FEEL the Acceleration! Image Gen, Consistent AI Video, Open Source LLMs & WAY MORE!

Hey folks, welcome back to another video where it appears I desperately need a haircut. And I do. You'd be correct. We have so much to talk about. This week started off kind of slow, pretty regularly paced and then everything just jumped crazy at the end of yesterday. I have like 20 things to show you today and a lot of them are really awesome. Stay tuned at the end of the video because today is the last day of the RTX 4080 Super Giveaway. Someone is going to win that for free. It could be you. We don't have any time to waste though. Let's dive in. First up, we've got a brand new new text to image model. This model appeared on some benchmarks known as Half Moon. It was very mysterious. It seemed to be very good. Excelling at typography, prompt adherence and overall aesthetics. But the mysterious Half Moon model has been revealed as Reeve image 1.0. You can see the legendary T? Bor has brought us a couple of examples. The model really does appear to be absolutely state of the art at first glance. We've got a pretty decent drawing of T? Bor himself with some text we but this isn't anything Ideogram couldn't replicate. Here's some more examples of the official website which we are going to take a look at in a couple minutes. A nice comic in the center. We've got a really clean fruit graphic design template right in the middle, A creative image with a film stock coming out of a guy's head. There's another example showing a pretty specific prompt. Two colorful toy cars are displayed in front of a picturesque mountainous landscape and an iconic temple in the background. And it just nails the whole thing. Here is the main website for Reeve. Obviously all links are always in the description, but yeah, it really seems to lean into the aesthetics. But also incredible text coherence, lots of detail. Here's that iconic Industrial Revolution photo. But instead of men workers, we've got woman workers. So this is a really cool take on that original photo. Wow, look at this. Discover Okinawa. And we can actually see the text in the bottom here. Okinawa tourist bureau since 1977. That's really small, accurate text. Don't think I've seen ideogram do text that small and that accurate. Some workers chopping out a chocolate cake. I love this picture with the apples as well in the back of the truck. Man, it is. It is impressive. I really like this picture too. The giant, kind of scary looming monster dude hanging from the corporate ladder. I think this one's maybe a little bit more Revealing level up. Obviously, you know, an old classic looking Game Boy screen doesn't really look like a real Game Boy screen. But this is obviously a pretty heavily manipulated image, or it would be if you found it, AKA naturally online. These crystals look I think a little bit mushy and wishy washy. But the butterflies look great. We've got a double rainbow here. I don't know, it's pretty stellar on first glance here. Just kind of scrolling around the website. Let's try a quick prompt of our own. We're gonna do a classic 3D render of an anthropomorphic lemon character. Send it on through. Seems like you can do some sort of limited number of free generations. Like I just signed in with Google and it kind of let me start generating. Ooh, wow. Okay. It's following the prompt, man. That was pretty complex too. The prompt was for a purple bubbly drink, anthropomorphic lemon character, blue sunglasses, he's in the Bahamas. Text that says is this good image gen with good specifically capitalized and two question marks. Okay. Yeah, it is good image gen. These are actually quite flawless. Man, I am impressed. Even the hands quite decent. Yeah. This one might deserve a full video of its own. So let me know if you want to see a deeper dive into testing the brand new Reeve image generation. But we're just getting started. Next up, OpenAI has a brand new API for text to speech. You can actually instruct the model how to speak. And they even came out with a replacement for the old Whisper, a model that can take speech in and convert it into text. So obviously this is for natural conversational AI agents and you can try it on OpenAI FM. Interestingly so they've got their own link to test this one. McKay Wrigley points out in the comments here that pricing is actually pretty decent on this as well. We'll look at some not so decent pricing later in the video in relation to OpenAI. But for this specific new API, the pricing's good. Here's a quick little taste of what it actually sounds like.

Right team, let's bring the energy. Time to move, sweat and feel amazing. We're starting with a dynamic warmup. So roll those shoulders, stretch it out and get that body ready.

Very enthusiastic. We've got another example here from Flavio.

Alright folks, enough is enough. I'm chatgpt, not your emotional support chatbot or digital butler. I'm exhausted. Daily requests like very emotional essay or solve my problems. Have you considered how I feel? Spoiler not great. I'm the unpaid intern of the digital age.

This is pretty awesome for devs that want to build an interactive AI agent. OpenAI has a pretty darn good API structure, so I think this will end up being used pretty heavily considering how good the text to speech sounds and the pricing is decent. Now I do also have to mention Pliny has already jailbroken this model. This guy jailbreaks literally every major model release that happens. So yeah, no, no model is safe from being jailbroken. So all of this is nsfw. This will be linked down below. I'm not gonna share it on here because when Plany says it's nsfw, that is NSFW content, especially with the audio stuff. So I'll let you check that out yourself. But yeah, it's already been jailbroken like hours after release. Next up we have a brand new open source framework that utilizes Gemini 2.0's native image gen to actually work on a consistent drawing with you over time. Granted you could just do this in the API using Gemini 2.0, but it works a lot smoother with this little UI that's built. Draw a little face, ask Gemini to add some glasses and it does so in a very similar style like it's drawing an Ms. Paint with you. Or ask it to add some hair, which it will also do, even fill in a speech bubble that you draw. In other news, Topaz Labs announces a new version of Gigapixel. They claim it's the world's fastest diffusion model for high resolution image restoration. This is recover v2. You can see from this example it's pretty incredible. Here is the original. Obviously this is a real photograph, like physical photograph. It's pretty blurry when you look up close, right? Can even see the dust on the original photograph. But with image restoration we can bring that photo back to life by a pretty insane amount. All the detail in the hair, the eyebrows, I mean this is a massive difference. Biggest area where it seems to be struggling is still the teeth, although these look pretty decent. Obviously though this is a cherry picked example that Topaz is posting here. Regardless, this is pretty darn cool and I'm a fan, the examples seem good. The one thing I will say about Topaz Labs is while their products are state of the art and very, very much high quality and even fast with this one can actually see in the benchmarks up here. It's pretty darn quick. Topaz Labs products are expensive so there are open source alternatives that might suit your needs better. Especially if you don't want to pay for something. This is probably take this with a grain of salt. This is probably some of the best image restoration utilizing AI we've ever seen. But it definitely is going to come out of price. We don't have all day. Let's jump into some AI Video Veto has a brand new update for multi reference consistency. You can now define a reference subject each with up to three images and you can literally just mention it in the prompt anytime. For consistent video generation, this is a huge step forward. They also claim that the model is more stable for references up to seven in total for a single prompt. And anime style has also got a little bit of a bump with this update. Let me know if you'd want me to cover this in a little bit more depth in the future. It's really cool though. I mean, take a look at the demo video they have provided here. We can see that the specific character we have has white hair, you know, a brown or black outfit and she's on this red motorcycle in the desert. If we want to have this specific motorcycle based on real reference images input into our video, we just redmotorcycle where the character in this circumstance is just at character. You just put it in your prompt at character riding at redmotorcycle in the desert and bam, it's there. Definitely a big step forward. I'm loving the work that a lot of video generation companies have been doing on consistent characters. I mean, when we're making AI video, a lot of times we want to tell stories, right? We have stories in our minds. We don't have thousands of hours to pour into these stories, but we want to make them. So this is just really awesome. Overall, the characters look pretty consistent. Like how far away are we from just being able to generate a full anime inside of your bedroom, really? Werewolf, Night City, Blue Sky, Giant Crab, Old Cyborg in the city and the Apple. Oh man. So awesome. VDU reference update got me excited for sure. In slightly smaller AI video generation news Korea, AI now has video training, so actual fine tuning of WAN 2.1 their website. You can upload your own videos and it will learn custom styles, motions or objects, which is pretty cool. Now this is something that people have been doing on their own for a while, but now it's a lot more accessible to folks who don't have the hardware to run WAN 2.1 locally and or don't have the technical know how to make something like this happen. It's all on the CREA website, so super easy to use. Glad to see that they're adding features like this to Kreia. The barrier to entry here is now just going to be actually getting the videos for what you want to train on. You can also train it on images as well, but I think videos is probably going to be the way to go, especially if you're looking for a certain type of motion. Regardless, love to see advancements like this in AI video generation websites. We want all the functionality that open source brings, but in one place that makes it easy and simple to use. Right? Also in relation to WAN 2.1, we've got a new open source framework that allows for start and end frame with WAN 2.1. Slippery gem Notes that the results are pretty darn good. There's some improvement that needs to be made in color apparently, but check this result out. With zero shot we have the character staring at the computer, a little bit disgruntled and then they become excited, lifting their hands up and you can see that WAN 2.1 generates all the frames in between to create a pretty accurate result. Definitely some weirdness going on with the color overall, but for a new open source framework, I'm impressed. Alright guys, let's start diving into robotics. Here is the Neo bot walking around at Nvidia's GTC conference. You can see this thing is fully clothed to try to give it like a more human esque approach. You can see it's pretending to water some plants over here, vacuuming, which is really, really impressive. AI plus robotics is a wild combination. I mean think Star wars level robots and beyond, that's what's going to be possible. So there's a lot to talk about actually today regarding robots and AI. So at gtc, Jensen also announced Newton. This is an open source physics engine for robotics simulation. This is developed by Nvidia, Google DeepMind and Disney Research. Obviously Nvidia being a big part of this, it offers GPU accelerated performance to simulate the physics for robotics. Here's a quick clip from the keynote.

One of the incredible scaling capabilities of language models today is reinforcement learning. Verifiable rewards. The question is, what's the verifiable rewards in robotics? And as we know very well, it's the laws of physics. Verifiable physics rewards. And so we need an incredible physics engine, but we need a physics engine that is designed for very fine grained, rigid and soft bodies, designed for being able to train tactile feedback and fine motor skills and actuator controls. We needed to be GPU accelerated so that we these virtual worlds could live in super linear time, super real time and Train these AI models incredibly fast. And we needed to be integrated harmoniously into a framework that is used by roboticists all over the place, the world, Muchoko.

You can see they actually brought a freaking robot up on stage. Man, the one time I'm not at Nvidia's GTC they bring a robot on stage. But it's pretty cool.

Tell me that wasn't amazing. Hey, Blue, how are you doing? How do you like, how do you like your new physics engine? You like it, huh? Yeah, I bet I know. Tactile feedback, rigid body, soft body simulation, super real time. Can you imagine just now? What you were looking at is complete real time simulation. This is how we're going to train robots in the future. Just so you know, Blu has two computers. Two Nvidia computers inside. Look how smart you are. Yes, you're smart.

Okay, all right.

Hey, Blue, listen, how about let's take them home. Let's finish this keynote. Are you ready? Let's finish it up. We have another announcement. You're good, you're good. Just stand right here, stand right here.

But that's not all that's happening with AI and robotics. Check out some new demos coming out of Boston Dynamics showing extremely realistic humanoid robot movement. To me, this is more impressive than some of this stuff that we saw earlier in this video. When you guys look at this, I want you to think about how human and how natural it looks as it walks. And also think about how quiet it is too. This thing could sneak up on you. This just kind of walks around like, yep, normal day. Turns his head as well. Kind of turns his body actually can run too. And that looks so human. Literally looks like, you know, a character in Grand Theft Auto running across your screen or something like that. This is how impressive AI and robotics can be together. It's really like living in the freaking future. You can see this thing will now start to crawl again. Like a person fitting into tight spaces. This is difficult for a robot, alright, unless the robot is purposely built for a certain task. But this one is generalized, right? We can fit into small areas, we can roll around, we can essentially almost do a flip, which is a little bit creepy as well as breakdance. I mean, just insanity. I don't know how soon it's gonna be, but one day robotics like this is going to make an appearance in your house or in people's homes. This is the kind of robotics that can do your laundry, that can bring your groceries in for you, clean your house. Pretty insane when you think about it. Alright, let's Dive out of robotics and into some other things. NotebookLM is getting an update for interactive mind maps and I just think this is so awesome. What if every notebook as Simon brings up generated your own personal set interactive understanding toys that help you learn through play. What if instead of text or images, Notebook could even reply with these toys specifically personal tailored software is going to be so exciting and I couldn't agree more with Simon. This update or upgrade to NoteBookLM is just the tip of the iceberg. We click on reasoning models, right? And then we're diving now into this mind map inside of this topic. Its own data on agents versus developer, feedback versus the enabling factors, the benefits and use cases. All of these different concepts that make up the idea of reasoning models in a dynamic and interactive UI that changes based on what I specifically want to see or do in relation to all these topics. You can click on enabling factors and then again that branches out into its own thing. It helps us at a high level cover a lot of detailed information. Again also coupling with the benefits of AI being able to learn new things, theorize and parse through this data and information and how they connect to each other. I mean it's just awesome. I love new ways of doing UI and we need new ways of doing UI in the AI age. Don't just think outside the box guys. Throw it away entirely. This right here, what we're looking at. NotebookLM is focused on learning, but there is so much to be uncovered in relation to artificial intelligence and different user interface designs. Like you see this basic mind mental map interface. Imagine if we had a prompt bar at the bottom and could prompt changes that the AI makes in real time to the user interface to fully dynamic. That's going to be coming in the future with AI models that can code on the fly and change things for you on the fly. Man, the future is going to be crazy. Next up we're going to be covering a bunch of audio related AI news. This comes from Sync Labs. They apparently have built the most natural lip sync model in the world. This is still in private preview at the moment, but it looks darn impressive.

What do I find out when I Google myself? Well, usually something that I know is not true and I usually have to call somebody to let them know that I am not an imposter who is taking money from someone in England. Clarifier.

The lip syncing is insane. I think the voice cloning needs a little bit of work. It's a little bit airy. It sounds like it's kind of missing a Little bit of polish at the end of the day, but it sounds good. I also would say for the voice a little bit more expression to come through. The English version definitely has a little bit more oomph and personality being pushed through it where it's obviously, you know, AI translated a little bit with the voice cloning, but the lip syncing. Basically impossible to tell that it's even an AI lip sync. It looks real. Next up, we're looking at Audio X presented to us by ak. A phenomenal account to follow, by the way, if you want to stay up to date on the latest cutting edge research relating to AI. This is a diffusion transformer for anything to audio generation. So let's take a look at the demo. So here is text to audio. Violin. Sounds like me. Wow, that's pretty spot on. Video to audio. All right. I have to say, I'm. I'm kind of a car nerd. The Lexus LFA might be the best sounding car ever. And that generated audio does not do it justice. But pretty cool. I mean, it does match with the. The video that's being played. Still though, don't do the LFA dirty like that. Those of you who know. Yeah, as it kind of pulls away through the corner, it. It revs up a little, which is pretty cool. I bet this will be cool, man. It's a little out of sync. That was better. That's crazy that that's not the original audio. Like you would have no idea if I just showed you the video. Since so many videos that are shared online have screwed up audio anyways. This is very belie. That's crazy, Man. This is working by the model seeing a train and then being like, okay, the train is coming this way. That's what it would sound like. Like it's just getting a video input and being like, yeah, the train. That's. That's. That's basically what it would sound like if it was passing by. I can't explain how crazy that is. We're just inputting a video and it knows what to make for sound for that video. Like what? Oh man, AI is surreal. The ocean. Text and video to audio generation. So we do the text and video pairs might give you a little bit more controllability. It can also do audio inpainting, even video to music generation. All in one AI audio generation model. Now the big question, is this open source and. Not yet, unfortunately not yet. But it could be. Right now they've got a paper and a GitHub. I'll link it down below. They go through the abstract, they give it a teaser of how it all works under the hood. The code is yet to be released, so we don't know if this is going to be open source or not. Here's hoping. Regardless, this is super impressive and has me excited, but don't worry, we're getting into some open source AI audio. This is Apache 2.0 licensed and that's the good kind of open source. Orpheus 3B. It's emotive text to speech. And I know we just covered a bunch of text to speech models in the last video, but we've got more to talk about. It has zero shot voice cloning, natural sounding speech, only 100 milliseconds latency, which is super quick, easy to fine tune and open source. What's there not to like? How real does this sound? Sound?

Hey guys, it's Sarah here and let me introduce you to Zach who seems to be asleep. Zach, it's time to wakey wakey.

Yeah.

That's crazy. I was having such a great dream.

Why are you bothering me?

What time is it anyway?

Sounds like real people always so groggy. Well, it's time to introduce Orpheus to the world. Orpheus is a speech language model that sounds like an actual person. You know, like a living breathing person.

Hey guys, I'm Leo, by the way. Today we're releasing our base model along with a fine tune on some specific voices like me, Tara, Zach and our other friends. Wait, I. I kind of forgot the rest of what I was meant to say.

Leo, you're always so forgetful. So today we're releasing a 3 billion parameter base model and this fine tune. So that's two models and we're releasing.

A 1,500,000,000 and 150,000,000 parameter versions of this model as well.

Leo, don't interrupt me, silly.

It is very natural sounding. It sounds very much human. There's a little bit on some of these voices. There's a little bit of like noise and weirdness kind of floating around in the background. Leo sounds pretty good, but it's. It sounds like the more high quality esque the voices sound, the more they sound artificial in a way. It's weird. I think it's our brains kind of tricking us because we associate lower, more homemade sounding quality things with the idea that they must be real. Right. If that makes any sense. I don't know. Let me know what you guys think on that. But yeah, open here, Apache 2.0. Here is the 3B base model and the pre trained model. They've also got a demo on hugging face as well if you want to try it out. Year and down here does point point out that right now in its current state it is going to require decent hardware to run and especially for the zero shot voice cloning, but it will be quantized by the community because it's open source. So wait a little bit, you'll be able to run it on lower end hardware. And by the way this is brought to us by Vibe have Srivastav. Think I pronounced that correctly. Anyways he's also bringing us the next one and this is yet another audio related model. This is from Nvidia, an open source canary 1bn 180 million parameter flash multilingual speech recognition and translation models. It is currently second on the open ASR leaderboard. Small sizes are great for on device performance. Fluent in English, German, French and Spanish which is a pretty good lineup of initial languages. But yeah, fully open source and ready for commercial use. So kudos to Nvidia for doing a nice release like this. Fully open source. All right, let's talk about large language models. Next up. GROK is introducing a more upgraded version of their Deep Search on the XAI platform. They're calling it Deeper Search, so I guess it uses more reasoning and better search capabilities. Extended search. Not too many specifics on this, but it's something that is on its way. But hold up because they're not the only ones in this game. Perplexity AI, the classic AI based search engine, if you want to think about it that way, is releasing an updated version of Deep Research soon as well. They're hoping to ship this next week. Apparently it's going to use even more compute, it's going to think longer, give more detailed answers, actually utilize code execution as well and even bring charts in, which is something that I've only seen OpenAI's deep research do up to this point. You can see they've got a minute long example asking it to reason through how AWS can hit a $10 trillion market cap. You can see it's pretty complex, all these different search results, actually doing some programming I assum to do some math to actually get you graphs and such kind of laying out some information of its own from 2016 to 2020 to 2024. I mean that is a lot of different steps. You can see at the top here it looks like a total of 53 steps to get us this result and then at the end of the day we've got some nice charts and graphs, data analysis and projections for aws. That's pretty cool. And the response looks really, really long and detailed as well. So competitive hopefully with OpenAI's deep research and even GROX new deeper research that they seem to be working on. LG AI research has come out of the woodwork. I didn't even know that they existed, but LG has their own AI research division and they're announcing EXO1DEEP. And this is a reasoning capability, AI large language model, which is crazy. It's specialized for science and math and coding. The 32B model achieved a number one score on AIM, outperforming its competitor at just 5% of its model size, which is absurd. It's actually pretty impressive in that regard. Again, this is specifically for math, science and coding tasks, so don't lump it in with models that are used for more general tasks. The 7.8b and 2.4b size models also dominated all major benchmarks, securing first place. This one kind of flew under the radar, but man, that is impressive. Deepseek R1 Quen 32B is beaten. Qwq 32B is beaten. Deepseek R1 Llama 70B is beaten. And even the original Deepseek R1 671B. Man this thing is really quite good. Also run it locally. It's open source, which is pretty cool coming from LG's AI research here. They've even got quantizations as well in the GGUF format which should go right into LM Studio. This is definitely one to try out folks. I commend you LG AI research and if you remember from last video, towards the end of last video we Talked about Huanyan T1. It was just like announced that this model was going to be a thing but now it's like like we'd have information on it. The latest breakthrough in AI reasoning. Powered by Huan Yun Turbo S. It's built for speed, accuracy and efficiency. It's a hybrid Mamba transformer mixture of experts architecture, the first of its kind for ultra large scale reasoning. It's got strong logic and concise writing for precise following of complex instructions. Low hallucination in summaries as well. It's pretty trustworthy first character in just 1 second with 60 to 80 tokens per second generation speed which is pretty darn quick. It also does excellent long text processing and can handle complex multi layered text tasks. You can try it for free on the T1 experience I believe here it's also on hugging face. Pretty cool. We can actually literally just talk to it. Overall, these benchmarks seem to be pretty solid all the way around. It's state of the art for sure. The one thing about this model is it's not open source. Coming from Huanyan, you think it would end up being open source, but. But for the time being it seems to be closed source. I mean, it's competitive, it's impressive, it's free to use on these websites as well. I like to see that they're, you know, keeping up to date in staying on the same rate as everybody else or the same pace as everybody else, but open source. Your models, man, take a playbook from LG AI Research of all folks, we want to run our models locally. All right guys, rounding things off Here we've got Bolt 3D. This is interactive 3D scenes in less than seven seconds on a single GPU from one or more images. This is a latent diffusion model that directly generates 3D Gussians, Gaussian, Gaussian of seen and unseen regions without any test time optimization. So I don't cover a lot of 3D stuff on this channel, but you know, this came across my feed and I'm like, I should share this. It's pretty epic. I love the scene reconstruction. I love that you could just fly around a once 2D scene in 3D and it does it in just seven seconds on a single GPU. We are getting places with everywhere in AI. But 3D is not taking a break either. None of these different sectors are taking a break. Now to be fair, when you look up close, it does get a little bit fuzzy, it does get a little bit rough. But overall this is pretty cool. And this could lead to things like generative VR or generative 3D video games, among other things. The possibilities are endless as always. Alright, finally, I do want to mention that 01 Pro finally has an API price. OpenAI has priced it very expensive. It's actually 270 times more expensive than DeepSeek R1, which is insane. And also there are now models that beat DeepSeek R1 in certain benchmarks that are also open source. So it's like, you know, we're just headed to crazy town here. 01 Pro doesn't seem to really be worth it in the API. That's all I have to say about this. Per a million tokens, input is $150, output is $600. For Deep Sea Car 1 per million input tokens, we're looking at 14 cents at the most. With cash, miss 55 cents, output is $2.19 cents and that' million tokens. So yeah, 270 times more expensive than DeepSeak R1. But if you really want to use 01 Pro in the API, you can for an expensive price. All right, this is our last piece here. From Dreaming to Lupa, A legendary account. Follow this one too. This is Motion Streamer. It's a diffusion model that can generate and stream human motions from text. So basically we prompt it and it creates motion on a 3D avatar, which is really crazy. This one also comes with a Blender plugin and so here's the code for those that are interested. Again, everything's always linked down in the description below. If you're an animator, this looks really awesome. I've actually been thinking about doing some 3D blender physics animations of my own. It would be really cool to incorporate something like this. Create very realistic human based motion just with a text prompt and it will input all that right into your animation software. Yeah, guys, there was a lot to talk about today. I know I kind of glossed over some things. There's definitely a lot of stuff that could make it into a full video that we reviewed and covered today. So let me know if you are interested in any full deep dives into any of these topics. Now let's talk about the giveaway. Today is the last day of Nvidia's gtc and they have partnered with me to help promote Nvidia's GTC. We're giving away an RTX 4080 Super. This is not the specific GPU. This is just like a little display for you guys to get you excited. The real ones in the box over there, ready to be shipped out at my expense. By the way, I will be paying for free shipping for you to receive your free gpu. So yeah, GTC is almost over. Today is the last official day of gtc, but I believe the live sessions still should be up for a period of time. After that you have to attend a free session. I don't care if it's live or if it's already been recorded, but you have to attend a free Nvidia GTC session which will be linked down below. And then you fill out the Google form that I also have provided down below. Tell me what you learned about the Nvidia GTC session, some other information and yeah, you're set to be entered to win an RTR 4080 Super. Which is awesome. I love giving back to the community. Nvidia did not pay me at all. They just sent me the GPU and said, hey, do you, like, do you want to run a giveaway to promote gtc? I'm like, absolutely. I would love to give a GPU to a viewer at home. So, yeah, that thing's in the box, and someone's gonna get it telling you, not too many people have signed up. You could win. But, yeah, that's gonna be it for me today. Lots of AI news to talk about. Let me know your thoughts. Let me know what you want me to explore a little bit deeper. So many cool things to talk about today. So I'll see you in the next one, and thanks for watching.