Огляд AI-інструментів для відео: масштабування Topaz, 3D-генератор Gaussian GPT та корекція вокалу Suno

We have some big updates on Seed Dance 2.0. There's an AI tool that says it can fix your really bad singing voice. So, I am going to put it to the test. Google released a brand new AI video model. And Topaz and Magnific AI video upscaler that looks to solve the plasticky AI skin problem forever. We'll cover all of that and more in today's episode of AI Film News. Thank you so much for joining. Let's hop in. Now, be sure to like and subscribe if you want more videos like this in the future. Now, I want to kick things off by giving a huge hat tip to the people at the Hollywood Reporter for including the Curious Refuge community that includes you inside a featured story in this month's edition. Rick Carter, who is an Academy Award-winning production designer, gave his quotes about his experience at Curious Refuge inside of the story. And I got to say, it is just wild to hear so many kind words coming from somebody who worked on Avatar, Jurassic Park, and The Force Awakens among so many other incredible films. I really appreciate it. I highly recommend checking out the story if you get a chance. You'll find a link below this video. And also, thank you so much to Runway for inviting Dave to be a keynote speaker at their recent event. Okay, let's get to the news. So, I want to kick things off by talking about two brand new AI video upresers that hit the market. Now, you may already be familiar with Magnific as an AI image upreser, but they actually released their own version of an AI video upreser recently, so we wanted to put it to the test. To use the tool, all you have to do is go to the Magnific website and go ahead and select the video upscaler button. From here, you'll have creative or precision. I like sticking with precision because I want my videos when they get upscaled to be close to the base video. I don't want too many changes or it'll be hard to make all the shots connected. And we'll go ahead and bring in our video here. I have this shot of this lizard just kind of looking at the camera and it looks okay, but there's definitely a lot of that AI artifacting present inside the video. So, we'll drag and drop that into the upload section. Now, from here, you can turn the strength up or down depending on your video clip. And then you can also select 4K as the output if you want. Now, there's all these other settings that you can select. I typically recommend keeping them at default because when you turn up sharpening, it can look a bit synthetic. And when you turn up smart grain, you can have less control over the grain that's present inside your video clip. So when you're ready, go ahead and hit upscale. Now, before we take a look at that final video from Magnific, I also want to talk about Topaz Precision 2.5. So, if you're not already familiar, Topaz Precision 2 was an AI video model that basically allowed you to upres your video with precise details and make it more in line with the uploaded reference footage that you put into the system. It was the best tool on the market, but they released 2.5 just this last week, and I wanted to test it out for myself. So to use the tool, all you have to do is go over to the Topaz Astro website and we'll select the precise model and we will go to Starlight Precise 2.5. And you'll see we have this new apply to all scenes button, which is kind of interesting. Topaz is now allowing you to drop in a video clip that has multiple shots broken down by scenes and you can upscale each one individually. It's really helpful because it allows you to bulk process multiple clips in a single video instead of having to break it up into individual clips. Now, inside a professional workflow, you probably still are going to go clip by clip, but it is cool to have that functionality built inside of the tool. Now, for the output resolution, again, we're going to select 4K. And you do have the ability to change the frame rate if you wanted to, but I'm not going to do that for this specific video clip. So, we'll take that same video clip of the lizard and drag and drop it into Topaz here. And when you're ready, go ahead and click render. So, to pull off this 5-second clip, it cost us about $1.50 inside Magnific $1 inside Topaz. So, let's compare and contrast the results and see which one's best. So, first off, we have this shot, which is version 2.0. This is the old version of Topaz, and it looks pretty good. There's definitely some softness and some artifacting happening here, but there are just some synthetic details on the eyes of this lizard that kind of pulls you out of it. So, next up, we have magnific. And this is the result that we got. I do think it's a little overly sharp for my taste, but it is better than the previous version of Precision 2 from Topaz. And finally, we have 2.5 from Topaz. And I think that this is the best video clip. While there's still some chromatic aberrations happening along the edges and just some artifacting that I think does need some cleanup, I think it has a lot more to do with our base footage that we uploaded rather than the 4K version of the clip that we're looking at now. So I think all things considered, Precise 2.5 is the better video clip. Now let's take a look at a couple more examples here. So this was upresed using precise 2.0. That's the older version of Topaz. Again, it looks pretty good. And I know it's hard on YouTube to see all of the fine details here, but there are just some kind of strange artifacting happening in some of the edges and some of the details here. Now, from Magnific, we have this video clip here, which yeah, looks looks pretty good. There's still some wonkiness happening, but not too bad. And then finally from Precision 2.5, we have this video clip, which again, in my opinion, looks like it's the best quality from all of them. And then our final example here, we have this archival footage of this kid walking. And this is the old version of Precision 2. Again, it's not bad. It's coming from an archive, so it's kind of harder to pull in some of those details. And we have this video clip from Magnific. And I like this because it feels like it's true to the original source footage while also providing more details. So, I think it did a good job at staying true to the original footage. And then finally, we have the shot from Precision 2.5. It looks good, although I do think it kind of softened the skin a little bit. It really took away a lot of the uh graininess that was inside the source footage. But I actually think inside a documentary context, some of that grain helps to reinforce and authenticate the realism of the actual asset. So I think in this case, actually, Magnific did a better job. So all things considered, between these two tools, it's kind of hard to pick and choose, but I do think that Topaz Precision 2.5 is a better tool, and ultimately it's about 50% cheaper. So I think for most of our test, we'll probably end up going with Topaz. But if you're running into some issues, I could see a scenario where you might want to hop over to Magnific for some backup shots depending on the individual clip that you're working with. The team at Adobe also has a pretty sweet deal that is on until April 22nd. For only $25 a month, you can get unlimited AI video generations using Cling 2.5 Turbo, making it what I believe to be the best deal for AI video that I've ever seen. They also have a Firefly Premium version that gives you VO3.1 fast and Nano Banana Pro 4K unlimited until April 22nd. The unlimited generations are only available until April 22nd and then it is half off for the very first year that you are subscribed. But the cool part is it's only monthtomonth. So you can unsubscribe at any time. If you're interested in testing out the Adobe platform for yourself, I highly recommend checking it out. A research team has announced a new project that looks to be nano banana for 3D worlds. So, if you're familiar with Gaussian splats, then you may have already played around with other tools like World Labs or the new World Labs integration inside of tools like Open Art. It basically allows you to either scan an environment or upload an image and it will automatically create that environment in return. Well, essentially this tool called Gaussian GPT allows you to type in a prompt and generate a 3D world and all of the assets inside the actual 3D world. So rather than doing inpainting and changing an image, you actually just communicate with the tool and it creates an entirely new 3D object. Over on the white paper, you can actually see some of this in action where we have only part of the 3D scene, but you can prompt in and ask for the 3D scene to be changed to all sorts of different environments. Now, do I think that this specific white paper and this specific tool set is going to be the future of 3Dbased nano banana back and forth 3D modification experiences? Probably not. But with recent updates inside of tools like WorldLabs or Google Genie that allow you to automatically create a 3D world, having the ability to edit and have fine-tuned control over what those 3D worlds look like is really important. Obviously, not only for animation or gaming development, but also for actual set design. Increasingly, AI tools are allowing you to create a virtual version of your set, put a camera in there, and then you can put your actors inside of that scene and work through that scene just like you were a cinematographer on a real world film set. So, that is absolutely a part of the future of AIdriven filmm. And it's interesting to see tools like this that are laying the groundwork for those further innovations. There were a couple of updates to everyone's favorite AI video tool that is technically not released but kind of released in only certain countries and I am of course talking about Seedance 2.0. So up to this point Seed Dance 2.0 was only available primarily in Asian countries but they are slowly releasing it into the larger world and Europe now has access. Now, of course, if you live in the United States, you do not currently have access, and you will need to use a VPN to access the tool, but you can get access to it over at Dramenia. And we've heard an interesting tip that if you actually VPN into Canada, you can get some of the best deals for those credit generations inside of Seedance. Now, the team at Seed Dance is also very aggressively coming after contracts in Hollywood, and we heard rumors this last week that they are charging up to $2 million for access to their tool for studios, which is uh pretty pretty expensive, pretty wild, and uh obviously not very userfacing, but I imagine in just a few weeks, even if you live in the United States, you'll probably be able to access Seed Dance at the very least inside of tools like Dramania. There's an open- source competitor to Meshi that just released this last week from the team at Microsoft called Trellis 2, and a lot of the test and results that we've seen up to this point have been really impressive. So, let's take a look and see how it stacks up. Now, you'll find a link below this video to download the assets for yourself. Obviously, if you're working inside a PC environment, you can run this locally on your machine, but if you don't have a PC, you can actually run this online via Hugging Face. So, let's take a look at using Hugging Face, which basically allows you to charge per generation inside a virtual machine. So, for our example, I have this image of a 3D mech suit robot, and let's say that we wanted to take this image and create a 3D model. Obviously, it would take a long time with all of the details to model something like this inside a traditional modeling workflow. So, let's see if we can at least fasttrack some of the process using this tool. So, what we're going to do is drag and drop that image into the image upload section. You do have the ability to change the resolution, although when we did maximum resolution, it was throwing some errors, at least inside of hugging face. So, not sure what's up with that. So, we'll keep it in the middle of the road resolution. And you can change your seed, which is pretty cool. And then adjust some of the other settings as it makes sense. And when you're ready, go ahead and hit generate. Now, I'm pretty shocked because it only takes about 30 seconds to generate these models. And after 30 seconds, we have this little guy here. And uh yeah, it looks pretty good. There's a lot of detail there. Obviously, it's not perfect, but it's pretty darn impressive, especially for an open- source tool. So, when you're ready to actually get that 3D asset, just hit that extract GB button and it will give you a 3D file that you can then bring inside of tools like Blender and you can go in and further refine the model, adjust your lighting, your shading, and everything else. Now, I am curious to see how this stacks up against some of the better image to 3D tools that are on the market. And so, I'm over here on Meshi and let's go to image to 3D. And you can see we have the drag and drop section here. So I'll just take the image of our robot and drag and drop him into the image section. Now from here you do have the ability to adjust further settings as you see fit. But we'll just go ahead and click generate. And after texturing this is the result of the 3D robot from Meshy. And it is definitely better. It's better in so many ways. But it is interesting to think about how an open-source tool is already creating 3D assets that are comparable to some of the previous versions of Meshi. It seems like there are AI tools that are constantly evolving, but just one or two iterations behind is an open- source tool that is completely free. So, it's definitely a neverending race here in the AI world. And it's why I think a lot of people who develop AI tools are pretty tired right about now. The team at Google released Google Vo 3.1 Light this last week, which is an even faster version than Google 3.1 fast. The big value proposition here is that the tool is half the price of the previous fast model. So, it really allows you to iterate faster and get some sort of moving video on screen as quickly as possible. For a lot of quick prototyping projects, you don't need the final video quality to be picture perfect. Sometimes you just need a visual representation of what you're looking for. And in this case, you will be able to do that even faster. Now, 3.1 Light is currently only available inside of Google's AI Studio and through their API. It's not available inside of Flow just yet, but I think it's really cool to see how we have different models to pick from depending on the creative applications that we're working on. Obviously, at Curious Refuge, we tend to want to focus on the most cinematic, epic, great quality video models possible, but not every single creative team needs the very best video model. And so, tools like this allow them to get creative results very fast. The team at OpenAI came out with some news that would have been pretty shocking 12 months ago. So Sora, everybody's favorite slop factory, has unfortunately come to a close. The team at OpenAI released a post on their social media saying that they are discontinuing Sora. Now, the reason why I think this is actually a good thing for the creative industry is because a lot of the content that was created on Sora is what most people would categorize as AI slop. It was just really cheap and fast content that was just kind of weird. And a lot of the focus on AI creativity in the video world got focused specifically on Sora outputs. But I think that's the wrong way to think about AI video generation. I think AI is a tool that can enhance creativity. And if you're taking your existing skills, you have the ability to create stories that would have been otherwise impossible. And those stories can be really refined and beautiful and resonant with us at a human level. And I wasn't really seeing a lot of that from Sora. So, the fact that they're suns setting this tool doesn't really change a lot for creative professionals. And my hope is that the quality of AI video generations that you see on a day-to-day basis will only improve. The team at Sunno, the AI music generation company, released version 5.5, which allows you to take your own voice and put it inside the songs that you generate. The value proposition is that it can fix your really bad singing voice and make it sound like you are singing songs that you've never heard before. I'm super excited and moderately nervous to test this out for myself. So, uh, let's hop in. To use 5.5, all you have to do is go to the create tab on the left. And you'll see they have this button here for add your voice to the mix. So, let's go ahead and select try now. So, we have a little section here that says, let's add your voice to the mix. I am uh a little nervous here to uh sing in front of you, but let's get started. Okay. So, all you have to do is click on the record button. And when you're ready, start singing. So, here we go. 1 2 3. Okay, so we finished singing our song. Let's go ahead and select use voice. Sing a song with a gentle flowing melody today. And now it's asking me to describe my singing voice. Is professional the right button? N I think I'm probably just going to select uh still finding my voice there. Okay. So, let's give our voice a name. We'll say Caleb's beautiful singing voice and description is uh very bad singing. Okay, let's go ahead and click save. So, from here, we now have the ability to create our song. So, I'll go to the simple song generator here. And we have Caleb's beautiful singing voice. So, now let's uh create a song. So, I'll say an indie pop song about our brand new Curious Refuge membership. I'm sweating. Let's go ahead and click create. Okay, our first song is ready. Let's take a listen. >> Why do I sound like a vaudeville villain? Okay, let's listen to the second one. Scribbled out everywhere. How do I break through? Okay, that one's not it's not good. Like that is not good, but it is at least better. Okay, so what does this mean? Well, having the ability to create custom songs based on your voice is we'll call it fun if not a little strange. But the music world is always a test of what's coming in the video world. So, we've talked for a long time that in the future you might have the ability to cast yourself and your friends inside a film and watch it together. Well, that technology is probably not too far away. We've already seen tests of tools like this already in the larger creative ecosystem. And I imagine that within 12 months, you will be able to upload images of your friends and will be able to create a film automatically similar to Sora, but probably with much better quality. And if you're wondering, one, no, that voice did not sound like my singing voice, and two, I am 100% signing up for singing lessons. Inside the community, we have so many events popping up. I do want to give a hat tip to the folks at Machine Cinema for hosting such a fun party this last week inside Los Angeles. We have our first to know office hours, which is available to the entire community on April 2nd. That's a virtual event that you can join from anywhere in the world. We have a Curious Refuge community meetup happening in Vancouver on April 4th. The winners of the Feelgood Film Competition will be announced on April 9th. And our brand new membership at Curious Refuge is officially launching on April 7th. We've had a ton of questions about what you actually get with access to the Curious Refuge membership. Well, in short, you'll get access to every single course at Curious Refuge. That includes live technical support and feedback on your homework assignments. You'll have access to one-on-one career coaching where you'll connect with real life members of the Curious Refuge team who can assist you in creating a plan for you to help you achieve your creative goals. You'll have access to exclusive events and speaker series at Curious Refuge along with discounts that you won't find anywhere else. And to top it all off, we're releasing our distinguished network of artists. This gives you a direct pipeline to get on the list that we use to tap artists whenever we are looking to hire people inside of our studio and recommend others to the larger creative world. We've heard so many messages from people who are already enjoying the beta version of the membership and who have said that Curious Refuge has changed their life. We would love to have you inside the program. You can click the link below to learn more and join the wait list. And that brings us to our AI films of the week. The first film that I want to shout out is called Fieldnotes by Aerial Co. It's experimental. It kind of utilizes ChateBT, Midjourney, and Sunno. But we really appreciated the timing on the editing and the pacing. It was really good, and the visuals kind of have this like dark and dreary vibe to them that we just thought was really nice. So, very good job. The next film is called With AI is Love Blind by Adam Palmer. It's hilarious. It's a really, really good project that uh made us laugh out loud. We really felt like the editing was really good. It really did feel like a TV show. And the final film I want to give a shout out to is called Keep Cooking by Simon Meyer. We really enjoyed it when we saw it just a few weeks back. It's very well edited and it touches on a topic that a lot of AI artists probably face. So, I highly recommend checking it out. Okay. Thank you so much for watching this week's episode of AI Film News. Of course, be sure to like and subscribe if you want more tutorials and news updates like this in the future. And of course, if you want to join us for our weekly live office hours, you'll find a link below this video to RSVP. Thank you so much for being an incredible member of the community. We'll see you in the next one.