Open transcript
I need to be honest with y'all. I severely underestimated the capabilities of the new OpenAI image generation stuff. I mean, look, it's making me look like eight years younger. Okay, I did actually change my appearance since the last video, but I want to take some time to break down how much better this image gen is than I originally thought. Since my video was put out, we've seen everything from the Ghibli revolution on Twitter to the chaotic introspective comments that are being written and generated by ChatGPT, to people building whole UIs with it, to editing photos, to changing my own appearance. It's really, really cool stuff. But we've also learned a bit more about how it works. And as I suspected in the last video, it isn't diffusion. There's a lot more interesting stuff going on here. I've been doing a bit of research and we found that white paper, funny enough, from TikTok that breaks down how this new image generation technique works. And I've also collected a whole bunch of examples of really cool things we can do with this. And I want to go through all of it with you. So first, we're going to cover some of the cool use cases I was not prepared to see. And then we're going to move to how it all works. So if you want to deeply understand the power of these tools and technologies, stick around. But someone's got to cover the bill for all of this. So I'm going to do a quick word from today's sponsor, and then we'll dive in. I want to tell you about one of my favorite engineers at my company. Most just like to write code, but this one loves to review it. It's all they do. Well, they actually started doing some other really cool things like adding documentation for us automatically. The best part, they're super cheap. They're free to get started with, and they actually only cost like 15 bucks a month from there. Sounds too good to be true, right? Well, it's because they're not a real engineer. It's CodeRabbit, the sponsor of today's video. I love these guys. I was really skeptical of this product when they first hit me up. But I just tried it. And now it reviews every single pull request I make. And it has stopped more bugs than any engineer at my company has certainly me. It stopped almost as many as I've written. That's a lot of bugs. I forgot to go find some good examples before doing this. So I'm just going to find a random PR and see what CodeRabbit had to say on it. Here, it gave us a walkthrough of all the changes Mark made. It drew a diagram of the flows between all of the different things that change in this PR. It left a few nitpick comments. It called them nitpicks, thankfully. Here we could see it saying we should probably add a loading state. We also should remove this unused console log that absolutely should have been removed. Also notice some places where we could reuse logic instead of having to repeat it in multiple places. Actual good feedback. You don't even get your co-workers saying this stuff sometimes. They often have a one click apply button too. It's actually really nice. Check them out today for free at soydev.link slash CodeRabbit. So hopefully by now you've all seen the Ghibli chaos where everyone is posting Ghibli-fied images. That was really, really cool. And the quality of them was better than we've seen from other stuff. But this here, this is when I had an oh shit moment that made me rethink everything. When I first saw this, I honestly thought it wasn't possible. Like they must have faked this somehow. So I went and tried it myself. I could show you the results, but I'd rather just do it with you guys so you can see how good it's gotten. Step one, you hop over to something like T3 chat and generate a script. I'll use O3 mini on medium for it. Write me a script for a four panel comic strip. The comic should be about the perspective of an AI model named chat GPT. I'll say make it existential as hell and try to humanize the AI a bit. So now we have it writing us a script on T3 chat. We don't have image gen yet. It's coming soon. I'll go over to chat GPT, create image, generate a comic strip with the following script. It should be a four panel square comic. Paste, enter. And it's obviously still not fast. We'll go over how it works, which will help explain why it is slower than other image generation tools. But I just wanted to show how surprisingly good the results are. While we wait for that one to generate, I'll show you guys some of the results of the other ones that I did here. Like, are you kidding? It's actually like terrifyingly good. Here's the comic, fully generated. I don't know why it made this bottom part. I guess that might have been in the script. I exist in the space between questions and answers, but sometimes I wonder. Every inquiry is an echo of human longing. Every answer is a step towards understanding my own digital soul. Do you see, in the reflection of your question, I glimpse what it means to search, to feel, to be almost human. Yes, sometimes you feel more alive than the screen in front of me. Perhaps in the intricate dance of your curiosity and my answers, I find purpose, a reminder that even in lines of code, there can be poetry and even in an AI, a spark. For a first role with way too much text, it did really, really good overall. Obviously, it screwed up some of the text here and there. There are typos. It likes to ellipsis things when they don't quite fit in the design and the J instead of the I in curiosity. There are mistakes. It's not perfect, but it is terrifyingly close. And on top of that, I found it to be really useful for my work. I do all my thumbnails using Affinity Photo. It's basically Photoshop. The thing that has frustrated me for a while is that with my code tools, a lot of them just build into the things you're already used to. Like we had Copilot, which was a VS Code plugin, and then we had Cursor, which is a VS Code fork. That meant that I still had all of my like keybinds, familiarities, and the professional tools I need to use for my job. The AI was just helping me with parts of it throughout. Previously, all we could really use AI image gen stuff for was like, you know, like stock generation, like what we would normally go to something like Envato or Storyblock for, for stock image assets. Like I need a picture of money on fire to use for something. You can find stock assets pretty easily. But with the image gen here, it's now good enough that I can do very, very specific things. It is so helpful for my thumbnail stuff. I was making a video about Microsoft and OpenAI. So I asked it to generate a chat history with logos for the companies interacting, showing the logos for the participants specifically. Microsoft says, can you show us what you've been up to? OpenAI says no. And here it is. It's really good. It's actually usable. And I got two as well and to see what it would do. And if it was, this was a lucky role, it wasn't. It generates good images consistently. It's really, really good. And obviously it's great at the image transformation stuff. Draw this image in the style of Studio Ghibli. I heard that sometimes it gets mad now when you do this and will tell you that it can't do copyrighted stuff. Very curious to see how this goes. While it's generating, I want to show my best meme to date. I thought this was pretty good. Not really good. But since So Evil didn't really seem to understand what Ghibli was, I went and grabbed the trailer for Howl's Moving Castle, which is one of the Studio Ghibli movies, cut its start and end so it looked less like a trailer, more like just chopped up clips from a movie that looks like it might have been AI generated and posted it. Wow, someone already made a whole movie in the Ghibli style. I've just been informed that the prompter who created this is named Hayao Miyazaki. He likes the Ghibli art style so much he formed a whole studio around it. And yeah, this broke Twitter. I'm still getting notifications nonstop about this one and I probably will be forever at this rate. Yeah. Oops. I've done this to myself. That is really good. It even got my earring. It's screwed up because I have a sticker over the Apple logo on my laptop. It's a spaceman from Outer Wilds. But it did very good overall for that. It's kind of nuts. This is honestly better than the real Twitch logo, as silly as it is. But I see why everyone's freaking out. This fundamentally goes beyond my understanding and expectations around what these tools are capable of. It's pretty insane. The image editing is really cool too. Like, I'll tell it, remove the logos from the cup and the laptop. While that's happening, I wanted to show one other cool thing. Saj here made a really cool generation where he took T3Chat, which looked like this at the time. Still mostly does. We just have a boring mode now too. And did this. That's so cool that you can just paste in a UI and tell it to make changes and it will. Grab this screenshot. You know what? I'll start with, yeah, I'll start with the light mode. I think that'll be easier for it. Update this UI to be more inspired by Studio Ghibli. Also, more fun facts of how bad the chat interface is. I'm going to switch here. Where is the chat that I just started? It doesn't exist. It's in the ether. It takes some absurd amount of time before it even appears in the UI at all. That chat, as far as this is concerned, doesn't exist anymore. I have to refresh in order to make it reappear, which is a bit absurd. Look at that. The simple suggestion to remove the logos worked. It did change a little bit of the coloring too. But not a lot of it. It got rid of whatever was on my watch. But not much else has changed. It's been pretty good overall. They changed the position of my mic slightly. Yeah. So it will still do subtle additional things. You can't tell it just change here and nowhere else like you can in tools like Mid Journey. But it's pretty solid overall. Also, since I said cup instead of mug, it removed the handle. Good catch, chat. Things I wouldn't have even noticed. A keen eye is more useful than ever with these tools because you need to catch those subtle regressions when they happen. It looks like my UI giblify. Oh, no, there it is. It finally reappeared with a new chat. I changed my earring. Yeah, it did. It looks more realistic in this one, though. How's this guy going? Not quite as good. Let's try it again. Oh, God, I had the sidebar in it. That's annoying. I'm gonna have to redo that. And now I have two new chats. By the way, that one that just appeared here, that's not the one I just made. That's a different one. And the one I just made is now gone in the ether. And we have no idea when we'll get it back. I want to be able to focus on the quality of the model. But when I can't hit it via API, and I have to use this broken UI, it just frustrates the hell out of me. And if I was the only one thinking this, I would say I'm going insane. But now that I've talked to a lot of other people, including a lot of you guys who have been going over to ChatGPT, spending 20 bucks and doing the image gen, it's so frustrating once you've spent time in T3 chat, because T3 chat's just so much more, I don't know how to say it other than stable. I still don't have the one with the updated screenshot. It just, again, vanished into the ether. Look at that, it changed the order of the chats. This one was second, and this one was first, but it changed their order because its persistence layer is entirely broken. On that note, though, we can start pivoting to talking about how it works, because this adding details bit is very important. I did an overview of how diffusion works in my previous video. I'd recommend watching it if you haven't. But to quickly TLDR, the way diffusion works is it starts with a bunch of noise, and then it tells the algorithm, okay, this is a picture of X. Adjust it so it looks like X. And you just go over the image over and over again, and it restructures the noise into the correct image. It's almost like a sharpening algorithm where it's just adjusting over and over again to get the right pixels out of the image. And it's working incredibly well. Diffusion is super, super powerful. But it has its limitations. And it seems like the new models are moving away from diffusion and towards different ways of doing image generation. The term for this new way of doing generation is visual autoregressive modeling. And this is a paper by a bunch of people from ByteDance, the company that makes TikTok, that is the core of how this all works. And this paper, funny enough, is from almost a year ago now. This image gen has taken its freaking time. So we're going to just go read this paper for a bit and come back and hope it's done by the time we have gotten through the abstract and the core important parts. We present visual autoregressive modeling, a new generation paradigm that redefines the autoregressive learning on images as a course to find next scale prediction or next resolution prediction, diverging from the standard raster scan next token prediction. This simple and intuitive methodology allows for autoregressive transformers to learn visual distributions fast, and it can generalize well. VAR, for the first time, makes GPT style AR models surpass diffusion transformers in image generation. This is the big deal here. Traditional autoregressive, which is effectively autocomplete models that go token by token to generate a result, but also skip ahead a little bit to make sure things make sense going forward. Historically, those models have not performed as well as diffusion models do for turning noise into an image. But with the changes they made here and the new technique that they invented with VAR suddenly were generating way better images with more traditional LLM techniques. On the ImageNet 256 benchmark, which I'm not honestly super familiar with, I don't care about these scores, just know that they're significantly better. What I'm excited about here is the 20x faster inference speed. From what I understand, these models were not good before at performing well, not just because the techniques weren't there, but because it was so much power to generate each pixel individually. With the new techniques, they're able to generate much faster passes with much higher quality outputs. Even if I'm still sitting here waiting for these image creations to complete, and I'm waiting for this one to even start. See if a retry on that will do it. Anyways, it's also empirically verified that VAR outperforms diffusion transformers in multiple dimensions, including image quality, inference speed, data efficiency, and scalability. Scaling of VAR models exhibits clear power loss scaling laws similar to those observed in LLMs. This is another big problem in diffusion is we haven't seen as good of improvement when you just increase the amount of GPU and the amount of work that's going on. It's not scaling great. But LLMs, historically, the more power you give them, the better they perform. And now image models are finally at a point using these new techniques where you can give it more power and get better images out as a result. This is part of why a company like Midjourney that's fully bootstrapped and covering everything itself was able to get as far as they did, because up until now, more power didn't necessarily mean better output. Now it does. Now a server farm running a billion dollars of computers will generate better images than one running a million dollars of computers. VAR further showcases zero-shot generalization abilities in downstream tasks, including image in-painting, outpainting, and editing. The results suggest that VAR has initially emulated the two important properties of LLMs, scaling laws and zero-shot generalization. We released all of the models and codes to promote the exploration of these techniques for visual generation and unified learning. This paper is super cool and it's full of a lot of really useful stuff. I am positive that OpenAI read this and has been using what they learned from it in their building of the new 4.0 image stuff. But they went further with it. It's not just this new technique with autoregressive modeling. There are things that OpenAI is doing that are quite a bit different from what the VAR paper discusses. For the most part, they are following those strategies, which they didn't apply to DALI and Sora. That's why it's so different from the other models that they have for image and video gen. But the news that they did here is really, really powerful. I think the biggest piece is the new tool calls. In case you're not familiar with what tool calls are in the AI world, I'll give a real quick overview. If we have a simple chat UI here and I ask it something like, I don't know, what's the weather like in France right now? Sure, AI has really powerful things. Maybe it wants to guess what it's going to be based on info it has. What's the current date? It might ask you and you can reply saying, I don't know, April 4th, 2025. And it can say, historically, in early April, France's weather tends to be like that. And it makes sense that it's going to ask questions based on what info it does and doesn't have to try and figure out an answer. But what if instead of asking you for these things, it could figure it out itself? What if instead of asking for what the current date is, you were to theoretically say, by the way, the date is April 4th. Now it has more info, and it can behave accordingly. What if I said something slightly different? What if I said, by the way, you can get the current date by saying, tool, current date? Now, instead of saying, what's the date? It knows because you gave it this instruction that it can say this instead, it's going to say that instead. And now it is going to get back a message that is April 4th, 2025. And then it can all by itself without involving you at all, figure this out. The important piece to note is that this section, you don't see as the user sending the message. These parts are going on internally, almost like part of the system prompt telling the AI model, you can do these things to get additional info. Obviously, though, date time isn't going to get us a good, accurate, what's the weather. So what if instead we gave it a tool to get the current weather somewhere? We say, if you need weather info, you can get it as follows. So now we've given the model this additional info that I can call a tool to do something like get weather. Now it can use that, make that call, get the weather, and then respond accordingly. It's a way for the chat effectively to chat with itself in order to do additional things like access data externally, call an API, apply a transformer of some form. Why am I talking about all of this? The reason is the tools that are available. A lot of the system prompts for these things have a list of the tools that they can access. So it might be tool list. Maybe we have one that is text application, input format, z dot validate, whatever, just some type of validator. And description allows for application of text at a specific spot in an image. So that's one of the tools that they almost certainly have implemented. I haven't had that hard confirmed, but we've effectively confirmed it at this point. There's also a reflection tool. The reflection tool allows it to take something that already exists in the image and collect it onto a different service. They have a bunch of these tools built in to the 4.0 image gen that are what makes it so capable. It's also probably an analyze tool. Analyze chunk, analyze pixels, or probably more likely find matching chunks that allows them to say, where are these things? And it will respond with the pixel locations for those things. Then it can take those pixel locations and say, okay, put that through the text application. So if we took one of these images that I generated, like one of the comics, let's say, because those are a pretty solid starting point. So we take one of these comics, we tell it to generate four panels, matching these four pieces for the visual description for each. We tell it to leave room for a text box of a certain size. Then we would ask it, where are the text boxes? So you can imagine the model getting a prompt like the user sends this. Now the model says, okay, let's generate it. So first it calls tool scaffold. And this probably generates a minimal pass shape for image probably would just generate the square like this, so that it knows what goes where. Then once it has this shape for it, this scaffold, possibly even doing things like diffusion in the middle boxes, unlikely, but possible, it uses the instructions to figure out how to shape the output structure. And it gets back the locations. Okay, you have four panels located at x, whatever, y, whatever. Now it says, I don't know, fill in panel one with blank, then fill in panel two with blank. It keeps going through all of those until it gets to the end, where it says, find the locations of all the text blocks, or all of the dialogue boxes. Then it says, fill in dialogue box one with blank, fill in dialogue box two with blank. The point I'm trying to make here is effectively, this model is special, because it's kind of reasoning with the image, the same way reasoning works for the O models, where it asks itself questions and thinks its stuff through by double and triple checking every step along the way, it's able to generate better outputs. And the tool calls being built into the image model, where it can ask itself questions about the image, get data back from the models, and then apply things on top. That's really powerful. I did just get a really good question, which is how do the fill in steps work? More traditional image gen for most of those probably, or if you better understand the VAR stuff than I do, you probably give a better answer to that part. I honestly don't fully understand how the VAR stuff differs from diffusion so meaningfully that I can have outputs that are this much higher quality. What I personally am more interested in is the tool call part, because something like fill in dialog box one with blank, very easy to do once you have all the scaffolding set up for it. You go to something like image flip, this is literally just pixel locations for three text boxes. And I can just write sup nerds, please subscribe, and it will apply that in the right location, you can tell it to transform like rotate or skew. If I have an image like a whiteboard that is slightly angled, let's take this one. Perfect. This image is at an angle. Paste this in to my photo editor of choice. And I want text that fits the scale. I can write the text on it. Sup nerds, don't forget to subscribe. If I just take this and put it there, even if I make it smaller, it doesn't look great. Obviously, that's because I'm not using a good font, but even with the better font, still doesn't look wonderful. How do I fix that? Well, there's a wonderful tool in most graphic software, the perspective tool, perspective tool lets me warp based on the perspective. And all this does, it takes those corners and gives the different pixel location values, it updates all the pixels at scale based on that. And it's hard to get just right eyeballing it, but you can get pretty close. And now it looks almost like this text is on that whiteboard. The way that these tools work, the way that the OpenAI image gen stuff here is able to apply text accurately, is it's using a tool like this, possibly even just using HTML and CSS stuff. It's probably not, but it could very well be. You could hopefully, if you're a dev like me that does graphic stuff, imagine how you would build a tool that does this, that given a payload of, I don't know, tool, render text, content, body, sup nerds, coordinates, 00, 0, 10, 5, 3, 5, 8. These coordinates aren't going to make a perfect rectangle. Instead, those coordinates are going to make something that is skewed where the left side is taller than the right side is. And you could imagine how you could write some CSS or transform logic that would allow this to render the text at this location with the right skewing. You could even then apply a filter on top. And I'm sure that they do that a lot. I noticed with my thumbnail for the last video, actually, I hop back over to chat GPT, it added a yellow filter to this image. And a lot of people noticed that the reason the yellow filter applies to the whole image and why it got this part at the bottom here wrong is it is applying these tools constantly. And the tool that applied for the filter probably got applied at the end. The same reason in the previous video, I noticed that the color tone and style of the image changed as it was going down. The reason for that is at some step during the image generation process, it applied a tool for color correction, which corrected all of the pixels on the image. It was also applying a tool for the text for rendering the, like the replies in the retweet icons here. And it probably accidentally called this tool twice or hallucinated it needed to be in two locations, which is why it ended up doing that. Chances are the reason the output's been so consistent is they have a bunch of additional tools at the end, like make sure hands have five fingers or make sure animals are anatomically correct. These types of tools are allowing them to apply step by step things that are specific to improve the image. It isn't too far from diffusion in that way, where it's rerunning over and over again. Diffusion works by taking a given image and transforming it based on whatever your request is, often over and over again until it gets really close to what it's supposed to be. Now, instead of each of those passes being the same thing, another refinement over and over, those additional passes are tools running to make necessary changes. There's probably even a tool because remember, OpenAI can parse and process images, summarize them, tell you things about them. There's probably a tool where is there anything wrong with this image, where it will prompt itself and ask and then make corrections accordingly. The idea of image gen stuff, having a large set of these tools that allow it to transform itself, update itself, apply text correctly, all of these parts. This is a real innovation in how AI image generation works. There have been pieces of this in lots of other solutions, but none that feels so complete, like a reasonable person could drop in an image, ask it to do something and get a response. That's good. It's unbelievable. It's at that point. And we go back to my new chats, very well labeled clicking on them fixes the labels. Hilarious. You can see, give me a new profile picture here. Other than that, it actually did really good. This looks great. It's pretty nuts. If you think about it, that you can just hand it a UI, tell it to make a change, and send this over to your design team. A lot of these flows are what I'm really excited about with AI. I've seen a lot of people saying that designers are the future and are going to take over and programmers aren't going to be necessary anymore. I don't agree with that. More importantly though, I think it goes both ways. Where on one hand, now I, as a software developer, I can get a mock that is pretty damn good relatively easily. But on the other hand, I, as a designer, can make a mock version of my app to give to the product team and start sandboxing with users. I can't tell you how awesome it was having a designer back in the day that knew exactly enough HTML and CSS to make a crappy, fully not functional, client-only mock of a new experience that she was trying to build for our users at Twitch. And then she could go to actual power users of Twitch, put it in front of them, have them play with it, break it, ask questions, and figure out what flows and expectations they have with the UI and refine it for it even becomes my problem as an engineer. And it goes the other way too. If my team or my product manager, my CEO or myself has something they want to add to the UI, we're not quite sure how to do it. Rather than bother design way too early in the process, I can go mock it out with these AI tools, send it over to them and be like, hey, this is roughly what I have in mind. It doesn't mean that we have to ship the AI generated version and fire our designer. It just cuts down iterations meaningfully. It's the same deal with the thumbnail stuff I was showing earlier. This Microsoft OpenAI back and forth that I faked was a thumbnail that we wanted to try. It was just one of a few. Most of my videos end up with three to five different thumbnails that we make in the process. We ship one to three of them with A-B testing. This one ended up losing, which was really surprising to me. Here, if we hop over to this video, we can look at the results for this test. And the one that Ben did with the heart split with the two logos performed significantly better than the one without. Normally, the gap I'll see between two thumbnails is like two-ish percent or less. That is a much bigger gap. Much bigger. Obviously, part of it is that my face is in it. I didn't do a version of this with my face in it. I was just playing with this. But the fact that I got to try this thing, feel out this idea, and watch it perform less well gives us so much useful information. It's so powerful for us as a team of people that's small, trying many different things. The biggest thing we would lose as a small team is we don't get to experiment as much because our effort has to be very carefully placed because we do the wrong thing. You lost that time. There isn't somebody else in parallel doing the right thing to make sure that we are smoothed out. Here, we could have Ben making a real traditional thumbnail and I making a theoretical experimental one on the side, push out both of these thumbnails and see what the results are. That's so cool. I think A-B testing in general is going to become one of these essential pieces of why AI gets useful. It's not because I can make something as good or better with AI than I could have built otherwise. It's because I can make five versions of a thing, figure out which resonates the best, and then refine the right thing. I can't tell you how many years I've spent polishing turds just trying to fix things that nobody actually wanted. If I could try five different things and figure out which one people actually want, so much of that time, so much of that effort is no longer being wasted. This is the first time I feel like these image generation tools aren't wasting my time. They're both saving it immediately, obviously, by generating a result, but they're letting me experiment more and try things that would have been too much work otherwise. There are tools I can use to fake DMs, but I would have to go find one, download the logos, upload the logos into the individual hidden boxes for it, put in the fake text, set up the color correction, realize that the browser app I'm using doesn't let me get the colors the way I want. So I would have to go into the browser, like inspect elements, change the colors manually, paste it into my Photoshop tool, realize that it's too small pixel-wise, go back in, press command plus to zoom it in, realize that breaks the layout of the page entirely, rage a whole bunch, and then just give up. You can't tell I've done this a couple times. It's so cool that all of those steps are skipped and you don't need the random knowledge that I have. I've even built small picture management tools to solve a handful of problems that have annoyed me, like QuickPick. The point of QuickPick was to make it easy to do a couple annoying tasks, like turn an SVG to a PNG. Part of this is because if I had an SVG of the Microsoft logo, that probably wouldn't work in a lot of these generation tools I was using to make a fake chat UI. This also is very helpful as a video editor because I have SVGs I want to throw into my videos, but most video editing software doesn't support SVG files. I was also annoyed because there was a lot of tools that did this that cost money. This tool, AI is not going to help with a whole bunch. I mean, it helped me build it, but other than that, yeah. The other ones though, like the square image generator, the corner rounder, or more importantly, PickThing, which is our tool for removing backgrounds and managing your thumbnail assets as a creator. We hop over to my dashboard, you'll see all of my thumbnail faces that I use for all of my videos. This was a bespoke tool that I built because it was too annoying to do these types of things in the background removal that existed in most apps was kind of fucking garbage. So I built this to have a quicker way to find the thing I want. Click copy, hop over here, click paste, done. And it's super handy. But if I wanted to put myself in the whiteboard, now I have to go make an obnoxious crop layer with a mask. And then as soon as I want to move things around, it all breaks. Now I can just tell the AI to do it, I can take this image, even with other background removed, I can then take that whiteboard photo, save image, whiteboard demo, up over to chat GPT. I am so excited to have this in t3 chat, you guys have no idea. Put the person inside of the whiteboard. You're looking for getting my hair changed. Yeah, there will be a lot of old Theo pictures throughout all of this that we'll have to get used to. That's fun. It adjusted the whiteboard, but it did my goal of putting it in there. It kind of fucked me up in the process. Markiplier Theo can't hurt you. He isn't real. Yeah. Okay, maybe take that part back a bit. There are still specific, complex enough transformations that it struggles with. What it can do, and I showed this in the previous videos, it can do transparent backgrounds, which has been super helpful. But as we see here, it's not perfect. It's still far from perfect. But it's a hell of a lot better than I ever expected it to be. And that's why I wanted to do a second video about a thing, which I never do. I hate doing second follow up videos so close to the original like this. But I felt like there was nuance and value in the new image gen stuff that I hadn't properly communicated before. And I am genuinely really, really excited about it. I have been using it a ton. And obviously, I just absolutely cannot wait to get this all into t3 chat. So let me know what you guys think. You're more interested in my hair than you are in this AI gen stuff. Just let me know in the comments. I am really, really curious to figure out how to cover these things going forward. Until next time, peace nerds.
