AI News Drops to Blow your Mind! Google 2.5 Pro, Hunyuan Custom, & More!

Welcome back everybody to the Matt VidPro AI YouTube channel. I'm back with another AI news roundup and this is definitely a good one. Plenty to talk about. In fact, there's so much to talk about I don't want to waste any more time on the intro, so let's dive in. LTX Studios new AI video generation model LTX V13B. This is a fully open source model that actually generates very quickly and cheaply. So yeah, 13 billion parameters. It's got pretty smooth motion, not too much artifacting going on, but it still is AI video of and it's sharper than the previous models. Obviously this thing's not going to compete with Google's VO2, but the point is that this thing is fast and runs on low end GPUs while still being absolutely usable for a lot of different applications. Now the way they were actually able to achieve all of this, this speed while maintaining decent fidelity is through multi scale rendering. As they describe, it analyzes scenes at multiple spatial resolutions at once so the model can understand scale structure while preserving the finer details. So smooth motion, more consistent frame to frame coherence, and plenty of detail. And you can see this is an example of the different spatial resolutions at once. And yeah, I have tried this model out for myself. You might have seen the sponsored short I did for them. It does generate super, super fast, especially in LTX Studio itself. But for usable AI video at scale, this is a fantastic model and I gotta give LTX huge props here for releasing it fully open source like this. They've got a full GitHub page to get you started using this sucker in Comfyui. If you install the quantized version of this 13B model, you can get this thing running on only 8 gigabytes of VRAM. And there are a plethora of cards that at least have that. Next up, hey, gen just dropped avatar 4. All you need to do is give the AI a single photo of your avatar and then a script and it can generate very realistic looking video. Yeah, I mean just looking at the video itself alone, very difficult to tell that that's actually AI generated. Once you hear the voice, it kind of gives it way. But it's pretty astonishing how far AI avatars have come. Take a listen. I'm just a small town girl living in a lonely world. Sorry I'm a terrible singer, but did you know you can turn a single photo and script into a high quality avatar video using our new Avatar 4 model? That's kind of crazy. Hey, Jen's been doing this for a while, but the avatars seem to get better and better and better. And at this point I can barely tell, at least visually, that this isn't just a video of a regular dude. Not going to spend too much time on this one. If you want to learn more, Angry Tom's thread going to be linked down in the description below, along with everything else we're going to be talking about today. Next up, we've got a bunch of Google News that I'm excited to get into. Google just dropped a brand new Preview of Gemini 2.5 Pro and this is the 5.6 checkpoint. So that was only three days ago. Anyways, Google shared some fun simulations that people have been making with this new model and I thought it'd be a pretty great way to demonstrate some of the crazy capabilities with these high end LLMs like 2.5 Pro Preview. Before we dive into the rest of today's news, I have a quick message from today's sponsor. You're already ahead of the curve if you're using AI technology to make your work life more efficient. So why not get even farther ahead by using Zapier in combination with AI technology? All those manual tasks for me, it's posting links, updating my calendar, sorting through vast quantities of emails. They no longer need to be manual. I don't know about you, but I've had it juggling link tabs. My everyday workflow can be turbocharged by using AI powered zaps. With Zapier you can connect 8,000 plus apps, Google Sheets, Notion, Slack and build no code workflows once. Then your AI powered automations will just run themselves. Zapier now has AI steps built in directly. I can auto draft video descriptions, summarize comments, and even flag certain trending topics, all without lifting a finger. Here's a quite useful one built on my workflow. It's pretty simple. Get the transcript script from a YouTube video and use AI to create descriptions, timestamps and a thread draft for Twitter. No code whatsoever in leveraging the power of AI working smarter, not harder. Thing is that Zapier has been around for a while. It is completely free to start, no credit card required, and only upgrade when you actually need the horsepower. I've been on the free plan making AI powered zaps. No charge. If you're ready to take automated workflows with AI to the next level, sign up for Zapier completely free with the link down in the description below. And hey, if you make any cool AI powered zaps, I'd love to know about them, so feel free to share them either in the comments or on my Discord server. Thank you so much for Zapier for sponsoring today's video. Now back to your regularly scheduled content. Thanks so much for putting up with sponsors, guys. They make this channel a full time reality for me. Alright, so first up, we've got the Shape Visualizer, a quick program again made entirely with Gemini 2.5 Pro Preview. Fully 3D. All these different shapes that are even slightly textured in terms of having the ability to like reflect, reflect light and such. I'm no coder myself, so I'm just shocked that AI can even do this in the first place, you know. But we're not just talking circles and squares here. There's all kinds of different shapes that are very unique in here. In Shape Visualizer is a pretty perfect name. And as you can see, the UI that this is actually in is literally just the Google ui. So this is just chatting with the model straight up, generate some code, you click preview and boom, there. This is entirely. It's not like we had to export the code and then put it into a project ourselves. It just runs in the browser like this. Absolutely mind blowing how far AI coding is getting. Still has a long way to go before it's making, you know, AAA games or something like that. But I'll tell you, it is far beyond what I could ever hope to do. Next up, Pietro here made one gorilla versus 100 men. I'd like 2D simulation again made entirely with Gemini 2.5 Pro. Just uses emojis for the graphics here, but you can kind of see the damage here. The gorillas going after all the men. I don't know. I think the men might have it with this simulation though. We'll see. This is. This might be very accurate here. If you guys want to know the truth behind 100 Men vs Gorilla. I don't know. I think the gorilla is kind of going off right now. It's even calculating how much health the gorilla has and then how many men are fleeing at any one time from the gorilla. Men have numbers, but low morale. Unarmed. Dang. There it goes, man. I don't know. We'll see. Like, yeah, some of the men are. These guys are being a little bit smart, staying away from the Goril. This is just a funny simulation. This is so sad that we don't actually get to see who wins. By the way, if you want the official Matt Vid pro stance on 100 men versus gorilla, I think 100 men are definitely taking it all day. But I've also got a friend who thinks 20 men can take out a gorilla. And that's just a lulu to me. Especially if our AI generated simulation is anything to go by. Check the throat out for some more gorilla versus men style simulations. Wesroth also made a 3D traffic simulator with Gemini 2.5 Pro Preview. A good way to put like this into perspective is to think about how difficult something like this would have been to create not that long ago. Like a 3D city with actual simulated traffic would have been impressive on the Nintendo 64 for example, or the GameCube. Just going by video game consoles now you can just spin this up with your browser using AI. You don't even have to make it yourself. It can just be built in minutes. It's pretty astonishing how far technology can go. Here we've got another one. This is cubes inside of cubes. They're kind of like building each other or building themselves out of lines. It's very interesting. It reminds me of those like old school wallpapers that you would see on computers. But yeah, I mean, tell me guys, would you be able to code this and would you also be able to do it in just a few minutes? Because that's how fast Gemini 2.5 Pro does it, man. It's crazy. Yeah, they're all individual cubes. Building larger cubes, a little bit trippy, but it all is actually just one big cube at the end of the day. I don't know, man, this is just so far beyond me. Like if I had to make a project like this, I wouldn't even know where to start. This thing could just plop it out right inside of the Gemini interface. Just stellar, mind blowing. Also a fun musical physics simulator as well, where each ball is like a different instrument and then they sort of go up and down in the tones. It looks like you just click to like spawn a ball in. It's pretty cool. I remember seeing an app on the app Store that was kind of similar to this. But yeah, now you can just vibe code it with Gemini 2.5 Pro, a 3D city one. This one actually shows the whole process of being inside of the Gemini app, actually building this and then previewing it. I mean, that is insane how fast it's able to generate like in a whole city like this with like individual moving parts and stuff as well. Look at all the little trees and the cars. Yeah, there's more to see for sure. I won't spoil the whole thread though, so link in the description if you want to check out the rest of them now with this Gemini 2.5 Pro preview model. This 5.6 release, they actually dethroned Sonic 3.7 on the web Dev arena leaderboard and apparently not even OpenAI's O3 model was able to do that. But yeah, at least on this benchmark you can clearly see that this is quite a leap even compared to the old 2.5 Pro experimental from March. Jumping up quite a few ELO points here, Google is also planning to roll out some form of their own computer use agent inside of AI Studio. We don't have too much information on this, just a few snippets from the website. It will be cloud run though. It spawns virtual desktops similar to OpenAI's computer use. Gemini will be able to control them via a computer use tool, so expect something very similar to OpenAI's version of computer use. I gotta agree here, I'm not as excited for this. I'd rather use something that is actually can take control of your whole Google Chrome or something like that. That would be a lot more useful. But nonetheless I'm excited to see what this has to offer now. Gemini 2.0 flash from Google also got an upgraded version of its own native image generation similar to OpenAI's huge drop not that long ago with their native image gen. This definitely is a bump in quality. Even users on my Discord are messing around with it and noting that still not as good as GPT4 omni, that native image generation is second to none. But this is definitely a step up and more usable for a variety of use cases like text or putting logos on products and just generating images that actually look, you know, realistic at the end of the day. Here's a fun generation from user Thakano on my Discord server. Yeah, I mean it kind of looks like me for sure. I think it gets some of the details in the face a little screwed up. It looks like someone tried to clone me, but half of my brain is missing in the new clone. But hey, I've got a pretty dope Pikachu so I like that. At least I got the hair and the eyebrows pretty spot on as well. Not too bad, not too bad. But yeah, you can see that folks in my Discord server are always messing with the latest model updates. You can see on this one we've got a pretty standard character, then we add some new glasses, a fake nose and mustache and then like this beanie thing onto him. But it preserves like the overall details of the character quite well, swapping between them. I'm pretty impressed with this brand new Gemini update. But yeah guys, if you want to play around with the latest models, really stay on the pulse of AI, I recommend joining my Discord server. Users are always messing around with the latest and greatest. Alright, now let's get into some speech recognition. Something we don't really talk too much about on the channel, but this is a drop from Nvidia. It's parakeet TDT06B. So apparently this is the best speech recognition model. At least if you want to go by the open ASR leaderboard lightning fast. It's crazy like these transcription type of model. Speech recognition used to be something that was so slow, clunky, kind of not super reliable. Especially when I was a lot younger and it was like brand new technology. But now, yeah, transcribe 60 minutes of audio perfectly in one second with this open source and video model. Sure, 600 million parameters and yeah, open licensed. This is a great release. And it's fully open source beating all of the major closed source models on the open ASR leaderboard. So yeah, pretty bananas. And respect to Nvidia for releasing it open source in the first place. It's not something they had to do, but that they did do. Since this is open source and has the ability to run on a consumer gpu. Theoretically you could see a model like this integrated into video games. Imagine talking to AI characters in real time, all locally running on your own system. Definitely possible. I mean 60 minutes of audio in one second, it could absolutely do it in real time. We continuously are seeing rapid development of AI technology that could power all sorts of insane science fiction style ideas. I mean, you name it, and it's definitely at least somewhat in the cards with the AI technology that is being born every single day. Blown away by how far we've come. But let's move on. In other audio news, eleven Labs finally launches sound effects in their studio. It's built directly into their long form editor. All you got to do is describe a sound and it's generated. And usually they're pretty realistic high quality sound effects. I mean the more weird you get obviously the less good it's going to be. But if you want to tell a story right, like narrate a book out with 11 labs and include sound effects to make it more alive when you're listening to it. This is absolutely going to fit the bill for something like that. We'll take a quick peek at the demo they have provided. Sarah lit a match, threw it into the fireplace and poured herself a glass of wine as the fire began to crackle. That's not bad. Now let's listen to a guided meditation example. Welcome to this guided meditation. Close your eyes and focus on your breath. Notice any ambient sounds around you. Rather than engaging with the sound, let it gently wash over you like water. Again, pretty impressive. I like Luke's comment down here. Excited for audiobooks to turn into audio dramas. I mean, correct me if I'm wrong, but the majority of audiobooks don't have sound effects in them. It might depend, right? But most of them, it's just different characters. Get different speakers for this. Not only can you have different speakers for different characters generated with 11 labs, but you can also throw in sound effects to actually really put the listener right inside of the story. Let's move on to Imagegen. Apparently ideogram enhanced their 3.0 model better realism, more versatile styles, and improved prompt following. I did notice when I was testing ideogram 3.0 prior to its full release and prior to this update, of course it kind of lacked a little bit in the prompt following even compared to ideogram 2.0 sometimes. So I'm glad to see that they did ship somewhat of an update to this to give it a little bump ahead. And since this update, I definitely have noticed the difference. I am a regular user of Ideogram. I often use it for thumbnails on the channel. While I still have been able to use 3.0 for my thumbnails, this is a very welcome upgrade. And also finally, if you do use either Magic Fill or extend with 3.0, that's also going to be available in Canvas now. Okay, I really like this comment from Brent. Wait, an upgrade without a decimal or a letter? You can do that? I know they just updated the model without like saying 3.1 or 3.0 R2O plus. All these AI companies constantly have crazy model names, but it looks like Ideogram was like, let's just ship an update to 3.0 and leave it at the same name. Yeah, it's not that deep. I'm with you guys. I feel like Ideogram's constantly kind of been on top of the game with their models. Like if you had to ask me what's your favorite, you know, standard image generation model? Traditional diffusion, not auto regressive stuff like the GPT image gen. I would say Ideogram is like my top tier and has been for a long time now. It's my go to for traditional image gen for sure. Next up let's talk about fantasy talking realistic talking portrait generations. This is a fun open source project for talking characters. No, it's not going to compete with the likes of hey Gen which we took a look at earlier. But it is open source and I think it's worth at least taking a peek at. The demo videos here are actually pretty impressive, but it does take a little bit to get a good generation from this one. I did try to use it myself and we'll talk a little bit more about that. But first let's just take a look at the demo. Ah, welcome, welcome. You're just in time. I made a fresh batch of my world famous how to reset your life when you are feeling lost. Dear friends, good day to everyone. That's a good one. This long journey of oh, I know what to do. I like this guy. I can slide and die. You want a real piece of advice? Instead of hating your job, try loving the fact System Online diagnostics complete. You seem this nourishing and repairing face cream is rich in natural ingredients and antioxidants. It's time to turn up the quality of your characters. Pretty interesting, right? Not as good as the closed source stuff yet. If you want to give it a shot for yourself. They do apparently support this in Comfyui to some degree, but it's yeah, fully open source. We got inference code and model weights. They've also got a hugging face space here. But when I tried to use it for myself, it took like ages to generate just 10 seconds of this talking character. Like we're saying maybe eight minutes to get 10 seconds of a talking character. Not really worth it. And also it didn't really look too great. Definitely got to kind of mess with the settings to get things perfect and it just takes too long to dial this one in for me just yet. I think it needs a little bit more time in the oven. Still though it's open source, it's a pretty cool project. I thought I would share it with you guys at the very least. Now moving on to probably our largest piece of today, Quan Yan Custom A multimodal Driven Architecture for Customized video Generation fellow YouTuber Tim from Theoretically Media just did a full video diving into this one I believe. So if you want to learn more about that, I'll probably link his video down in the description of this video. However, we're gonna take a quick glance at it for sure today. You can already tell that this is all about customization in your videos. Let's add custom objects, let's add custom characters. That's Been a huge theme for AI Video this year, and I love to see it because it's all about storytelling, right? We need to do custom characters. We need custom objects. Things need to be consistent from scene to scene if we're going to tell a story, right? RunwayML has been working on this. Vito has been working on this. Kling has been working on this. And now Hunyan. Lots of demos to take a look at too. We've got a girl playing in the living room. You can see there's a little bit of weirdness going on with her toys. But, yeah, that's pretty much the same girl. Has similar teeth, eyes, nose. Same with this one. This girl is just taking some selfies out in the middle of the street. She's got the shirt that says Hwanyeon. She's got even the chain necklace. The same hair, the same hat. This one's pretty impressive. We've got this particular dog, I believe, like a goldendoodle. Am I right in that breed, Playing with a cat. The golden doodle is kind of merging with the cat in a creepy way. Not a huge fan of that cat. Seems to be, like, losing limbs every once in a while. So a little bit of weirdness, but definitely still the consistent dog. Now they show off combining multiple objects. So we've got this girl and then this cat. So she is painting the cat on an easel here, and it does a pretty great job with that. I will say, though, it does keep the cat essentially the same. Even though the easel is pointing at this lady, the cat is still facing perfectly straight, at least to our perspective. So that's a little bit strange. But still, both characters are in here quite well. We've got this dude sitting at a pool with a lace chip bag. This one's pretty perfect. Not too much to say other than great job with that one. We've got this doctor with a lady in the dress just talking on a building of some form. And then here is a big one. Custom clothing. Of course, we've got our custom character lady. Definitely got the black hair down, the lipstick and her overall face. But we are giving her a specific piece of clothing and again, replicates it quite well. Changing wardrobe is a huge part about telling a story with consistent characters, so I'm glad to see that this model can handle that. Now, these last two examples also show off using reference videos and then just adding in an extra object. Like to see that this is also something that is a possibility. Again, opens the creative doors up a little bit more. Maybe you could take some real videos and then add fake objects in. So we're literally just taking this magician here and we're adding in a Mickey Mouse hat. Does a pretty good job though. Like you wouldn't really be able to tell too much that that's not really always been there as a hat on his head. And same pretty much goes for the most part with this little plushie. Pretty stellar. They have plenty of examples, even comparing to Vidoo, Pika, Hylju and Skyreels, which I haven't heard of. Some pretty deep comparisons in here. I'm not seeing Runway ML in here, which is a little disappointing. It doesn't exactly work the same way. I think the architecture setup's a little bit different for runways, so I suppose we can cut them some slack with that. There are quite a few other competitors that they have listed in each of these examples. So yeah, this one is open source, but it's under a custom Huanion license, so it's it looks like it's going to depend on where you are. For example, it says this agreement does not apply in the European Union. So sorry Europeans. But yeah, May 8th they released the inference code and model weights. Pretty dope even though it is under a custom license. But if you take a look at the peak memory required oh man, you're going to need a lot of VRAM. This is certainly past consumer grade graphics. 60 gigabytes for the lower res and the higher res need needs 80. Oh man, that's pretty crazy. Definitely not great for local setups just yet, but at least it's open source. At least the the community can do something with this. This is better than having it closed source. If you want to do custom local characters, you probably want to stick with image generation and doing the custom images and then converting them to videos with like WAN or something because that is a crazy amount of vram. Nonetheless though, pretty awesome project. Finally, to wrap things up here, we've got a little bit of OpenAI news Sean Rawlson brings to us reinforcement. Fine tuning is now available with the O4 miniseries of models. RFT uses chain of thought reasoning and task specific grading to improve model performance. Especially useful for more complex domains. Pretty cool fine tuning for this new O4 series of models. Not something I thought I would see from OpenAI so soon after they released them. If I was gonna do a fine tune of 04 mini, I definitely would focus it on use of a very specific tool. Maybe try fine tuning it to use like Windows or something. That might be very intriguing guys. Or fine tuning it to play a specific game. We really got to start taking these LLMs in more of an agentic way. Fine tune them on very specific tasks. Who knows, we might learn something. And also this one came as a surprise. OpenAI has added GitHub integration to the deep research tool inside of ChatGPT. And again, I believe Deep Research is its own separate model. So even if you select O3 or 04 mini, there is a separate O3 variant that uses Deep Research. It's its own thing, but apparently now it can actually analyze real code bases, break down product specs and summarize GitHub repos all in natural language, which is pretty cool. Deep Research has continued to be the most impressive research agent I've ever seen from the AI world. Now that it can directly access GitHub for coders and devs, this might be pretty huge. Definitely something I think I'm going to mess around with. At the very least, maybe we can compare certain AI projects, feed it some open source ones and say hey, come up with a way to optimize this or make it even better, right? At least come up with the theory in natural language and then paste the theory into Gemini 2.5 pro. I don't know guys, just some thoughts. Just some thoughts. Alrighty, well that is all I have for you guys today. Thank you very much for watching as see blazing fast pace for the AI world. I thought there might be some large drops this week. I heard some rumors surrounding like Deep Sea Car 2, but it looks like nothing too insane happened this week, just a lot of cool like smaller drops. Tomorrow I've got a live stream at 2pm for the map Vid Pro community Minecraft server. Again, join the discord if you want to be more clued in to that stuff, but it's going to be interactive and pretty fun. So I'll see you guys tomorrow if you want to join that. Have a good one. Goodbye.