# This FREE AI Previz Tool is Awesome & a Secret Video Model!

Channel: Theoretically Media
Video: https://www.youtube.com/watch?v=qlSI_ZbRdEc
Duration: 18 min
Language: English
Words: 3542
Transcript page: https://viewrankai.com/tools/youtube-transcript/qlSI_ZbRdEc

---

[0:00] Some pretty cool stuff this week on the creative AI side of things, including a new secret video model, plus free 3D workflows that I you know kind of do that pre-vis thing, very much simplify the process getting you the shot that you want. Uh plus I've got a bit of a helpful tip on that front as well. Kicking off we have a new mystery AI

[0:20] video model, which is always exciting. This one is code named Iron Goddess, uh which probably I mean to me at least indicates that it's likely a Chinese model. Now admittedly we don't have a lot to go off of. Uh Iron Goddess briefly appeared over on arena.ai, which if you're not aware of is well I mean it's the arena. You can run a couple of video generations per day. Uh you never know what you're going to get. Just make sure you're in battle mode and then uh turn it on to video and then uh yeah, type in a prompt and you'll get a couple of results. Uh sometimes those results

[0:51] will be like secret models that are being tested. Anyhow, earlier this week a mystery model began appearing there code named Iron Goddess. Uh it certainly does look pretty good. Synthesia gives us a fluid dynamic test. Let's take a quick look.

[1:09] [music]

[1:14] Now granted it is a fluid dynamic test, so it kind of has that like screensaver-ish look to it, but I do think that this very much showcases the model's ability to like hold consistent structure and well I mean well how it handles fluid dynamics I guess. For something that's a little more character driven, uh Leo ended up getting I don't know kind of a stunning result here. So no sound on this one. Uh Iron Goddess is capable of generating sound, so maybe Leo just didn't uh upload it here, but you know overall I think that this looks really fantastic. Now for sure there are some minor issues here and there, but

[1:44] I'm not necessarily going to nitpick here. I mean it's overall I think this is a great generation. Sliding over to some more like well dynamic tests, uh friend of the channel Brent Lynch gives us our best look at Iron Goddess in terms of how it handles action. Uh Brent got two generations off. Uh this is the prompt for one of them. Uh you know, you

[2:02] can pause it and read it if you want to. In the meantime, uh well, let's go check out the generation. [music]

[2:15] I mean, overall, I think that this looks really interesting. It is not without its problems, so we're going to go over that in just a minute, but you know, it's actually it's nice to see a model like really swinging for the fences in terms of like all of this motion and movement. Uh it is something that kind of reminds me a little bit of SeaDance 2.0 that 2.5 kind of lost. We'll move into some issues with the model, but first uh Brent actually managed to get a second version of that same prompt off,

[2:39] so let's take a quick LOOK AT THAT.

[2:50] AGAIN, A VERY FUN GENERATION. THERE ARE SOME issues with some of the physics in there, but everything is kind of like, you know, being cut and moving so quickly that, you know, I think the idea is that you might not notice. The problem is that speed does buy us is a lot of like decoherence and chop in the frames. Like I mean, if you kind of look it's very like smudgy and on the verge of full decoherence uh when you go frame by frame. And because this is moving so quickly, a lot of times like you're just not going to have time like your eyes are just not going to

[3:18] have time to register like uh you know, any kind of like weird issues like this. You have to know that while we see some chop in more here, obviously, I do think this is a very smart model. Namely, as we have a shot like this where our runner stays, you know, again, fairly coherent. Uh what I'm impressed with is that all of the background there, that

[3:36] all stays uh very uh solidly consistent. So again, this is, you know, there's like a little bit of that attention is all you need uh going on with this thing. And the model does seem to know when to slow down when we're coming into like the ending sequence here where she's getting on the elevator. Like we're not seeing a lot of like chop and de-coherence here. And in case you were wondering, yes, Brent was nice enough to run the same prompt in Stable Diffusion

[3:58] 2.5 for comparison. [music]

[4:07] [music] So, look, absolutely, that is a very solid output. And when you scrub through, you know, the Stable Diffusion 2.5 output, you can clearly see that like our runner character here, like, you know, I mean, she she definitely stays much more coherent um all the way through. That said, I mean, I will say that that does come at the cost of, you know, sort of the movement being a little less dynamic, I guess. Moving over to an output that I got in Iron Princess, uh this is our man in a blue business suit uh walking down a busy city street and running away from a jaywalking ticket. Uh this one is

[4:39] text-to-video.

[4:49] [music]

[4:55] So, look, there are some issues here. Uh it is text-to-video, and it was a fairly minimal text prompt on top of it. But, there's stuff that I really like in here, as well. I think the city actually looks really good in here. Um and when the chase sequence begins, although again, yes, we do have some some issues like the one police officer is actually outrunning the person that he's chasing before, you know, deciding to slam into a a bunch of boxes. Both the movement of the running and the chase camera, I I mean, that that actually is really pretty solid. Now, as to who is behind

[5:26] Iron Princess, well, I mean, we we don't know. Um that said, if I had to speculate, like, if I had to guess, I'd put money down on Alibaba, considering that at their App Sarah conference a few weeks ago, they did showcase a new video model that will be released sometime in November. Now, it's not really clear if this is a video model that is uh a successor to the Wan line or if this is something coming from happy horse. So, I'll be keeping an eye on this. I mean, obviously given the November release date, likely tied to the release of Kuaishou 4, it does make sense that they

[5:59] would be out and testing this right now. So, you know, possibly what we're looking at isn't even like the final version of Iron Princess. You know, on a side note, on the Kling side of things and yes, I will have like a full deep dive into Kling 4 next week. Um, you know, this is a bit inside baseball, but I I I thought it was pretty interesting. So, Kling will actually be moving into a Hong Kong IPO next year. They're valued approximately at 18 billion. But interestingly, investors into Kling include Alibaba, Tencent, and Baidu. Now, it is important to note that Kuaishou, Kling's you know, parent company does hold 68%. So, they

[6:38] are minority investments, but yeah, interestingly, those three along with Kling considering that you know, essentially they kind of like forming a Voltron AI video models to compete against ByteDance. Sliding over to the 3D side. Look, I know this always loses a bunch of you, but honestly, stick with it. There's some good free and not hard to use stuff here. But before we do that and relevant to your interest, let's go check in with our friends at Recraft to

[7:07] check out their new fast image model. You know, it's interesting that we've hit near real-time speeds for video, but a lot of times images, they're the slow ones. So, Recraft just solved that with their V4.1 flash and it is fast, about 1.3 seconds per image. Now, that might be slow for a real camera shutter, but for AI image generation, I mean, that's quick. As a quick showcase here, of course, is our man in the blue business

[7:33] suit walking down a busy city street. I'll generate in Recraft fast and yeah, I mean, there starts popping in. I mean, it is pretty quick. Now, I should say that the reason that it is moving so fast is that these are, you know, low-res images. But, if you do find something that you like, you can, of course, come over to the enhanced feature here, and I actually like it on subtle. And we slammed that off, and indeed, although not as fast as the flash model, obviously. I mean, it is still pretty quick, but we get an upscaled, cleaned-up version of that image. Now, where things get really

[8:05] interesting, exploratory, and fun is when you start adding in styles, which has always been of a Recraft secret power, in my opinion. Um so, again, this is just the exact same prompt with a number of styles applied to them. To illustrate that, playing around with 4.1 flash, and actually one of the older prompts that I used to run here on Recraft all the time, which is our

[8:24] cyberpunk woman with long white hair. You know, we end up with some images that you know, are starting to look pretty interesting. So, from there, I began adding some styles to it. In this case, it was the V4 futurism style, and ended up with these images, which started to tickle my fancy. From there, I ended up iterating a bit more by combining a few other older images to come up with looks like these. We're going to go over that in just 1 second. But, from here, once I once I had these, I was able to combine that stark contrast look with the more colorful, illustrated look to come up

[8:57] with images like these. In particular, I actually really ended up liking this one. Creating styles in Recraft has always been fun, and you know, continues on with version 4.1. You know, you just come over here and create a style. And from here, you can add in references. Now, the fun of doing this, in my opinion, is actually grabbing images that you wouldn't necessarily think would work together. I mean, obviously, you can do some complimentary stuff, but yeah, it's always sort of fun to kind of come up with some interesting, different variations. The new thing here is that they actually provide some prompts in here, as well. So, you can sort of test

[9:28] out your style, as well. So, this gives us a good idea of how our new style is going to look. We just simply hit save, and as you can see, when we return back to our man in the blue business suit walking down a busy city street utilizing quiet ink work, the style that we just created, well, I mean, there is a pretty big difference. An additional cool thing is that you can now share your styles as well. So, I'm I will have the link to this style down below if you want to check it out. So, all of this is

[9:54] available in Recraft Studio right now. And if you've been watching the channel for a while, you'll know that I'm I'm I'm pretty big Recraft fan. In addition to their own models, I do want to let you know that practically every image and video model is available here as well. Plus, they are one of the few platforms to still offer a free plan just to try them out. So, for sure, at at least sign up for that. I mean, Recraft is a great tool to have in your back pocket any day of the week. The link is in the description as is the quiet ink work style. As always, my

[10:22] thanks to Recraft for sponsoring today's video. Moving on. So, I know the whole like Blender to Cascade and like Mini Max workflow seems to be the latest in like OP workflows. And look, I should say, while you can get some really great stuff out of this and even like as we're moving along, taking it even further, apparently people are now messing around with Cinema 4D with a very similar technique. And like yeah, I mean, this looks kind of it looks really remarkable. But, going back to the Blender side of things, you know, it can be very time-consuming, especially in the renders. And might not, especially if you're not overly familiar with

[10:54] Blender, get you with the exact results that you're looking for. For example, this was a this was a experiment that I was playing around with a little while ago taking the original the opening to a Touch of Evil and having it re-rendered out in in Blender. So, for the you know, the exact same camera move. It got close. Like it wasn't quite 100% there. As you can see, even there's some drift in the sync as well. But, definitely did over overall, it got the idea here. It also just took forever to render. Well, to that, we have Cozy Clay, which is again free and

[11:26] open source 3D pre-visualization tool. And legitimately specifically built for the purpose of creating 3D scenes to translate into AI video. Now, there are a few different ways to go about using uh Cozy Clay 3.0. To me, uh in my opinion, the best way is to hop over to Pinocchio. Now, I have covered Pinocchio a number of times on the channel in the past. Uh once again, Pinocchio is free as well. So, you just download Essentially, what this acts as is a very easy way to uh you know, download and get up and running on open-source models, uh anything on the video, audio, or image front. Um actually, they even

[12:04] have like God's eye view over here at the wall's um thing, which is very cool as well if you haven't checked that out. Now, I do have to say that this does take a hot minute to download, so you know, just make sure that you're aware of that. But yeah, I mean, once you uh get it up and running, this is now just running in a web browser, like locally on your machine. Uh and what's kind of cool about this one, uh besides the fact that, you know, essentially, you have like your clay characters over here, uh you can uh move them around if you want

[12:30] to. Um you know, any object in here, obviously, is a is 3D, so you can just kind of play around with it. Uh but, you don't know what you're doing, there's actually an uh agent automation down here in which you can uh hook in your uh OpenAI Codex, for example. Um if you have uh an Anthropic key, I guess you can um you can use your Claude models here as well. But uh if you just have, you know, your regular OpenAI ChatGPT subscription, you can actually hook that up uh here. So, essentially, you're chatting through, you know, Codex through uh um through Cozy. Um so, yeah,

[13:06] it's pretty cool. And there's definitely a lot that you can do manually in here as well. So, we have like fully Well, not fully rigged character, but a rigged character here. So, we can move uh our character around if we want to, if we want to take manual control. Or, obviously, um we can just like Yes. Yes. Yes, you can take manual control. Um

[13:26] Well, where did his head go? Um, yeah, so um um, but obviously if if you don't want to go through the hassle of, you know, uh, Pinocchioing in Pinocchio, uh, you can uh, you can just ask your agent to do so. And you know, it's pretty fast on top of it. Uh, here I had it build me out a little private investigator set. Um, you know, the plus side here is that, uh, you know, it's it's all browser-based essentially. So, um, there's no you don't have to worry about GPUs or anything like that. It'll run basically on anything. Um, one little thing that I do want to note here

[14:00] is that, um, just cuz it took me a minute is that you have to, uh, right click on your mouse, uh, in order to move the camera around or WSD AD around as well. Um, and then again, you can, um, you can animate in here as well. So, if we find an angle here, um, you know, I just had, uh, ChatGPT do this, but we get a walk cycle coming up to the murder board here and then when we turn around our detective behind the desk stands up and walks over to the window. So, from here, I could have it go out and, you know, cut angles, uh,

[14:33] for me onto the timeline so that we essentially, you know, get all of our, um, our shots in here and then, you know, comp them together all into one thing. That said, one of the issues that you might run into and something that I was talking to friend of the channel PJ Ace about, um, is that, you know, when you go this route that a lot of times your, um, your outputs, you know, when you run them through something like Cines or Mini Max are going to kind of look a little bit on the stiff side. So, PJ actually ran into this issue with his latest, uh, short AI film and, um, the

[15:05] solution, uh, is kind of amusing. So, after pre-vizing the entire, you know, uh, short film in Blender, the solution was to take screenshots of that pre-viz and then lay it out into a storyboard. So, in in some ways, you know, what we've kind of gone all the way around the circle uh, to is uh, is we're back to image to video. Now, that said, of course, you know, the bonus in going this direction is that you can then choose, you know, essentially the you know, the shots, the exact shots and composition that you want. And if you're using something like like Cozy Clay over here, you can even

[15:41] change out things like you know, essentially the the lens length and the aspect ratio and all of that stuff as well. So, I mean, definitely a lot more flexibility in terms of what you're trying to accomplish as opposed to just relying on say like Nano Banana or GPT image 2. So, you know, I mean, any way that you want to use it, I do think that Cozy Clay is definitely worth a download. Again, totally free. Pinocchio as well, if you don't have that already

[16:05] as well. You know, overall, I again, I don't think it's for every shot. It's definitely not a Blender replacement, but I think that if you have a very specific shot that you're looking for or if you just want to work and iterate very quickly, Cozy Clay definitely worth it. Rounding out with some quick hits, Ideogram just released their first vector model. So, those of you who work in SVGs in Illustrator, you can now generate on Ideogram and you know, do all of your Illustrator-y stuff. Have not gotten a chance to play with this. I always feel bad with Ideogram. I'm always like, "Hey, they just released a

[16:38] thing. I haven't gotten a chance to play with it yet." In other news, it looks like SeaArt Dream 5.5 is coming out. This is ByteDance's image model. This information coming via Emily, who has always been really good about you know, and reliable about info drops. So, it looks like this will come up to 100 layer seeing decomposition. So, up from 10 that was in the last version. That's that thing where you can, you know, sort of layer foreground and background

[17:02] objects. 80 rounds, 80-plus rounds of drift resistant local editing. 4K. Yeah, I mean, we'll see. I mean, like always been a bit mid when it comes to the SeaArt Dream models, the image models, but you know, always willing to take a look Uh and you know, hoping for a home run. So, that's it for this week. I will be back next week again with the overdue look at Kling 4. Plus in the background, I am still cranking away on Planet Hell. So, hopefully have that for you guys next week. As always, I thank

[17:36] you for watching. My name is Tim.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt