Visual Generative AI Ecosystem Challenges with Richard Zhang - #656

0:10All right, everyone, welcome to another episode of the TwiML AI podcast. I am your host, Sam Charrington, and today I'm joined by Richard Zhang. Richard is Senior Research Scientist at Adobe Research. Before we get going, be sure to take a moment to hit that subscribe button wherever you're listening to today's show. Richard, welcome to the podcast. Yeah, thanks so much for having me. It's really exciting to be here. I'm excited to be speaking with you. We're going to dig deep into some of your work on image-based generative AI, and we've got some interesting things to talk about. I'd love to hear a little bit about your background and how you came to work in computer vision.

0:51So I started grad school back in 2012. That was kind of a very seminal moment we had there. You may recall the Alex Knight paper came out. That was actually my first semester there at Berkeley. That's when we started really seeing that deep networks were really working for computer vision tasks, such as image classification. That really got me excited about that field. Awesome. So it was just good timing. Yeah, really fortunate timing. Yeah. That's awesome. And what were some of the first interests you had in the field? Yeah, definitely. So after the AlexNet paper came out, you saw the immediate reaction was everyone was applying deep networks towards discriminative tasks.

1:25But on our end, we were kind of interested to see if deep networks could be used for image generation instead. And right now, obviously, they can be. But back then, that wasn't so obvious. And the reason for that is, instead of predicting like a low dimensional label for image generation, you're predicting kind of a high dimensional signal. And there's simply a lot of ways you can go wrong. And also, you know, you can, this just takes more compute to predict a high dimensional signal. And back then we were much more limited in that. So kind of one of my early experiences in a kind of generative task was that of image colorization.

1:55So for that, you may have some, you know, old family photos that are in black and white, you may want to add kind of a splash of color to liven it up. That was kind of one of the first projects that I worked on and kind of an early experience that I had. Yeah, I think a lot of us will remember seeing those first colorized photos and later on colorized videos, of course, obviously predating the term generative AI, but they were generative in nature in a sense. Yeah, definitely. So we were calling it generative modeling, but now that it works, it's generative AI and it's not a bad term. I'm happy to go with it.

2:24Awesome. You know, one of the things that we talked a little bit about before we started recording was just the way you think about and describe visual generative AI as this ecosystem that you have tried to, I guess, balance in a sense in your work, the parties and that ecosystem being the creators, the consumers, and the contributors. Can you elaborate a little bit on that idea? And what are some of the main concerns you're thinking about from each of those perspectives? Absolutely. So maybe I'll take us back to that project on colorization. So back then we were really trying to push on improving visual quality because that really was the bottleneck towards making things work.

3:06One area that I found was really challenging was having a good loss function that actually described that could kind of predict visual quality as to what a human would think looks good. So when we were trying to do colorization, we were taking kind of off-the-shelf networks, or sorry, off-the-shelf loss functions, trying to train the network and getting kind of very dull, sepia tone, blurry results. And the reason for that was we didn't have a good loss function. So one thing I was really working on after that. And by that, you mean you were not able to kind of accurately represent what a human would perceive as a good result.

3:38Is that the idea? Yes, precisely. So deep networks fundamentally are trained with some sort of loss function that have to guide the network towards either a good result or towards a better result. But if that function doesn't describe what humans think look good, it's not going to be guided towards that good result. It's going to be guided towards something else, which is kind of what was happening back then. Yeah. So that kind of inspired me to push on creating better loss functions that were more perceptually aligned with people. And thanks to some of that work, along with a lot of developments in the field, gender modeling or Gen AI really started working.

4:09And we really started seeing that, you know, it wasn't just going to be this kind of laboratory thing. We were perhaps on the cusp of something big that, you know, gender models could be underlying all sorts of different kinds of content creation tools, not just creating these kind of one-off applications, you know, which are fun, like colorization, but are kind of done piece by piece. And so I think that got me and a bunch of colleagues thinking about all this technology from a more holistic standpoint. So it wasn't just about writing papers and improving quality a bit. We were going to have all these images all over the place, which fundamentally is really cool, but we also want to make sure that the consumers of all this visual content have tools to be able to discern what's real and what's synthetic.

4:47And beyond that, a lot of these systems are, of course, based on large-scale training data. And that's just not like a static thing that comes out of thin air. Training data comes ultimately from people, right? And so we want to make sure that those data contributors are also adequately recognized for their contribution. So yeah, I think that's the way I kind of think about this ecosystem that we want kind of creators, consumers, contributors all to be happy to take part in this technology together. So let's maybe dig into the perceptual work, which is from that creator perspective. That paper, if I'm remembering, It was quite a while ago, but I'm wondering if you can kind of share broadly how it is that you think about creating better perceptual metrics and loss functions to guide generative AI.

5:31Absolutely. Yeah. So as we mentioned, deep networks have to be trained with some sort of loss function that's guiding it towards an answer. And so the way that works is the deep network is going to predict some sort of answer. And perhaps you have a ground truth that you want to compare it to. So you have a prediction and a ground truth. You need to get some sort of number as to whether that prediction was good or not. So this is just a function that takes in two images and tells you how close they are. But it turns out that this is actually a very difficult problem. It's actually an open problem in a whole bunch of fields like image processing, graphics, and even neuroscience.

6:01It's actually just really difficult to write down some sort of function in closed form that describes all the nuances of the human visual system. And so I can give you just kind of a very simple loss function, which is just L2 or Euclidean distance. You just compare the two results pixel by pixel, and you sum all the differences across the whole image, and then you get some sort of number out. If the two images are exactly the same, you're going to get zero. So that's at least a good property to have. But it turns out this doesn't really describe the human perceptual system well. And I can just give you a simple example.

6:26If you have an image that's just like a vertical edge here, and if you were to just move that vertical edge over by one pixel, you and I, looking at those two images, wouldn't even see a difference. But to L2, that's a huge difference. It's a huge lot, yeah. Yeah, it's zero in one image, it's one in the other, and vice versa, right? And so because of that, if you use L2 and there's a whole bunch of different possible outputs, it's going to really just hedge its bets and average over all these possible results. And we've seen this effect in a whole bunch of different problems. So we wanted kind of a loss function that could correspond better to human perception and really drive you towards a mode.

6:57So I can go a bit into how we went about doing that. Fundamentally, we kind of realized that just trying to describe the human visual system in math is really difficult. So instead, we took a data-driven approach. So what that means is I coded up a whole bunch of different types of distortions. For example, you can take an image and you can like blur it, you can shift it to the left, to the right, up or down, you can warp it, you can change the saturation, the contrast, you know, different photometric changes, you can add different checkerboard, different artifacts like checkerboard patterns. So I had like a big set of distortions.

7:25And then what I did was I took a patch and I distorted it in two different ways, let's say A and B. And then I asked people like on the internet, or I could ask you, does image A look more like the reference? Or does image B look more like the reference? And they would either choose A or B. And that gives you some sort of a signal towards perceptual similarity. And what that means is if they chose A, a good loss function would put image A and the reference closer together than image B and the reference, right? And so by collecting this kind of data set, we actually got like half a million judgments, I think.

7:54We could then evaluate how different embeddings agree with humans. And on top of that, we could also even improve those embeddings. So for example, you could take AlexNet or VGG, and you can select which channels are more quote unquote perceptual. And so that's how we end up getting a better loss function. So that's called the Learn Perceptual Image Patch Simulary Metric, which is a bit of a mouthful. So I think it's just known as LPIPs. And so that's used quite often in evaluating different results, like in colorization, of course, and super resolution, view synthesis, compression. Basically, anytime you're trying to reconstruct a signal, this is something that's kind of serving as a tool for people to use.

8:29Got it. And when you looked at AlexNet and VGG and identified the channels that were perceptual, is that a function of the architecture or is it specific on a trained model of that architecture, kind of specific to the data? Yeah, that's a great question. So when I was doing this work, actually, people were using VGG, trained on ImageNet. In the literature, that was called a perceptual loss. One thing we were kind of not satisfied with was people were calling it a perceptual loss, but no one had really done the experimental validation to actually, you know, validate that it is something that corresponds with human perception.

9:03So that's part of what motivated our study and as to kind of evaluating VGG or AlexNet or ResNet and all these things. One thing we found that was interesting was that it was a more of a function of the architecture than you may have imagined. And it was actually backwards from what you may have assumed. So the name of the game was you stack as many layers as possible. Like AlexNet is seven layers, VGG is 16 layers. So it does need to be deep, right? Yeah, they need to be deep. ResNet is 50 to 100. And computer vision performance just gets better and better with more layers. But it turns out the correlation with human perception actually goes down, like it goes up, up to AlexNet and hits a maximum and then goes down.

9:38And that was kind of unexpected. And I think the hand wavy kind of explanation is AlexNet is like seven layers, the human brain is seven layers ish. I mean, there's some feedback and stuff, but seven layer ish. And so by having that kind of kind of correspondence, maybe that's the reason why Alex ended up correlating the best. So that was kind of an unexpected result that we had upon. Interesting. And how does that reconcile with the fact that we went much deeper with these networks? Is it just like a point of diminishing returns after seven or are those additional layers just doing other things or?

10:07Yeah, it's not clear why. I've also seen results from neuroscience from Jim DiCarlo's lab at MIT looking at correlating deep network responses to neural responses in monkeys. And they kind of solve the same sort of thing where it goes up to a certain point and then goes down. And part of that might be just because the really deep networks for the kind of narrow tasks that they're trained on are perhaps just superhuman. Like a ResNet 150 is probably better at doing ImageNet classification than I am, right? So at that point, maybe that's why you actually start seeing the correlation go down. Okay.

10:37And just to be clear, what is going up and going down? Are we talking about the overall performance of a model on a given data set for a given task? Like it sounds like you're saying that basically seven layer networks are optimal. And that's what I'm finding hard to believe applies like super broadly for perceptual tasks. Am I interpreting your assertion too broadly? Yeah. So maybe I'll map out what we're correlating. So think on the y-axis, the agreement with the perceptual data that I collected in the Alpips data set on the x-axis is computer vision performance, right? So if computer vision performance were perfectly correlated with perception, it would just go up and up.

11:14But what you see is that it kind of goes up and then it goes down. So there does seem to be a point where by getting better computer vision performance, you actually reduce performance on corresponds to perceptual similarity. Got it. So within a certain range, computer vision kind of follows along with kind of tracking this perceptual metric that you outlined. Beyond that, you're not tracking the percept, you're going the opposite direction from a perceptual perspective, but maybe you're still kind of just brute forcing performance through these deeper and deeper networks. Maybe you're memorizing more.

11:46Other things are happening, but it's not because perception is getting better. Yes, exactly. Definitely the deeper representations are better. They're more useful for a whole bunch of tasks. It's just part of the correspondence to human perception, using it as a loss function, it starts kind of getting worse at a certain point. Got it. Oh, that's super interesting. That's really interesting. And so is this method, LPIPs, is it used in kind of some of the contemporary gen AI things that we've all been hearing about recently, stable diffusion and the like? Yeah. So there's a little bit of irony in the story where like I was working a lot on these loss functions because, you know, L2 wasn't adequate, but now everything is based on diffusion models for now.

12:24And you may know diffusion models, they're just trained with L2, right? And so somehow in that setting, L2 doesn't end up blurring things. And the reason for that is diffusion models will take a big, very difficult problem of going from noise to image, chop it up into like like a thousand little steps. And by doing that, L2 is actually perfectly happy towards driving towards modes rather than towards some blurry mean. So it turns out that, yeah, LPips isn't used in the central part of diffusion, but actually it's still very practically useful in a lot of systems. So for example, you asked about stable diffusion, that's based on LDMs or latent diffusion models, which is an academic paper, I think from Munich.

12:57And so a part of LDMs is this compression step. So the idea is that you could run diffusion on raw pixels and that works really well, but that's going to be really expensive because pixels are, you know, very high dimension. Instead, you could go from that high dimensional space to a compressed code and do the diffusion on that compressed code. Now, the way you train, you know, this compressor is you need an encoder and a decoder that reconstructs the original image. And the way you judge whether that reconstruction is good or not, in part, is without PIPs. Perceptual models. Yeah, exactly. And there's also other stuff on top of that, like GANs and KL Divergence from VAEs.

13:29Basically, kind of a list of generative technologies are all kind of actually used in LDMs on top of the diffusion core. Awesome. So we've talked about the colorization task. We've talked about in stable diffusion image generation tasks. Are there other areas that you've worked on, things like super resolution come to mind or other aspects of visual generative AI or visual generative computer vision? And again, kind of staying with this creator realm before we jump over to the consumer side of things. Yeah, definitely. So one thing I'm interested in is really adding a lot of controllability to generative AI systems.

14:05Yeah, so that's something we're working on. So the reason for that is if you look at GANs, you know, you're either going from a random noise vector to an image, or if you look at these text image models, right, they're going from a short sentence into generating a whole image. But really, that's not giving a creator all the control that they may want to have. So you've probably heard the adage, like a picture is worth 1000 words, but in DALI 2, you only get 76 tokens. That's why you're limited. So what am I supposed to do with all that other detail that I need to add in? So we're really thinking about what types of control can we actually add to the system.

14:39So text image is really great for getting you started and for getting kind of a nice foundational model that knows a lot about how natural images and different art styles look like. But we also want to add in control into that process for creators. And it sounds like you see that control coming from not expanding the number of tokens available, but via other like out of band types of things. What are some examples of like specific examples of the types of controls that you're envisioning or have worked on? Yeah. So even if we gave you a thousand words, you probably don't want to write an essay for every single possible thing in the image.

15:15And there may be some things that I would want to standardize without having to write about like every image that this model produces is in the theme of starry night or something like that. Yeah, exactly. It's just text in some sense for a lot of operations is going to be very clunky, right? And so what we want is just a whole spectrum of, I think, different ways or kind of think about them as hooks for interacting with a system. So if an expert user wants to go in there and start like adjusting the gray level of specific pixels, they should be free to do that. And there should be also kind of more mid-level things, of course.

15:47Like if you were to create an image on pen and paper, you would start sketching, right? That's kind of a very natural way of making an image. And so that could potentially be an input modality that we can add. There's also different things like style transfer, right? You may have a repository of little paintings that you made that you want to apply towards making a new image. And there's also adding kind of personal objects, like maybe I want to take my face or take my, like a pet or my toy or my car, something like that, and add that in. So you can think of just a whole lot of different types of controls that could be added.

16:16I think one key thing is having enough controls and preserving kind of the state of the results so that people can iteratively interact with the system. So it's kind of a back and forth. It's not you type in some text, you wait five seconds, you may or may not get what you want. And then you have to adjust the text and it gives you some other complete random image, right? We want to have some sort of permanent state that allows you to iterate with it kind of meaningfully. And how far along are you in this direction? Have you Have you published on it? And what kind of early results are you seeing in this controllability domain?

16:47Yeah. So one area that we worked on was that of customizing models. And so there are concurrent works out there, such as Textual Inversion and Dreambooth, which are all very pretty well known. Our method was called custom diffusion. So that's kind of one method of interacting with systems where you can take a small collection of your own photos of a certain object or style and basically add that into the network. It's like doing a little bit of network surgery. So you can add some of your content in there. content in there. You don't have to just stick with the pre-trade model. So that's one aspect we're playing with.

17:15We're creating kind of more interactive systems as well. It's not out yet. So hopefully for a CBR or SIGGRAPH or something. Yeah. So we're working hard on this. Okay. So watch the space. Yes. Awesome. How about on the consumer side? You talked a little bit about who that consumer represents in the ecosystem, but what are the things that they care about most that are driving your research? Yeah, fundamentally, it's really great that all these Gen AI systems are working now, and we have a lot of that type of visual content. We do want to make sure that if there's going to be some percentage of malicious actors, that we have some sort of solution that scales up to societal level for dealing with that in some way.

17:52And so we kind of wish we want people to have tools for being able to tell what's real and what's fake. And fortunately, there's actually a lot of areas to, I think, contribute to in this space, because I think ultimately we really need a multi-pronged solution for dealing with this. So I think about kind of three different areas. So there's provenance, there's detection, and then there's just plain education of the public as to what tools are out there. I'm working more on detection. I can describe provenance a little bit. I'm not an expert, but I can kind of describe the high level idea, which is that if we had some sort of kind of open standard for recording metadata as to the kind of origin as the origin and subsequent edits to every piece of visual asset, that would be really useful, right?

18:30That would be kind of like getting a nutrition label for a piece of visual content. And the nice thing about this is if this were to become very standardized, even if you got a visual asset that didn't have that metadata on, that itself would be suspicious, right? So that would also be some sort of signal. In what ways is that an AI challenge as opposed to a software engineering slash standardization challenge? Yeah, I think it's more the latter, which is why, because I'm not an expert in that field, I'm not working on it as much, but I think actually that's hugely important. So I'm glad there's really smart people working on it.

19:04I'll point you to the Content Authenticity Initiative or CAI. Got it. It was like a coalition that was established, I think three or four years ago. It has like a huge number of member organizations now. It's looking at kind of creating an open standard so that this type of metadata can proliferate out in the ecosystem. I think that would be really great. So it's the CAI. Yeah. Yeah. Sounds like the big challenges there are getting folks to agree on the format and what kind of information and then some kind of set of cryptography standards to make sure that the information can't be tampered with once it's injected into the image or applied.

19:36Yeah, absolutely. There needs to be probably some standardization. Just like be bad if all our wall plugs in every state were different. Same kind of thing. There needs to be a little bit of coordination here. Yeah, absolutely. And so there was the provenance and then there's detection. This is certainly an issue that predates generative AI. Like Photoshop became a verb a long time ago for manipulating original images. Have you been working on it predating? Or maybe the question that I'm trying to get at is, do you see generative AI as a kind of a concept that has flourished over the past several years providing unique challenges and or tools in the context of detection?

20:17Or is it kind of the same work that's been going on around detection for a long time? Yeah, that's a great question. I think we need detection tools for all different types of content creation tools. And there is some fundamental difference. So we had a project kind of on detecting if an image was Photoshopped and more specifically, a specific button in Photoshop called FaceWare Liquify, which can kind of warp a person's face, which is kind of it's not really Gen AI based, right? It's just kind of some simple warping for that type of tool, because it's kind of static, like that button is not going to change from today to tomorrow.

20:48It's going to be the exact same. I could create a very specialized detector just for that. But the difference with Gen AI systems is that today we have this type of method. Tomorrow you check archive and something new has come out. And of course, tomorrow's archive is not completely different than what was written today, but it's going to contain some differences. So really, we're always trying to catch the next method that's coming in the future. So it's just going to be a bit more fast moving in nature versus the kind of more traditional. And, you know, thinking about some of the technical approaches that may come into play, historically, you could do simple things like edge detection, trying to find kind of imperceptible edges that may exist in a photograph that may indicate that it was manipulated with and wasn't an original image.

21:33Whereas if you're generating the original image from scratch, that's a different paradigm for detection. How do you think about the way to approach detection when you're generating photorealistic images in their entirety with an AI? Yeah, so one thing we've been looking at is taking a very data-driven approach. Because some of that motivation is that AI-based systems have democratized content creation so much. But it was really kind of this asymmetric situation where anybody like me or some kid in their basement could create deepfakes. but thanks to like the advancements in AI, but those advancements hadn't been applied to forensics yet.

22:05Really for forensics, you almost had to be like CIA or a major research lab to kind of have that expertise. So we want to see if we can actually use data-driven techniques for helping with forensics as well. So we did this kind of pilot study three years ago. Like I mentioned, generalization is extremely important. We want to see if to train a detector on one type of method, will it actually work on tomorrow's? Wanted to do some sort of study with that. So we collected like 12 different types of generative methods. So back then they were mostly GAN based, but some of them were like outpips or perceptual loss based as well.