, you know, when people like comment, you know, when people like comment, sometimes like they want very something sometimes like they want very something sometimes like they want very something very specific. They want to generate very specific. They want to generate very specific. They want to generate like a poster or like a birthday card. like a poster or like a birthday card. like a poster or like a birthday card. In that case, like Chat GPT-2 or Nano In that case, like Chat GPT-2 or Nano In that case, like Chat GPT-2 or Nano Banana Pro like excellent solution, but Banana Pro like excellent solution, but Banana Pro like excellent solution, but sometimes like when you're like a sometimes like when you're like a sometimes like when you're like a creative studio, you don't quite know creative studio, you don't quite know creative studio, you don't quite know what you want yet and want to like what you want yet and want to like what you want yet and want to like slightly explore what kind of like slightly explore what kind of like slightly explore what kind of like visuals you want to make. That's visuals you want to make. That's visuals you want to make. That's something that we wanted to uh focus on something that we wanted to uh focus on something that we wanted to uh focus on a little bit. a little bit. a little bit. So, So, So, you know, you know, you know, with that in mind, like how do you with that in mind, like how do you with that in mind, like how do you actually like train a diffusion model? actually like train a diffusion model? actually like train a diffusion model? Uh Uh Uh I hope most of the people in the I hope most of the people in the I hope most of the people in the audience like know how diffusion models audience like know how diffusion models audience like know how diffusion models work, but just to give a small recap. work, but just to give a small recap. work, but just to give a small recap. How it works is that you have an image How it works is that you have an image How it works is that you have an image and you add a little bit of like noise and you add a little bit of like noise and you add a little bit of like noise and then you ask the model to like, and then you ask the model to like, and then you ask the model to like, "Okay, like how here is an image with "Okay, like how here is an image with "Okay, like how here is an image with some of the noise added in? Like how do some of the noise added in? Like how do some of the noise added in? Like how do you like remove this noise uh you like remove this noise uh you like remove this noise uh remove this noise to get like a valid remove this noise to get like a valid remove this noise to get like a valid image. So, like at a high level, this is image. So, like at a high level, this is image. So, like at a high level, this is how diffusion models are like trained.

how diffusion models are like trained. how diffusion models are like trained. And once you train a model to do this, And once you train a model to do this, And once you train a model to do this, you can just you can just you can just like get a random noise and then tell like get a random noise and then tell like get a random noise and then tell the model to like progressively remove the model to like progressively remove the model to like progressively remove the noise uh the noise uh the noise uh noise to generate like an image. So, at noise to generate like an image. So, at noise to generate like an image. So, at a high level, like this is how a a high level, like this is how a a high level, like this is how a diffusion models are like trained. But diffusion models are like trained. But diffusion models are like trained. But right now, most of the production uh right now, most of the production uh right now, most of the production uh production diffusion models, at least production diffusion models, at least production diffusion models, at least the open-source ones that are very the open-source ones that are very the open-source ones that are very competitive, including ours, use a competitive, including ours, use a competitive, including ours, use a autoencoder. So, this is typically autoencoder. So, this is typically autoencoder. So, this is typically called a latent diffusion model, where called a latent diffusion model, where called a latent diffusion model, where you use a autoencoder to like, instead you use a autoencoder to like, instead you use a autoencoder to like, instead of taking the raw pixels, you first like of taking the raw pixels, you first like of taking the raw pixels, you first like spatially compress it and then spatially compress it and then spatially compress it and then decompress it uh decompress it uh decompress it uh later. But the generation works in a later. But the generation works in a later. But the generation works in a spatially compressed uh latent space, so spatially compressed uh latent space, so spatially compressed uh latent space, so that is a little bit more efficient to that is a little bit more efficient to that is a little bit more efficient to train. One of the main reason is that train. One of the main reason is that train. One of the main reason is that most people use diffusion transformers most people use diffusion transformers most people use diffusion transformers uh transformers or like some variant of uh transformers or like some variant of uh transformers or like some variant of that to train diffusion models. And as that to train diffusion models. And as that to train diffusion models. And as you know, transformers, at least the you know, transformers, at least the you know, transformers, at least the non-sparse ones, uh non-sparse ones, uh non-sparse ones, uh they tend to have like all like all been they tend to have like all like all been they tend to have like all like all been squared uh squared uh squared uh all been squared uh time complexity. So, all been squared uh time complexity. So, all been squared uh time complexity. So, if you try to like model every single if you try to like model every single if you try to like model every single pixel, that's very expensive. So, now uh pixel, that's very expensive. So, now uh pixel, that's very expensive. So, now uh and that was the initial motivation for and that was the initial motivation for and that was the initial motivation for like why people kind of like started like why people kind of like started like why people kind of like started with like latent uh latent diffusion with like latent uh latent diffusion with like latent uh latent diffusion models, because now they can actually models, because now they can actually models, because now they can actually like model this a little bit in a more like model this a little bit in a more like model this a little bit in a more efficient space.

efficient space. efficient space. So, one thing that I'll say is that like So, one thing that I'll say is that like So, one thing that I'll say is that like really like data is like quite really like data is like quite really like data is like quite everything that goes into the model. everything that goes into the model. everything that goes into the model. Like typically, you lock in your Like typically, you lock in your Like typically, you lock in your architecture and then you just architecture and then you just architecture and then you just a lot of work just goes into like just a lot of work just goes into like just a lot of work just goes into like just feeding the model like what it wants. feeding the model like what it wants. feeding the model like what it wants. Like, I mean, this sounds stupid, but I Like, I mean, this sounds stupid, but I Like, I mean, this sounds stupid, but I cannot iterate this more. That's why I cannot iterate this more. That's why I cannot iterate this more. That's why I put like, again, like really data is put like, again, like really data is put like, again, like really data is quite like everything. Like quite like everything. Like quite like everything. Like again, like this this just comes up again, like this this just comes up again, like this this just comes up again and again. And typically, after again and again. And typically, after again and again. And typically, after you lock in your architecture, like most you lock in your architecture, like most you lock in your architecture, like most of the work actually goes into like data of the work actually goes into like data of the work actually goes into like data curation, like making sure that the data curation, like making sure that the data curation, like making sure that the data is good. And in our case, like we wanted is good. And in our case, like we wanted is good. And in our case, like we wanted to focus on like stylistic diversity, so to focus on like stylistic diversity, so to focus on like stylistic diversity, so making sure that like we don't making sure that like we don't making sure that like we don't necessarily like remove like necessarily like remove like necessarily like remove like remove like remove like remove like remove like data data unintentionally to remove like data data unintentionally to remove like data data unintentionally to like cut stylistic diversity. Like that like cut stylistic diversity. Like that like cut stylistic diversity. Like that was also very important focus for us. was also very important focus for us. was also very important focus for us. For instance, like some people like For instance, like some people like For instance, like some people like think I know like low-resolution CRT think I know like low-resolution CRT think I know like low-resolution CRT videos are like videos are like videos are like a bad image, but some people like that a bad image, but some people like that a bad image, but some people like that kind of like aesthetics, so making sure kind of like aesthetics, so making sure kind of like aesthetics, so making sure that we have like good coverage and that we have like good coverage and that we have like good coverage and don't just rely on like very standard don't just rely on like very standard don't just rely on like very standard like aesthetic scores or like image like aesthetic scores or like image like aesthetic scores or like image quality scores to like cut like quality scores to like cut like quality scores to like cut like oversample uh typically what are oversample uh typically what are oversample uh typically what are considered like conventionally good considered like conventionally good considered like conventionally good images was also something that we had to images was also something that we had to images was also something that we had to like take into account. So, with that in like take into account. So, with that in like take into account. So, with that in mind, I'm going to describe some of the mind, I'm going to describe some of the mind, I'm going to describe some of the things that I think we did uh things that I think we did uh things that I think we did uh quite cool quite cool quite cool when we came to like data and kind of when we came to like data and kind of when we came to like data and kind of like things that we like kind of like things that we like kind of like things that we like kind of consider bad data or like duplicated consider bad data or like duplicated consider bad data or like duplicated samples, overrepresented concepts. So, samples, overrepresented concepts. So, samples, overrepresented concepts. So, these are typically

these are typically these are typically taken uh taken uh taken uh care of by like deduplication and and uh care of by like deduplication and and uh care of by like deduplication and and uh clustering uh based uh rebalancing. And clustering uh based uh rebalancing. And clustering uh based uh rebalancing. And then, you know, there are certain then, you know, there are certain then, you know, there are certain samples that like visual language models samples that like visual language models samples that like visual language models we used to generate the captions, they we used to generate the captions, they we used to generate the captions, they sometimes constantly fail to like sometimes constantly fail to like sometimes constantly fail to like capture important aspect of the image, capture important aspect of the image, capture important aspect of the image, which leads to certain biases, which leads to certain biases, which leads to certain biases, uh which is also the third point. And uh which is also the third point. And uh which is also the third point. And then, like you know, sometimes when you then, like you know, sometimes when you then, like you know, sometimes when you train on like very low resolution, train on like very low resolution, train on like very low resolution, it doesn't make sense to like put train it doesn't make sense to like put train it doesn't make sense to like put train on an image that has like I don't know, on an image that has like I don't know, on an image that has like I don't know, like 20 100 like characters on a 256 by like 20 100 like characters on a 256 by like 20 100 like characters on a 256 by 256 pixels because that's just going to 256 pixels because that's just going to 256 pixels because that's just going to be like too hard for the model to like be like too hard for the model to like be like too hard for the model to like learn at least at like low resolution learn at least at like low resolution learn at least at like low resolution stages. So, that's one uh example. And stages. So, that's one uh example. And stages. So, that's one uh example. And then, obviously, AI images uh then, obviously, AI images uh then, obviously, AI images uh distillation is like a very big like distillation is like a very big like distillation is like a very big like topic, but we tried very hard to like topic, but we tried very hard to like topic, but we tried very hard to like remove any AI images like at all because remove any AI images like at all because remove any AI images like at all because it does like provide you a shortcut to it does like provide you a shortcut to it does like provide you a shortcut to to get you like a good model, but to get you like a good model, but to get you like a good model, but synthetic data is like so sticky to the synthetic data is like so sticky to the synthetic data is like so sticky to the model that once you like start training model that once you like start training model that once you like start training on AI image data, sure your model's on AI image data, sure your model's on AI image data, sure your model's good, but you kind of lose the point good, but you kind of lose the point good, but you kind of lose the point because then you're going to get like because then you're going to get like because then you're going to get like very like ChatGPT or like Nano banner very like ChatGPT or like Nano banner very like ChatGPT or like Nano banner aesthetic. And at least for me like I aesthetic. And at least for me like I aesthetic. And at least for me like I can tell when like a model has been very can tell when like a model has been very can tell when like a model has been very heavily trained or like distilled on heavily trained or like distilled on heavily trained or like distilled on like ChatGPT 2 or like Nano banner Pro.

like ChatGPT 2 or like Nano banner Pro. like ChatGPT 2 or like Nano banner Pro. And in the long run like that's just not And in the long run like that's just not And in the long run like that's just not how you want to like do like you don't I how you want to like do like you don't I how you want to like do like you don't I and you know, as a researcher it always and you know, as a researcher it always and you know, as a researcher it always slightly hurts my ego if all I'm doing slightly hurts my ego if all I'm doing slightly hurts my ego if all I'm doing is distillation. So, is distillation. So, is distillation. So, yeah. yeah. yeah. And then, you know, captions are very And then, you know, captions are very And then, you know, captions are very important. So, like just to briefly important. So, like just to briefly important. So, like just to briefly describe our like captioning pipeline, describe our like captioning pipeline, describe our like captioning pipeline, how we we take the images and then we how we we take the images and then we how we we take the images and then we run OCR because text rendering is quite run OCR because text rendering is quite run OCR because text rendering is quite important. Uh so, we make sure that we important. Uh so, we make sure that we important. Uh so, we make sure that we first extract out all the text that are first extract out all the text that are first extract out all the text that are like visible in the image. And then we like visible in the image. And then we like visible in the image. And then we also add like optional metadata if we also add like optional metadata if we also add like optional metadata if we know this is like a picture of a famous know this is like a picture of a famous know this is like a picture of a famous person, we make sure that that's person, we make sure that that's person, we make sure that that's included. And then we do a second pass included. And then we do a second pass included. And then we do a second pass with our vision language model to with our vision language model to with our vision language model to generate like very detailed caption for generate like very detailed caption for generate like very detailed caption for this image. And once we have like a this image. And once we have like a this image. And once we have like a caption that sufficiently captures all caption that sufficiently captures all caption that sufficiently captures all the things that are relevant about the the things that are relevant about the the things that are relevant about the image, then we can like rewrite it to image, then we can like rewrite it to image, then we can like rewrite it to like whatever like JSON prompts or like like whatever like JSON prompts or like like whatever like JSON prompts or like other formats you would like to like other formats you would like to like other formats you would like to like uh feed the model. So, like this our uh feed the model. So, like this our uh feed the model. So, like this our like captioning pipeline. like captioning pipeline. like captioning pipeline. And this is actually uh one good example And this is actually uh one good example And this is actually uh one good example of what I consider like bad data.

of what I consider like bad data. of what I consider like bad data. It doesn't look that bad, but one of the It doesn't look that bad, but one of the It doesn't look that bad, but one of the issues that we had with this kind of issues that we had with this kind of issues that we had with this kind of images is that we would like try many images is that we would like try many images is that we would like try many things with the captioner. Uh things with the captioner. Uh things with the captioner. Uh and it would say that oh, this is a and it would say that oh, this is a and it would say that oh, this is a painting of blah blah blah blah blah painting of blah blah blah blah blah painting of blah blah blah blah blah blah, but it would not mention the fact blah, but it would not mention the fact blah, but it would not mention the fact that it's framed on a wall that it's framed on a wall that it's framed on a wall with like a on and have a white with like a on and have a white with like a on and have a white background. And I know sometimes the background. And I know sometimes the background. And I know sometimes the captioner would like consistently like captioner would like consistently like captioner would like consistently like fail to mention this fact. So, when you fail to mention this fact. So, when you fail to mention this fact. So, when you try to generate a painting of whatever, try to generate a painting of whatever, try to generate a painting of whatever, it'll be always hanged on a wall, it'll be always hanged on a wall, it'll be always hanged on a wall, on a white wall, which is pulling that on a white wall, which is pulling that on a white wall, which is pulling that what the user wants. So, this is an what the user wants. So, this is an what the user wants. So, this is an example of like, you know, the image is example of like, you know, the image is example of like, you know, the image is fine. It's you can train on it, but if fine. It's you can train on it, but if fine. It's you can train on it, but if it's a kind of image that, you know, it's a kind of image that, you know, it's a kind of image that, you know, LLMs or vision language models cannot LLMs or vision language models cannot LLMs or vision language models cannot like consistently like capture important like consistently like capture important like consistently like capture important aspect of the model. Like, this is an aspect of the model. Like, this is an aspect of the model. Like, this is an example where we just had to like example where we just had to like example where we just had to like design filters and then threw this kind design filters and then threw this kind design filters and then threw this kind of like data out or at least undersample of like data out or at least undersample of like data out or at least undersample it. it. it. And, you know, deduplication, we mostly And, you know, deduplication, we mostly And, you know, deduplication, we mostly use like hash-based solutions. use like hash-based solutions. use like hash-based solutions. We when we like train these models, we We when we like train these models, we We when we like train these models, we need to use like like, I don't know, need to use like like, I don't know, need to use like like, I don't know, anywhere from like 2 to like 10 billion anywhere from like 2 to like 10 billion anywhere from like 2 to like 10 billion images. That's a lot of images to run images. That's a lot of images to run images. That's a lot of images to run like filters on. So, like, first thing like filters on. So, like, first thing like filters on. So, like, first thing we do is like just tip calculate like we do is like just tip calculate like we do is like just tip calculate like pHash or like MD5 hash to do basic pHash or like MD5 hash to do basic pHash or like MD5 hash to do basic deduplication. And then, once we get to deduplication. And then, once we get to deduplication. And then, once we get to like a smaller like size, that's when we like a smaller like size, that's when we like a smaller like size, that's when we bring in, you know, some of the bring in, you know, some of the bring in, you know, some of the embedding-based uh deduplication method, embedding-based uh deduplication method, embedding-based uh deduplication method, SSCD, like SigLip, to do uh SSCD, like SigLip, to do uh SSCD, like SigLip, to do uh set semantic de- duplication or like set semantic de- duplication or like set semantic de- duplication or like near du- uh remove near duplicates.

near du- uh remove near duplicates. near du- uh remove near duplicates. And like, typically how we like also And like, typically how we like also And like, typically how we like also like design filters that we use like like design filters that we use like like design filters that we use like large uh large uh large uh large vision language model. I know, for large vision language model. I know, for large vision language model. I know, for instance, you design a prompt to like instance, you design a prompt to like instance, you design a prompt to like get a large language model to like get a large language model to like get a large language model to like learn, does this look like a AI image or learn, does this look like a AI image or learn, does this look like a AI image or not? And then like, we can act after not? And then like, we can act after not? And then like, we can act after like we get a good like fine-tuned or a like we get a good like fine-tuned or a like we get a good like fine-tuned or a system from from a large uh vision system from from a large uh vision system from from a large uh vision language model, language model, language model, one thing that uh we can do is we can one thing that uh we can do is we can one thing that uh we can do is we can like distill this data, like distill this data, like distill this data, this kind of like decision and like to this kind of like decision and like to this kind of like decision and like to like a very small like SigLip like a very small like SigLip like a very small like SigLip classifier. And then, you can base you classifier. And then, you can base you classifier. And then, you can base you have like a very like cheap classifier have like a very like cheap classifier have like a very like cheap classifier that is somewhat like reliable. And, you that is somewhat like reliable. And, you that is somewhat like reliable. And, you know, like typically if you run a know, like typically if you run a know, like typically if you run a classifier over like a billion images, classifier over like a billion images, classifier over like a billion images, you do need things to be like you do need things to be like you do need things to be like SigLip-sized, SigLip-sized, SigLip-sized, uh for instance. So, this is one of And uh for instance. So, this is one of And uh for instance. So, this is one of And also like, this is if you guys are into also like, this is if you guys are into also like, this is if you guys are into like LLM literature, like this is one of like LLM literature, like this is one of like LLM literature, like this is one of the approaches that uh the approaches that uh the approaches that uh the essential web data is used. They the essential web data is used. They the essential web data is used. They also used a big LLM also used a big LLM also used a big LLM big LLM to like come up with like some big LLM to like come up with like some big LLM to like come up with like some kind of taxonomy or like classifiers to kind of taxonomy or like classifiers to kind of taxonomy or like classifiers to like judge whether this text data is like judge whether this text data is like judge whether this text data is like good or not in this and then like good or not in this and then like good or not in this and then distill that down to like I don't know distill that down to like I don't know distill that down to like I don't know like 500 uh million parameter model so like 500 uh million parameter model so like 500 uh million parameter model so that you can actually run this over like that you can actually run this over like that you can actually run this over like a pre-training level corpus. Otherwise, a pre-training level corpus. Otherwise, a pre-training level corpus. Otherwise, it'll be somewhat expensive and frankly it'll be somewhat expensive and frankly it'll be somewhat expensive and frankly inefficient use of your GPUs to do so.

inefficient use of your GPUs to do so. inefficient use of your GPUs to do so. And another thing which I'm a little bit And another thing which I'm a little bit And another thing which I'm a little bit proud of uh is that we actually use like proud of uh is that we actually use like proud of uh is that we actually use like sparse autoencoders uh for some of our sparse autoencoders uh for some of our sparse autoencoders uh for some of our filtering. Uh filtering. Uh filtering. Uh I don't know how many of you guys are I don't know how many of you guys are I don't know how many of you guys are still like into like sparse still like into like sparse still like into like sparse autoencoders, but autoencoders, but autoencoders, but I we did some work or at least I did I we did some work or at least I did I we did some work or at least I did some work in doing uh sparse autoencoder some work in doing uh sparse autoencoder some work in doing uh sparse autoencoder research on like clip or these kind of research on like clip or these kind of research on like clip or these kind of like image models uh vision models. And like image models uh vision models. And like image models uh vision models. And one thing that you can actually get out one thing that you can actually get out one thing that you can actually get out of SAE is a unsupervised tagging system. of SAE is a unsupervised tagging system. of SAE is a unsupervised tagging system. So, one thing you can do is that once So, one thing you can do is that once So, one thing you can do is that once you train a SAE on a vision model, what you train a SAE on a vision model, what you train a SAE on a vision model, what you can do is that you can feed an image you can do is that you can feed an image you can do is that you can feed an image and then it will give you like sparse and then it will give you like sparse and then it will give you like sparse features that get activated. For features that get activated. For features that get activated. For instance, instance, instance, let's say you feed this image to a let's say you feed this image to a let's say you feed this image to a sparse auto vision sparse autoencoder sparse auto vision sparse autoencoder sparse auto vision sparse autoencoder and it will give you like it'll get and it will give you like it'll get and it will give you like it'll get activated for like features like horse, activated for like features like horse, activated for like features like horse, black and white, blur uh blurry like black and white, blur uh blurry like black and white, blur uh blurry like image image image blurry image. So, blurry image. So, blurry image. So, once So, you can kind of use this as once So, you can kind of use this as once So, you can kind of use this as like off-the-shelf like uh unsupervised like off-the-shelf like uh unsupervised like off-the-shelf like uh unsupervised tagging system. And if one of these like tagging system. And if one of these like tagging system. And if one of these like have like something that you want to have like something that you want to have like something that you want to like filter on or like filter on or like filter on or oversample like easy things or like oversample like easy things or like oversample like easy things or like signatures or like watermarks or like signatures or like watermarks or like signatures or like watermarks or like some kind of like border artifacts that some kind of like border artifacts that some kind of like border artifacts that I was like talking about. So, if you I was like talking about. So, if you I was like talking about. So, if you have like a feature for that in your have like a feature for that in your have like a feature for that in your sparse autoencoder, like this is one sparse autoencoder, like this is one sparse autoencoder, like this is one nice thing to remove uh to like remove nice thing to remove uh to like remove nice thing to remove uh to like remove like like like data that's kind of like undesirable uh data that's kind of like undesirable uh data that's kind of like undesirable uh in your data set.

in your data set. in your data set. And then, one thing that I think like And then, one thing that I think like And then, one thing that I think like people have quite liked about our uh people have quite liked about our uh people have quite liked about our uh open source create model is like world open source create model is like world open source create model is like world knowledge. knowledge. knowledge. Frankly, I don't know how much this has Frankly, I don't know how much this has Frankly, I don't know how much this has helped, but this is one of the things we helped, but this is one of the things we helped, but this is one of the things we did. did. did. Uh is that Uh is that Uh is that you know, this was something that like you know, this was something that like you know, this was something that like clip, the original clip paper did, is clip, the original clip paper did, is clip, the original clip paper did, is that you can actually take the that you can actually take the that you can actually take the Wikipedia, the entire Wikipedia, and for Wikipedia, the entire Wikipedia, and for Wikipedia, the entire Wikipedia, and for each article or concept, you can compute each article or concept, you can compute each article or concept, you can compute the page rank of each of the concepts, the page rank of each of the concepts, the page rank of each of the concepts, and then and then and then if it's like if it has like quite high, if it's like if it has like quite high, if it's like if it has like quite high, it's in the top 90% percentile of like it's in the top 90% percentile of like it's in the top 90% percentile of like page rank on Wikipedia, it's probably page rank on Wikipedia, it's probably page rank on Wikipedia, it's probably something quite important for the model something quite important for the model something quite important for the model to know, and what we do is uh what we do to know, and what we do is uh what we do to know, and what we do is uh what we do is we kind of like take these keywords is we kind of like take these keywords is we kind of like take these keywords and then make sure that and then just and then make sure that and then just and then make sure that and then just run like I know like standard like plain run like I know like standard like plain run like I know like standard like plain text search or like embedding search text search or like embedding search text search or like embedding search just to make sure that like these kind just to make sure that like these kind just to make sure that like these kind of concepts are in our dataset. of concepts are in our dataset. of concepts are in our dataset. And during this little bit of And during this little bit of And during this little bit of exploration, like one thing that exploration, like one thing that exploration, like one thing that we did find was that there's a Barack we did find was that there's a Barack we did find was that there's a Barack Obama the horse. I actually don't know Obama the horse. I actually don't know Obama the horse. I actually don't know if this ended up in our dataset, but you if this ended up in our dataset, but you if this ended up in our dataset, but you know, like there are many interesting know, like there are many interesting know, like there are many interesting things in the Wikipedia article, so things in the Wikipedia article, so things in the Wikipedia article, so something fun to share, but actually I something fun to share, but actually I something fun to share, but actually I don't think we had this specific Barack don't think we had this specific Barack don't think we had this specific Barack Obama the horse, but it's uh it's a nice Obama the horse, but it's uh it's a nice Obama the horse, but it's uh it's a nice horse.

horse. horse. Anyways, uh yeah, and and at the end of Anyways, uh yeah, and and at the end of Anyways, uh yeah, and and at the end of the day, like we have like around like the day, like we have like around like the day, like we have like around like 32 we I counted, uh and then I think we 32 we I counted, uh and then I think we 32 we I counted, uh and then I think we ended up having around like 30 to 40 ended up having around like 30 to 40 ended up having around like 30 to 40 like like like custom in-house classifier, like custom in-house classifier, like custom in-house classifier, like different heuristics and like filters different heuristics and like filters different heuristics and like filters that we've used. And again, like data is that we've used. And again, like data is that we've used. And again, like data is very important because once you lock in very important because once you lock in very important because once you lock in your your your I think now we're I think now we're I think now we're I mean, you're probably going to train I mean, you're probably going to train I mean, you're probably going to train like transformer, you know what you're like transformer, you know what you're like transformer, you know what you're training, and then like data is like training, and then like data is like training, and then like data is like what really determines the quality of what really determines the quality of what really determines the quality of your model, so again, can't your model, so again, can't your model, so again, can't emphasize enough how important data is. emphasize enough how important data is. emphasize enough how important data is. And then like once your data is set up, And then like once your data is set up, And then like once your data is set up, like we kind like we kind like we kind kind of go through this like Barry LM kind of go through this like Barry LM kind of go through this like Barry LM inspired training pipeline, where we do inspired training pipeline, where we do inspired training pipeline, where we do we progress we do like we progress we do like we progress we do like uh low resolution to like high uh low resolution to like high uh low resolution to like high resolution pre-training, mid-training, resolution pre-training, mid-training, resolution pre-training, mid-training, supervised fine-tuning, preference supervised fine-tuning, preference supervised fine-tuning, preference optimization, optimization, optimization, reinforcement learning, and we also like reinforcement learning, and we also like reinforcement learning, and we also like train our prompt expander, which takes a train our prompt expander, which takes a train our prompt expander, which takes a user prompt and then expands it out to a user prompt and then expands it out to a user prompt and then expands it out to a long prompt. long prompt. long prompt. So, this is pretty straightforward, you So, this is pretty straightforward, you So, this is pretty straightforward, you know, at know, at know, at typically most people like train start typically most people like train start typically most people like train start training at low resolution because training at low resolution because training at low resolution because that's where the model actually learns that's where the model actually learns that's where the model actually learns like text-to-image capabilities like like text-to-image capabilities like like text-to-image capabilities like you know, like it needs to know like how you know, like it needs to know like how you know, like it needs to know like how a horse looks like. You can do train a horse looks like. You can do train a horse looks like. You can do train that at low resolution and then you can that at low resolution and then you can that at low resolution and then you can progressively like scale up your progressively like scale up your progressively like scale up your training resolution so that it first training resolution so that it first training resolution so that it first learns semantics and then learns like learns semantics and then learns like learns semantics and then learns like structure, detail, like these kind of structure, detail, like these kind of structure, detail, like these kind of things that can be learned at high things that can be learned at high things that can be learned at high resolution later. So, we trained from resolution later. So, we trained from resolution later. So, we trained from 256 to 1K resolution.

256 to 1K resolution. 256 to 1K resolution. And then once you have your like And then once you have your like And then once you have your like pre-trained model, you kind of have this pre-trained model, you kind of have this pre-trained model, you kind of have this very malleable base to like train on and very malleable base to like train on and very malleable base to like train on and similar to like LLMs like what if you similar to like LLMs like what if you similar to like LLMs like what if you pre-train LLM, it's just basically auto pre-train LLM, it's just basically auto pre-train LLM, it's just basically auto complete. But typically, at least for complete. But typically, at least for complete. But typically, at least for LLMs, you want it to do like chat or LLMs, you want it to do like chat or LLMs, you want it to do like chat or like agentic stuff, tool calling. So, like agentic stuff, tool calling. So, like agentic stuff, tool calling. So, you actually need to like mold this into you actually need to like mold this into you actually need to like mold this into like things that are useful for you. So, like things that are useful for you. So, like things that are useful for you. So, in our case, we curate like in our case, we curate like in our case, we curate like illustration, graphic design, illustration, graphic design, illustration, graphic design, photography, cinematics, you know, kind photography, cinematics, you know, kind photography, cinematics, you know, kind of like data you want you have in mind of like data you want you have in mind of like data you want you have in mind for your downstream use case. So, we for your downstream use case. So, we for your downstream use case. So, we curate some like large-scale mid-train curate some like large-scale mid-train curate some like large-scale mid-train data and then SFT data to kind of mold data and then SFT data to kind of mold data and then SFT data to kind of mold your distribution. your distribution. your distribution. And once this little bit of like molding And once this little bit of like molding And once this little bit of like molding is done, then we do like preference is done, then we do like preference is done, then we do like preference optimization, which is, you know, if you optimization, which is, you know, if you optimization, which is, you know, if you do ChatGPT, they'll ask you like, do you do ChatGPT, they'll ask you like, do you do ChatGPT, they'll ask you like, do you like this over this? So, we collect like this over this? So, we collect like this over this? So, we collect bunch of these pairs where we use this bunch of these pairs where we use this bunch of these pairs where we use this for doing a little bit of like for doing a little bit of like for doing a little bit of like preference optimization to just polish preference optimization to just polish preference optimization to just polish out the model a little bit. This is out the model a little bit. This is out the model a little bit. This is where we get a little bit more where we get a little bit more where we get a little bit more opinionated need opinionated about like opinionated need opinionated about like opinionated need opinionated about like the kind of model we want to train.

the kind of model we want to train. the kind of model we want to train. And then, you know, reinforcement And then, you know, reinforcement And then, you know, reinforcement learning is also like kind of learning is also like kind of learning is also like kind of is now extremely standard in LLMs. is now extremely standard in LLMs. is now extremely standard in LLMs. That's what we do in diffusion land, That's what we do in diffusion land, That's what we do in diffusion land, too. So, we have like a pretty much like too. So, we have like a pretty much like too. So, we have like a pretty much like a GRPO inspired method where a GRPO inspired method where a GRPO inspired method where where we have model like generate images where we have model like generate images where we have model like generate images and then we send it to the reward and then we send it to the reward and then we send it to the reward servers and then based on this like servers and then based on this like servers and then based on this like feedback we teach the model like I know feedback we teach the model like I know feedback we teach the model like I know like how to like improve text rendering like how to like improve text rendering like how to like improve text rendering like have like better anatomy structure like have like better anatomy structure like have like better anatomy structure these kind of things. these kind of things. these kind of things. And then the last step this is kind of And then the last step this is kind of And then the last step this is kind of become a almost essential step for become a almost essential step for become a almost essential step for production grade diffusion models you production grade diffusion models you production grade diffusion models you actually need to train a small LLM that actually need to train a small LLM that actually need to train a small LLM that takes in like user prompt and then takes in like user prompt and then takes in like user prompt and then outputs like a very long detail prompt outputs like a very long detail prompt outputs like a very long detail prompt because typically longer detail prompt because typically longer detail prompt because typically longer detail prompt they're more in distribution with your they're more in distribution with your they're more in distribution with your models like training data that tends to models like training data that tends to models like training data that tends to like make better images. like make better images. like make better images. And you know the kind of next step if And you know the kind of next step if And you know the kind of next step if you're into like LLM literature is you're into like LLM literature is you're into like LLM literature is something that we are doing is something that we are doing is something that we are doing is doing like multi-expert on positive doing like multi-expert on positive doing like multi-expert on positive distillation that we are like currently distillation that we are like currently distillation that we are like currently working on so we would like train working on so we would like train working on so we would like train experts that are like specialized in experts that are like specialized in experts that are like specialized in like photography text rendering and like photography text rendering and like photography text rendering and different capabilities and then kind of different capabilities and then kind of different capabilities and then kind of like merge all of these capabilities like merge all of these capabilities like merge all of these capabilities into a single student so into a single student so into a single student so so that like we can have like kind of so that like we can have like kind of so that like we can have like kind of all we can have a student that can all we can have a student that can all we can have a student that can effectively match the capabilities of effectively match the capabilities of effectively match the capabilities of each of the expert in like whatever you each of the expert in like whatever you each of the expert in like whatever you want the expert to be good at.

want the expert to be good at. want the expert to be good at. And like in my experience like kind of And like in my experience like kind of And like in my experience like kind of like things that mattered for me was like things that mattered for me was like things that mattered for me was like infrastructure to iterate fast and like infrastructure to iterate fast and like infrastructure to iterate fast and also again like data is important like also again like data is important like also again like data is important like masses can like change every time code masses can like change every time code masses can like change every time code is something you can change very easily is something you can change very easily is something you can change very easily but you know data is like eternal like but you know data is like eternal like but you know data is like eternal like you can if you have that's going to be you can if you have that's going to be you can if you have that's going to be valuable no matter what the hot new valuable no matter what the hot new valuable no matter what the hot new training paradigm is. Uh simplicity and training paradigm is. Uh simplicity and training paradigm is. Uh simplicity and scalability I we very much prefer like scalability I we very much prefer like scalability I we very much prefer like masses that have like low number of masses that have like low number of masses that have like low number of hyper parameters that we need to tune. hyper parameters that we need to tune. hyper parameters that we need to tune. Efficiency again this goes ties back to Efficiency again this goes ties back to Efficiency again this goes ties back to just fast iteration like speed. And then just fast iteration like speed. And then just fast iteration like speed. And then like you know our thing that I like to like you know our thing that I like to like you know our thing that I like to do is steal a lot from LLM research so do is steal a lot from LLM research so do is steal a lot from LLM research so that I can just reuse their kernels and that I can just reuse their kernels and that I can just reuse their kernels and like research and like literature so all like research and like literature so all like research and like literature so all of these are things that I kind of like of these are things that I kind of like of these are things that I kind of like found very useful to iterate. found very useful to iterate. found very useful to iterate. I think I only have like 1 minute, but I think I only have like 1 minute, but I think I only have like 1 minute, but kind of things that I find funny is kind of things that I find funny is kind of things that I find funny is that, you know, we started with like that, you know, we started with like that, you know, we started with like encoder and like decoder encoder and like decoder encoder and like decoder uh transformers. Now we are going the uh transformers. Now we are going the uh transformers. Now we are going the opposite way because we have a prompt opposite way because we have a prompt opposite way because we have a prompt expander which is all the regressive expander which is all the regressive expander which is all the regressive decoder and then diffusion which is the decoder and then diffusion which is the decoder and then diffusion which is the encoder, so it's kind of getting encoder, so it's kind of getting encoder, so it's kind of getting reversed.

reversed. reversed. And with that in mind, it's also And with that in mind, it's also And with that in mind, it's also starting to look a little bit like starting to look a little bit like starting to look a little bit like DALL-E 2 where, you know, you have like DALL-E 2 where, you know, you have like DALL-E 2 where, you know, you have like a model a model a model that will like generate conditioning that will like generate conditioning that will like generate conditioning using like auto regressive model or like using like auto regressive model or like using like auto regressive model or like diffusion model and then we feed that to diffusion model and then we feed that to diffusion model and then we feed that to our diffusion model, so it's kind of our diffusion model, so it's kind of our diffusion model, so it's kind of this prompt expansion pipeline makes this prompt expansion pipeline makes this prompt expansion pipeline makes reminds me a little bit of like DALL-E reminds me a little bit of like DALL-E reminds me a little bit of like DALL-E 2. 2. 2. So, that's kind of a funny observation. So, that's kind of a funny observation. So, that's kind of a funny observation. And you know, like And you know, like And you know, like there's bunch of stuff that go into like there's bunch of stuff that go into like there's bunch of stuff that go into like diffusion model training. I really like diffusion model training. I really like diffusion model training. I really like to, you know, simplify the stack so that to, you know, simplify the stack so that to, you know, simplify the stack so that we can get rid of VAEs and then text we can get rid of VAEs and then text we can get rid of VAEs and then text encoders and then just train a single encoders and then just train a single encoders and then just train a single clean transformer. clean transformer. clean transformer. Uh so, that's something I always look Uh so, that's something I always look Uh so, that's something I always look forward to working on. forward to working on. forward to working on. And then, you know, vision language And then, you know, vision language And then, you know, vision language models has gotten like more powerful models has gotten like more powerful models has gotten like more powerful like this from the original LDM paper, like this from the original LDM paper, like this from the original LDM paper, but but but you know, like bounding boxes you know, like bounding boxes you know, like bounding boxes before it was expensive to generate, but before it was expensive to generate, but before it was expensive to generate, but now you can perfectly generate good now you can perfectly generate good now you can perfectly generate good bounding box for every image and you can bounding box for every image and you can bounding box for every image and you can condition the image model on this. condition the image model on this. condition the image model on this. Uh this something that Ideogram rev uh I Uh this something that Ideogram rev uh I Uh this something that Ideogram rev uh I have had also like been working on. And have had also like been working on. And have had also like been working on. And then, you know, this is from actually then, you know, this is from actually then, you know, this is from actually 2017 from uh Fei-Fei Li 2017 from uh Fei-Fei Li 2017 from uh Fei-Fei Li uh uh uh this lab like you could do like scene this lab like you could do like scene this lab like you could do like scene graph to like image generation. So, graph to like image generation. So, graph to like image generation. So, again like image generation, I think again like image generation, I think again like image generation, I think it's really like a proxy for like BLM it's really like a proxy for like BLM it's really like a proxy for like BLM like progress, so kind of things I'm like progress, so kind of things I'm like progress, so kind of things I'm like excited for is like what are other like excited for is like what are other like excited for is like what are other what are interesting like textual ways what are interesting like textual ways what are interesting like textual ways to like describe an image now that we to like describe an image now that we to like describe an image now that we have more like more like powerful visual have more like more like powerful visual have more like more like powerful visual language models, so this something that language models, so this something that language models, so this something that I'm quite excited about like bounding I'm quite excited about like bounding I'm quite excited about like bounding boxes, you know, scene graphs. I mean, boxes, you know, scene graphs. I mean, boxes, you know, scene graphs. I mean, there's different things that could be

there's different things that could be there's different things that could be like quite useful here. like quite useful here. like quite useful here. And then, yes, and this is my shameless And then, yes, and this is my shameless And then, yes, and this is my shameless recruiting slide. So, if you enjoyed recruiting slide. So, if you enjoyed recruiting slide. So, if you enjoyed this talk, send me an email here, take a this talk, send me an email here, take a this talk, send me an email here, take a picture, and yeah, I think that's mostly picture, and yeah, I think that's mostly picture, and yeah, I think that's mostly it. I don't know if there's time for it. I don't know if there's time for it. I don't know if there's time for Q&A, but sorry that I went slightly Q&A, but sorry that I went slightly Q&A, but sorry that I went slightly over, but hopefully you enjoyed it. I'll over, but hopefully you enjoyed it. I'll over, but hopefully you enjoyed it. I'll stick around for another 30 minutes if stick around for another 30 minutes if stick around for another 30 minutes if anybody wants to talk to me, but yeah.