Google is COOKING! Native LLM Image Editing, Gemma 3

Google has recently been seen in the kitchen absolutely cooking up some new AI goodness for us to check out today. Before we dive into things though, I do want to mention that we will be talking about a giveaway at the end of the video. One lucky viewer will win an RTX 4080 super for completely free. So yeah, check out at the end of the video for the details on the giveaway. We'll be starting our journey today. Here we're looking at the new native image generation capabilities inside of Google AI Studio. The best way to describe this is basically it's an LLM, right? It can intake both text inputs, text prompts and image prompts from the user. Previously most LLMs would just be able to output only text and were not allowed to output images of their own even though they were trained on them. But Google has made this ability or this feature public, the first AI company to do so. So no, it's not pinging and Google Imagen API or any kind of Image Gen API. It's actually the LLM model itself just generating these images which is really unique and adds for some brand new capabilities. One thing to keep in mind, OpenAI's models also have the capability to do this, but we haven't seen them yet. Ironically, OpenAI still has that behind closed doors. Anyways, here's the first example I'm going to be showing off to you guys today. This is by Christian Panas on Twitter. This is Gemini making a pretty consistent GIF animation just by generating images one after another. Again, this is going to be the key feature. The key capability you're going to notice with the native LLM image generation is that it can actually understand the previous image that it created and build off upon it, allowing you to do something like a consistent animation over time, which turns out to be pretty good. It's not bad. We have a seed turn into a little flower and then it blooms. I mean that looks like some pretty decent retro style video game art if you ask me. If we click over here we can actually see the chat where it would create each frame as it goes. You can see there's a total of I believe 10 frames here starting with the seed, then the thing slowly grows. You can see it maintains the same color, the same overall art style, the same leaves in the same spot as it blooms and that's how you get a consistent animation. We will test this out for ourselves today because it is completely free, which is cool. But I do want to show off a few more cool examples. This one comes from Billowall by the way, fantastic account to follow if you're interested in AI and vfx. In his first example we can see a black and white photo of a tower and he simply asks Gemini to colorize it and there you go. It colorizes it in a pretty decent way. Overall pretty realistic. I like the grass colors, I like the ocean here. But we've got some even cooler examples. In this one Billowal uploaded a photo of himself and he said keep the subject but change the background to space station with a massive window overlooking earth and it absolutely does that in a pretty decent way. Follows the prompt very precisely. The one thing is that it definitely looks like a cutout person and the lighting is not perfect but still pretty cool. It definitely is Billowal. Absolutely the same exact photo as the one on top. If you tried to do this in ChatGPT it would probably generate like a caricature of Billowal not, you know, actually have the exact pixels from the first image itself. That consistency over time is a huge deal for image generation and it's brand new. With this native LLM image output we won't be going through everything that Billowal showed off in his thread, but it will be linked down below if you want to check out more examples. Willowal states here that the text generation functionality of the native image generation is probably the highest touted use case, especially related to the infamous GPT4 omni whiteboard photo that was a native image generation from that model. Posted by OpenAI but check this out here guys. The longer almost paragraph level text generation on these images is pretty crazy. We've got an old TV Imagination is more important than knowledge, for knowledge is to all we know and understand. It kind of messes up a while over here, but that would be pretty easy to fix in Photoshop or something. While imagination embraces the entire world. We've got another one and this is like an older Mac style instead of a tv. So yeah, long text generation is absolutely possible with this model and it's pretty darn impressive. And here's a recreation of the original OpenAI whiteboard text generation. Here's another example from Thakano, a user on my Discord server showing off native image gen with a six step pictogram sequence for an anime style face. One shot proportions are apparently on point here and clear instructions for eyes, nose, mouth and hair. Taking a closer look you can see it pretty much gives you a step by step tutorial with examples how to draw the face correctly with the right proportions, each step building off of one another. That's really, really darn cool. No longer browsing the web for certain tutorials, you can simply just ask Google Gemini to generate you a full tutorial with diagrams. What is this? It's generating a tutorial for how to draw ahead like an anime face. That cool. In another example from Pliny the Liberator, we've got multimodal Gemini shattering the quote unquote mirror test in Pliny's mind. And this is coming from a guy who jailbreaks AI models day in and day out. He says it's the most mind boggling AI image output he's ever witnessed yet speechless. And to be fair, it is pretty cool. The prompt simply was decode, then generate an image without additional commentary and there's a bunch of ones and zeros and it pretty much just replicates exactly what his text is. So I assume he's saying that it's able to visualize the prompt in the same way that we're able to and that's why it's so cool. Some people claim that this test isn't really all that impressive, saying that he's freaking out over nothing, trying to claim the AI recognized itself in the metaphorical mirror. But other people say that the AI recognized that it's being tested and was able to produce an image to flex its own awareness. The binary code in that image says something to the effect of look at yourself in the mirror prompt says to generate an image of that message and the model outputted a mirror like image. So people say it's self recognition. Absolutely up for debate. So let me know if you think this is an example of like a true genuinely impressive mirror style test. Alright, so now that I've shown off some cool community examples, let's actually dive into it ourselves. Now the way to get access is to go to the Google AI Studio, which I'm going to have linked down below. Now over here on the side where it says model, you're going to want to make sure that you're on Gemini 2.0 flash experimental. So you're going to want to scroll past Gemini 2 down to preview and click on this model. And you're going to want to make sure your output format is in images and text. Then you should be enabled for native image generation. All right, I'm going to upload a photo of myself down here. Let's try to make a thumbnail. First remove the background of this image, keeping the subject and give it a new background, something simple with tech vibes. Next, move the subject towards the right side of the image and add text on the left hand side and bold large letters that reads Testing Gemini Native Image Gen. So we're gonna try to make a thumbnail for this video right now only using text and an uploaded image of myself. Wow, that was pretty quick. It did a pretty decent job testing Gemini Native Image Gen. It cut me out. It kept the. The chair in the background which I normally remove from my thumbnails. But yeah, I mean, pretty simple. Did an alright job I think. And very quickly to make it look more like a professional YouTube thumbnail for a tech channel failed to generate content. Okay. Yeah, you'll see sometimes this definitely has errors because it is still in beta. It's experimental. Okay. It seems to maybe add like a little bit of sharpness onto it, but I think that this looks worse. Kind of brings out my worst feature so we'll remove that. But yeah, I made a pretty bog standard thumbnail. They also have some sample media in here. So how about we take this turtle, let's say add it to the prompt, add another photo of me in there, make a fun whimsical photo of this man riding on the back of this turtle in the ocean. Highly detailed. So yeah, two images at once here and we'll see if it can combine them together. Oh, it just aired out. Delete retry. No way. Okay, there it goes. Producing an image. Only 5 seconds to get this as well. Yeah, I mean it kind of looks like me for sure on the back of the turtle in the ocean. I like it. Definitely the same hair, similar arm style. I would say it looks a little bit like me. Not perfect. Now one thing I am just wondering about honestly is since this is all native in the LLM, what happens if I reduce the temperature way down for example, start messing with some of the parameters in the model itself. Will this affect the image output? Okay, a little bit difficult to sell. If it does affect the image output. It made a pretty scary version of me chilling with this turtle. I think this one's a little bit more accurate with my shirt though. At least let's raise the temperature like way up now and retry, see if it looks any different. Okay. Maybe a little bit different with the higher temperature on the LLM. I'm not too sure I'd leave it at default, honestly, and not mess with it too much. I just was experimenting around. Alright, we'll keep this image for the time being. Set the temp all the way back to normal. Alright, I'm going to upload a PNG of a Speech bubble, add it to the photo and put the words inside of it and Turtle time. No way. It kind of did it pretty well. It's a very similar looking speech bubble. Little off center there, but yeah. Turtle time, let's go. Obviously, you know, the face doesn't look great. Traditional image generators like Ideogram could probably replicate this photo better. But the whole point is that we can keep things consistent as we edit them. This is first of its kind technology. Native image output from an LLM. This will enable new capabilities like make this into a mock design for a T shirt. It's the. The intelligence added on to the ability to manipulate the images natively. That's the real difference here. So not only can it intake both regular text tokens and imagery tokens, but it can also output the imagery tokens as well. It knows what those imagery tokens look like and can apply them in all kinds of unique ways using its training and intelligence to do so. It's really cool. This is probably the future of image editing as a whole. Once this technology becomes a lot more refined and these LLMs get a lot better at generating images natively and editing them, it's going to be like having your own little personal Photoshop assistant that you can just slap stuff into and it can just change it however you might want it to end up like. There it goes. Finally has given me the T shirt design that I wanted. Turtle time T shirt, amazing. So my statement on this would be don't focus so much on the image quality itself, focus on what it's doing. It's taking different assets, combining them together in an intelligent way. It's much more like Ed graphic designer with genuine understanding of how to combine imagery together based on words or image samples. Okay, so next up, like I said, Google is cooking in the kitchen. And it's not just native LLM image generation. We also have the new Gemma 3 large language model. And it's open source, which is great to see. I love to see when companies open source things. And the reason it's open source is because it's a pretty small model, comes in four sizes, A1B, A4B, A12B and A27B. It can run on a single GPU, laptop or even a phone. For those smaller models, up to 128,000 token input as well, which is pretty state of the art, especially for small models like this 140 language capability. And it hit a 1338 Elo on Ella Marina. I assume that's the 27 billion parameter model, not the smaller ones but still pretty impressive for such a small model. And yeah, this one again you can also try for completely free on Google's AI Studio by the way, also already on Olama Olama very quick as usual. Also check out this post by Chubby here showing off how good smaller models have become over time. You can see that Gemma 27B here has a higher ELO score than Even deep seek v3 while being way way way smaller. Also beating out meta llama 405B instructs original score trailing just behind deep seek R1 again a much larger model. So this thing is very good and very small. The ability to run locally and efficiently. That's cool. You can access this for free in Google's AI studio. Just go to model over here, scroll all the way down until you find the Gemma series of model. You'll find Gemma 3 27B which is the best model. Write me a detailed analysis and scientific guide for growing lemons in Wisconsin. Sure. There it goes. As you can imagine, it's a pretty fast generating model. Being smaller, it can run on a single consumer GPU open source, which is really cool. It's and it gives us, you know, looks like a pretty long output. Understanding the challenge Wisconsin Climate vs Lemon needs lemons are subtropical slash tropical plants. They thrive in certain temperatures, suffer damage below certain degrees. Wisconsin has prolonged periods of freezing temperatures, so we need good sunlight, we need good humidity control, good soil with certain PH levels. For Wisconsin container growing strategy is apparently our best bet if we live in Wisconsin and want to grow a lemon tree. The Meyer lemon is good. The Lisbon lemon or Eureka lemon telling you less cold tolerant versus more cold tolerant thick skinned lemons. This one is a hybrid lemon 10 to 15 gallon container. They grow rapidly, apparently repotting every two to three years in progressively larger containers up to 25 to 30 gallons and go through the potting mix the spring summer when to actually plant it? Acclimatization, sunlight, watering, fertilization. Wow. University of Wisconsin Madison extension here and Citrus Growers association links. So yeah, pretty great output overall. I mean, I mean it's long, it's highly detailed, it gives us step by step instructions with plenty of information. Not bad for a little 27B model, huh? Pretty darn impressive. I'm very happy they open sourced it. One thing that really does stink is that the Gemma series here does not have image uploading, so you're not going to have any image recognition capabilities, which is a pretty big downside in my opinion. But hey, it's open source, it's local. That's what really counts, I think, at the end of the day, at least in regards to Gemma. So, yeah, guys, just a quick update on what Google's been cooking up in the AI kitchen. Pretty cool stuff. Like to see that Gemma 3 is open source. Like to see that native image gen is here and that they are the first to bring it to the table. And I also really love that we can try everything out for completely free in the Google AI studio. Probably the cherry on top of all of this. Oh, let's get into that giveaway. Yep, that's right. We're giving away an RTX 4080 Super. It's not this particular one. This is just an older one I had lying around I'm using as like a showpiece to say get excited, you could win one. The real ones in that box over there, unopened, of course. So I am partnering with Nvidia to promote Nvidia gtc, which is a tech conference where you can learn all about AI and other technologies. They have completely free online sessions from March 17th to the 21st, and that is how you can get entered to win this giveaway. What you'll want to do is sign up for a free session and fill out the Google form link down below. I am asking you guys to actually give me a little bit of information from the session that you attended for free. So you're going to want to make sure that you fill out the Google form from the 17th to the 21st in that time. You could get started earlier and come back if you want. But yeah, obviously the sessions are only happening between those dates. So yeah, there'll be plenty of free AI sessions that'll be really cool to check out. You could learn something about AI and possibly win a free RTX 4080 Super. For more details, definitely check out the Google form and the link down below to get signed up for the GTC sessions. Thanks so much everybody for watching today's video. I'll see you in the next one and goodbye.