MAJOR Upgrades to AI Art Tools! & Open AI's new Image Gen!

What's up guys? Today we're doing an AI image news roundup. We're gonna get a little bit hands on, especially with Ideogram Canvas, and then we're gonna do a decently deep dive into the rest of some of the image generation news that has been circulating. If you guys have been following AI at all this week, you know it's been totally insane. But if you want to stay up to date, I suggest joining my Discord server. It is an awesome community and plus they are constantly up to date and on. The latest and greatest in some of the most impressive generations I've ever seen have come from the Discord server. We'll put it that way. Also tomorrow I'm going to do a large AI news roundup. I'm going to be talking about everything that has happened this week. Try to catch you guys up. It's been crazy, as I said. Anyways, let's get into things. So the first big piece of image gen news here is Ideogram Canvas. Now you guys know I'm a pretty big fan of Ideogram AI. It's been my go to pretty much since it's came out. I've always liked their UI design. The models are top notch and it's always been easy to use and interact with. They've just now added another way to interact and use their Ideogram models and that's their Canvas feature. This is a pretty big deal because we've seen Canvas style image generation UIs in the past, but this has got to be the most fully fledged and actually useful one that I have seen. Now can it replace something like Photoshop? Maybe not yet, but let me show you what it can do. Let's just start off here by generating a new image medium shot of a cat dressed up in a Halloween costume. Let's say it's a mummy Halloween costume. The background is a typical suburban street on a Halloween night. Sure. And as you can see, like typical Ideogram, it makes you four generations. But you'll notice if you scroll the wheel for example, or hold down middle click, you can actually drag and move around on a larger canvas. And you can also drag and move your images around. So we have four to pick from. You know what, I really like this one down here. So I don't need any of these other ones. I could just press the delete key to now remove them all. And there we go. We have this image and we can actually still get to our other images just by clicking these buttons right here. So they aren't gone forever or anything like that. And if I hold control and scroll, I can also zoom in to see my image more up close. Now, this is a nice cat and all, but I want to see more of his surroundings. So we're going to use the Extend tool right over here. And this is a big part of the new Canvas feature. This allows us to essentially out paint any image we want. So let's change the aspect ratio to, like, something a little bit wider, like a 3x2. We'll drag that right over here. We can also just expand it by dragging it like this. So we'll do it like that. And for the prompt, we'll just say, a typical suburban street on a Halloween night. Now, what you're gonna want to make sure you do when you use the Extend feature is turn magic prompt off. Because we don't want it to infer anything. No, we just want the typical suburban street. Don't infer anything extra. It might try to add extra details that we don't want in there. And we'll click extend. And there we go. It extends it pretty darn seamlessly. I mean, it's pretty much perfect. So that's already pretty incredible. You can see it actually copied the original image of the cat. So we do have that original image, but we could basically remove it now if we so happen to want to, because we have the larger extended image. And it also gives you four options yet again, so you can pick which one you find best. I think I'll just stick with this one. This one's pretty good. Now, I want to give the cat a speech bubble. He should be saying something. And that's where the Magic fill tool comes in. And there's a bunch of different ways you can use this. There's a brush mask, which is essentially like a paintbrush, a lasso mask, which is basically just like a free draw, or a rectangle, which is, you know, something like this, for example, where you can select and highlight just a rectangle. Oh, yeah. By the way, Ctrl Z, as typical with a lot of editing software to undo. Let's use the lasso mask, though. I'm going to highlight this portion of the screen, and then we'll click. Next we'll just say speech bubble. Speech. Text reads. Speech bubble reads. Dang. These people are giving away candy, not cat treats. That's a little bit long. I don't know if it's gonna be able to do this. Perfect. Let's risk it. Oh, it's trying. Okay. Actually, it did pretty good there. And you can see it made the little speech bubble. So yeah, that's like, that's darn impressive. I mean, this is almost a replacement for Photoshop. It's all AI focused editing tools and they work really well. I'm really impressed by how good the extensions are and how good the magic fill is. It's crazy. And obviously you still have like remixing for example, and there are some other tools as well, like image uploading. So in this example I prepared earlier, I already have myself uploaded in here and we can do things like extend, so we can say, you know, extremely muscular torso of a man and you can extend yourself. Now keep in mind the extend feature is very similar to something that we already have available in Photoshop, but honestly, I think it might work a little bit better than the Photoshop one. Ideogram knows what it's doing. However, Photoshop has a lot of traditional tools that are pretty difficult to beat. So for thumbnail creation, for me, Exit, for example, I still do use Photoshop, but I don't know, this is getting pretty darn close to me just not needing Photoshop anymore. And hey, it made me an extremely muscular body. Look at that. For some reason it always wants to give me tattoos, which I've never understood. But let's continue to extend. You pretty much just use it willy nilly. Something like this should work fine. I think you can even just have no prompt at all. Oh no, the prompt cannot be empty. That's a difference between this and Photoshop for sure. In Photoshop you can just have it fill in general areas just by guessing. Just gonna say background and see what happens. I'm just trying to sort of fill out the scene. Oh, and see, this is where prompting becomes very important because I said just background and then it kind of did a little bit of filling, but now it's just making other people as well. This one's a little bit strange, this one's probably the closest match. But yeah, the prompting is very, very important and just filling things in can be a little bit difficult. With Ideogram Canvas, you gotta be specific. Anyways. Yeah, no, this is, this is a nice image, right? This is actually what I look like. Guys, just so you're aware, here's another extended image of me that I also did earlier and you could see the starting image was this. But I was able to add lemon themed headphones, which obviously don't exist. I was able to extend the background pretty decently. Gave it a weird desk with some weird stuff on it though. And I was also able to make myself holding a lemon. Overall, pretty impressive. You can do a lot of really, really detailed, small, fine tuned touches with Canvas and the user interface is nice and easy to use. I don't mind working with it. The only thing I don't like is that maybe it's a little bit laggy, you know, a little bit lower frame rate from, from time to time, but that's something that can be fixed and everything is automatically saved to the cloud, which is nice. I also did another Halloween cat image and I started with just a smaller cat wearing like a witch's hat or something like that. And then I expended all these pumpkins out and this beautiful scene in the background and I thought that this was also particularly nice. So yeah, that's a Ideogram canvas. It's pretty awesome. It's a big step. It is absolutely by far the best canvas style editor I've ever seen come out of any image generation company. And if I remember, the first actual canvas style editor was I think, Playground AI and theirs wasn't bad. But the technology has advanced enough and Ideogram's models are good enough that their Magic Fill and Extend work really, really well. Tough to beat the quality of the models that they've done. I assume it's a different model for Magic Fell. Assume it's a different model for Extend. Not exactly sure how it all works under the hood, but they did a good job. It's very seamless. Midjourney also has some brand new updates as well. They're testing two new features. They have a new image editor for uploaded images. So this is basically inpainting. I don't think they have out painting like we just saw with Ideogram Canvas though. But the inpainting, at least from their demos, looks really, really good. And they're also adding something called image retexturing, which allows you to explore different materials, surfacing and lighting. And this all works with the rest of their advanced features like style reference, character reference and personalized models, which is really nice to see. I'm not a huge Midjourney user. I used to use it a lot more. Haven't been subscribed in a while, to be honest with you, but this is something that I would consider resubscribing for, especially the retexturing. They've got a brush size, you can upload a photo of your dog or anyone really anything you want, remove a part of the image, erase it, just like we saw in Ideogram Canvas, then type a prompt in midjourney. Will Go ahead and actually generate and fill that out. Takes a little bit, but there you go. A nice little party hat on the dog. And it, it looks seamless. It looks good again though, I can't help but think, you know, I could literally just go do that right now with the diogram canvas and it would look just as good, probably. And of course, you still have four variations to pick from. Again, pretty standard with all image generators, but yeah, better inpainting right on their website, which is really nice. Now they're gonna do something more complex. They're going to erase the body of the dog and then give him a colorful sweater. And this also turns out really nice. Now this impressed me a little bit more because it's remembering what the body of the dog is supposed to look like. Or it's just really good at guessing what the body of the dog looked like prior because the sweater looks good, it fits very well, and it looks like it's on the same exact dog that we saw before. But remember, we erased it and then regenerated it. Don't know if it's looking back at the initial image or if it's just good at guessing one of those two. So now he's showing the retexturing. He's uploading a real photo of the interior of a house, but he wants to customize it so he erases a large portion of the original picture and extends the canvas. Okay, so they do have canvas extension in midjourney as well, but it seems to be a little bit more locked than the freeform canvas. That ideogram has different approach, but I think this one might be a little bit easier for new users to grasp, which is good. And then bam. Mid century modern home. It goes and renders that out and it looks great. We got a rug, a chair. Looks like it could have been there the whole time. I'm pretty impressed by that. And they obviously have a bunch of different generations to pick from and all kinds of different, different homes. I could actually see this being genuinely useful for interior design work. Anyways, he takes this one right, and then he's gonna go ahead and use retexture Midjourney does this thing where I'll play this animation. It actually is able to scan the various details of the room or of the image. Don't know what kind of technology they're using under the hood to make this happen, but it's pretty darn cool. And I haven't seen anything quite like it. So he's going to retexture it here it's a living room from the far future. And with that retexture, you can see that it completely changes the environment. It keeps the couches, the books and everything where it's supposed to be though. So all of the critical details that were in the original image, it's hanging onto and just changing it to something that would be from the future, for example. It's really cool. Like I said to me, this is more exciting because I haven't seen anything quite like this. It captures so much of that detail in and is able to faithfully recreate it with whatever text prompt you put in. It's really, really cool. It also changed, I think, the art style a little bit with this one, which I found particularly interesting. So it kind of moved away from the real life photo scenario into like more of a drawing. So that's like style transfer. Pretty cool. So, yeah, I mean, if you're already a regular Midjourney user, this is pretty great. If you like the features that Midjourney has over Ideogram, for example, this is great because Ideogram just introduced canv and this is very similar tools to that. There might be some slight differences between the two, but I think this is great. Both Ideogram and Mid Journey are really starting to evolve into quite useful tools. Like I said, I have personal. Personally a fanboy for Ideogram, not gonna lie. But I. I think this is awesome for Midjourney. And like I said, I might buy a Mid Journey subscription again just to try the retexturing for the initial release. It's only open though, these features for people with yearly memberships. And I'm definitely not buying a yearly membership. So to actually test this out for myself, looks like I'm going to have to actually wait. They've got some pretty cool examples though. We've got someone running down the street. And then you can add some colorful streaks. And also remove the annoying person who photobombed the image. It's pretty nice. That worked out very well. You can't tell that she was ever even there. We've got a nice picture of a tower, for example. Expanded it to the side. Although I have to say the expansion over here lacks a little bit of detail. Not great over there, especially this portion. But the whale looks really cool. Like, I like that. That's awesome. We've got a nice big bean in the middle of the train station, which is awesome. This, I think, is his Grand Central Station. It's gotta be, right? But yeah, we've got the bean in there, which is good. This was really cool. This is from Creative Gravity. He took a real sketch that he made and then turned it into an image using the new Midjourney tools and I think that's just awesome. It really did a fantastic job at converting it. That's so cool. In other AI image generation news, stable diffusion 3.5 large dropped so stability AI finally introduced an update to the not too well received Stable Diffusion three, and it seems to have redeemed itself somewhat. It seems to be a decent model now. Can it compete with Flux? Is the the question here? I honestly really don't know. A lot of these images seem to be of good quality, but it doesn't look like anything that couldn't be generated with Flux. And Flux is very good, but it's a little bit more difficult to fine tune and train apparently. And stable diffusion 3.5 is easier to do that with, so there might be some sort of a divide in that regard. Either way, it's another good model. It's definitely a lot more open than Midjourney or Ideogram, let's say, which is great for the community. It's free for research, non commercial and commercial use for organizations or individuals with less than a million dollars in total annual revenue, which is good. That's good. I think that's pretty reasonable. As some of the community members on the Reddit point out, Stable Diffusion 3 was built to support Lora's control nets, IP adapters and fine tuning right out of the box. And this is why someone might choose this new Stable diffusion model over something like Flux. They've also got a more distilled turbo version of the model, which generates images in just four steps, which is pretty good. I don't know. From the examples they're showing off, it seems to be pretty good with things like realism. So is Flux. It does decent text, decent prompt adherence, like, I don't know. According to the benchmarks here, it seems like Flux does Dev has a much higher aesthetic quality, but the prompt adherence is slightly worse than SD3.5 large. It also says that SD 3.0 large though is about right in line with Flux.1 Snell and a lot of people would probably disagree with that. I don't know. With these image and art generation models, it really seems to be something of user preference, what you want to actually end up doing doing with the model at hand and what your goals are for your end results. It really can depend all on those specific use cases. OpenAI actually has some image generation news as well. They're introducing SCMs, their latest consistency models with a simplified formulation, improved training, stability and scalability. It says image generators that generate samples comparable to leading diffusion models but require only two sampling steps. Only two steps to get this right, which is pretty crazy. The comparison on the OpenAI page here is pretty telling. You can see that the image generated by the consistency model happens quite a lot faster than the diffusion model, but the quality on the diffusion model, at least to my eyes, looks a little bit better than the consistency model. This is really more of a research paper than anything else, but I'm really happy to see OpenAI actually releasing some of their research and explaining to people how it's done. Now are we getting open source model releases? No, but at least this is somewhat beneficial to the community. The scaling is also really cool to see and I'm always surprised by what OpenAI gets up to. You know, just when you think they're done with image generation like Dall E3 happened and they're they're chilling with Dall E3, it's in chat, GPT, etc. Nope, they're still pursuing new image generation technologies. Anyways, I'm not going to get too far into the weeds with this. All you really need to know is that OpenAI is still working on image generation and they have a new type of model that generates faster than diffusion. Their largest model, 1.5 billion params, generates a single sample in just 0.11 seconds on a single A100. Absolutely crazy. And folks, that is my little AI image generation roundup for today. So stay tuned tomorrow for the larger news drop. We're going to be talking about some really cool stuff that has happened this there's plenty that I still have yet to talk about and I'm going to recover some of the other things I spoke about this week in that video tomorrow. But if you're looking for the deeper dives, obviously check out the videos I've already posted. Thank you so much for watching everyone. Don't forget to join the Discord and join the rest of the Mavid Pro community. It is awesome. If you really want to stay up to date in the world of AI, you want to stay on the ball on top of things. Check the Discord out because as soon as anything drops, it's posted in there. See you in the next one. Goodbye.