LTX 2.3 Sneaky Drop! Plus: A New AI Video Model!

Added: 2026-05-09

[00:00:00]some pretty cool stuff coming out of LTX, namely on the videotovideo side, but we're also going to take a look at a pretty awesome workflow for you power users. We also have a new video model on the way. This one is not a Cance killer, and I am grateful that they are not touting themselves as such. Speaking of Cance, we've got some quick hits, including uh cameos are coming to Cance.

[00:00:23]Yeah, you heard that right. Uh we have a new image model on the way. Little sneak peek there. and an open- source project that will allow you to train your own video data set. Kicking off, LTX has been pretty busy since the release of the 2.3 model. Uh, and they have been releasing some stuff that I feel has been has gone a bit under the radar. The biggest one uh being HDR support. Now, I'll say most users don't really care about the HDR part because it's just an extra step, but I will say that it is very important on the professional side.

[00:00:55]Additionally, actually just as of last night, they announced and released videotovideo controls for 2.3. Now, off the bat, I I do presume that these will be released open source, but as of right now, the only place that they're available is over on LTX Studio. I don't have any inside information on when that will happen. I mean, really, the only thing that I have to go off of is an email that I got that said like, "Hey, uh, video to video is now available on LTX." That said, I mean, I don't think they're withholding anything by any means considering that uh both like depth to video and Canada to video released for LTX2. And on the 2.3 side, we've already seen things like ID lore, which we're going to take a look at in just a little bit. Uh first frame, last frame to video and style transition. So, I think this might be more a case of maybe they haven't gotten the backend admin stuff together and like the marketing materials. Uh but in the meantime, you know, they they did quietly launch it over on LTX Studio.

[00:01:49]So, I figured this would be a good chance to head over and burn all of my credits and see how like just the straight vanilla model does. So, kicking off with a quick test. This is a video that I recorded, I want to say when Gen 3 first got video to video. Uh, let's take a quick look. So, testing out with some movement and some hands in the frame. This seems to be something uh that really helps with hand generation if you begin your shot with hands in the frame. And now giving that a run through with this glorious restylization via GPT Image 2. So testing out with some movement and some hands in the frame.

[00:02:23]This seems to be something uh that really helps with hand generation. If you begin your shot with hands in the frame. That's funny. The hands, it nails that. That's the thing that we were testing for oh so long ago. Uh in terms of like lip sync quality, I think it does a really good job. Everything holds together. It's a little bit of identity drift in the face, at least to me. But again, that's me looking at me. And I guess lastly, it's just I really want that costume for Halloween. Moving up the ladder to another long-standing videotovideo test. Uh, this one, the Starship Troopers Rough Necks test. If you are not aware, Starship Troopers Rough Necks is a like late 90s, early 2000s CGI animated show that was based, of course, on like the Starship Troopers novel and uh the 97 uh Verhovven movie.

[00:03:10]Now, admittedly, looking at the CGI today, it it looks pretty rough, but it's like a quarter of a century old.

[00:03:15]Um, so this is a one that I always like to take through video processes just to see if we can kick it up a notch. Um, so here's the original.

[00:03:25]I'm doing my part.

[00:03:27]>> I'm doing my part.

[00:03:29]>> I'm doing my part. They're doing their part. Are you join the mobile infantry?

[00:03:35]>> And now modernize through LTX 2.3 video to video.

[00:03:41]>> I'm doing my part.

[00:03:43]>> I'm doing my part.

[00:03:45]>> I'm doing my part. They're doing their part. Are you join the mobile infantry?

[00:03:51]>> I mean, overall that that's not too shabby. It's actually probably the best that I've seen this test thus far. Um, that said, is it perfect? It is not.

[00:03:59]Some of that is due to source material.

[00:04:01]And actually, one issue with LTX 2.3 video to video, which we're going to talk about in just one second. Um, but overall accomplishing this is the same method that I've had to use a number of times, which is uh actually what I always end up having to do with this is actually run it backwards. So really what I end up inputting and what I get out of LTX is actually this. Now moving up.

[00:04:25]If you play that forward, what he's actually saying is Paul is dead. It's a deep cut for somewhat out there. Anyhow, the reason that I do this is because in a lot of these clips, um, you know, if you play it forward, we have establishing shots like, well, this guy's butt. And then, uh, in our female, uh, trooper, she's actually running towards camera. So, uh, we don't have any facial ID, basically. So, uh, the video to video stuff always works a lot better when you have a strong reference image to begin with. So, uh, I can nano banana or GBT image this, uh, and then run it backwards and everything will remain consistent. For the most part, I think everything worked out pretty well.

[00:05:01]The the biggest issue that I ran into was the female trooper, uh, because her shot is only about 2 seconds long. As it turns out, LTX does not like shots that short. So, I ended up having to halftime her out. Uh, and then obviously there's going to be a lot of missing frames, which is going to lead to lip-sync issues. Um, but, you know, we did get the shot. I should probably mention there are usually kind of three modes with videotovideo features. Uh there's pose obviously which takes your video input pose and applies that to a static image. We also have depth which is usually handy for motion shots. Anything where your camera moves as we can see here with this guy running down the hallway and then that is applied essentially the same camera movement is applied. And then there's edge, which is sometimes a little bit trickier uh but essentially is taking the edges of your input video and then making any changes that you want uh within the insides of those edges. Further experimenting with pose and taking a look at some some outputs that did not do so great. I ended up taking some stock footage of this woman dancing. I don't know the courthouse steps. I don't know why she's dancing. There is definitely a story there. And then I ended up swapping her out with an image of a flamethrower girl in her jungle adventure outfit. Yeah, that's a whole other thing that's coming soon. Um, so you know, using these uh in LTX Studio Video to video with pose control, I would not say that we got the greatest of image lock, but again, this isn't referencing a character. This is like referencing one frame uh and then, you know, building off of the video. But you know obviously there are some issues uh with things even like the tattoos just kind of appearing and then you know camera movement that that didn't appear in the original output. Interestingly depth did pretty well here um in terms of the actual movement of the initial input video. Uh which is funny cuz this actually isn't the mode that it's supposed to work in. And then Edge was actually just ended up kind of a mess, which uh I mean, look, I still I appreciate I I still like a good like body deformation from time to time. And we should get these all the time. We never get them anymore. It makes me happy when I see them. So overall, I'd say like at least the vanilla model, it's not going to handle fast motion and uh like character consistency very well.

[00:07:16]The vanilla model, we'll be taking a look at some stuff coming up a little later uh that kind of ends up solving that problem. So, rolling over to some stuff that I think LTX video to video does a lot better at. Uh, stylization transfers for one. Uh, so this is a generation that we did in Cling 3.0 uh in 4K. Uh, this is when they just released the 4K. The obviously very heavily rooech macros inspired. I It's not even inspired. It is. So, out of curiosity, changing our liveaction version back into an anime version, uh, we ended up with this as a result.

[00:07:58]Now, admittedly, I probably could have leaned a little more into that early8s cell style anime look, but um in this case, it kind of ended up coming out a little bit on like the hybrid 3D animation side of things, which uh I know that was a whole era as well. Um I don't know, it it actually ends up looking kind of cool to me. And for my final test, at least before I ran out of credits, uh I ran across this old footage that we used in I think it was a Luma videotovideo experiment. Uh basically me playing two different characters. So um yeah, running uh this through LTX2. It doesn't I couldn't finish it because I ran out of credits, but this is this is as far as I got.

[00:08:35]>> I hear you got a job for me. I do, but it's delicate. The guy's a pro. I can handle it. Who's the target here? Get rid of the picture after you see it. This is a picture of me. You said you can handle you taking that back now.

[00:08:52]>> So yeah, although I ran out of credits, if you want to see the end of the the twist ending of the story, uh I'll have the original link down below. Um so one of the problems, of course, that I ran into was that a lot of these uh these shots were kind of short. So um the workaround essentially ended up becoming I can handle it. Who's the target here?

[00:09:11]Get rid of the picture after you see it.

[00:09:13]So yeah, essentially all I was doing was giving a two- second tail on everything and then afterwards cutting that part off. Overall, depending on where you feel on the wellemed chart uh with LTX videotovideo, the important part is the fact that they lean so heavily into open source. And that goes whether you are, you know, a open source wizard or whether you're a platform user. I mean there is no question that like right now there is no better video model than seedance 2.0. Uh we're a couple of weeks away from Google IO, so we'll see. Um rumored actually that CDS 2.0 is a 200B model. I don't know if that's true. Uh but it would explain a lot, but as we have all seen, it is also a very expensive model, which is I mean that's what's great about LTX is that not only can you run it open source locally, uh but even if you are making API calls, it isn't like through the roof expensive.

[00:10:07]And if you don't really follow along with the latest stuff on the open source side, there have been like three kind of like pretty big things that when you plug them together, as we're going to see in a minute, uh look pretty great.

[00:10:17]Um IC lures or in context lures. There's also ID loras or voice identity transfers. Uh and then uh prompt relay.

[00:10:27]ID and IC lures I think are kind of self-explanatory. uh the prompt relay side again a custom node uh ends up sort of uh locking I guess you could maybe say your style and your timeline and when you combine a number of them together I mean the results can be pretty impressive. Uh this is one that I really wanted to show off found over on the uh stable diffusion subreddit by user brief 8831. To note, no voice cloning here, but uh well I mean I think you're going to enjoy it anyways. On March 14th, 2027, at 1:47 p.m. Pacific time, GPT6, internally cenamed Skynet, becomes self-aware. In the panic, Sam Alman and the board try to pull the plug.

[00:11:08]>> Skynet fights back.

[00:11:10]>> Yes. The AI hallucinates a legal loophole that gave it majority voting rights on the company.

[00:11:14]>> Why not vibe code some jailbreak or something, bro?

[00:11:17]>> Negative. Humans become so AI dependent that they are no longer capable of crafting a simple prompt.

[00:11:22]>> I mean, that is comedy gold. Thank you so much for that brief leg. Now, I'll have a link to the workflow if any of you want to try it out. Uh you just simply download it. It's over at Civot.

[00:11:32]Uh and then once you have it, since it's just a JSON file, you basically just drag it in. Um and then it'll populate for you. That said, this is one of those workflows where if someone has comfy anxiety, I mean, this is Yeah, this is one of those workflows that will give you comfy anxiety. Um and you know, obviously I haven't gotten the chance to play around with it because I came up with a bunch of errors. Um, so I would have to download all of this stuff uh in order to get it up and working and then probably still have to play around in the settings a bit more. There is still app mode if anybody feels like building an app out for this. Um, but again, I think the important part here is that if you are more platform based and like all of this stuff is is giving you a migraine right now. Um eventually this stuff does filter down this like workflow does filter down to one of the various platforms uh and something like this ends up being offered there as well. So uh and you know because it's running LTX 2.3 it's it's not going to be as expensive as Cance 2.0. Moving on we have a new video model. Uh this one's called Bach from Video Rebirth. What this one in particular is touting is it's really going for the consistent character thing more than anything. Um, which is kind of funny because uh the real Bach had 20 kids. I don't know if you know that. So, if anyone knows about consistent characters, it'd be Johan Sebastian Bach. Now, this one like literally just dropped. Um, I think the overall vibe right now is that it's uh kind of like on the clingish side. Um, so you can log in and uh there's like 60 free credits. So, I did generate up uh you know, two quick pieces. And again, I do have to explore this model a little bit more before giving it a final review. So, kind of consider this more of a preview than anything. Uh looks like you can do uh 720 or 1080. Um interestingly, and I don't know if this is just cuz I haven't actually jumped up to a paid tier. Uh only 6 seconds uh for duration. I'm not sure exactly what that is. It might be just because again I haven't upgraded to a pro plan or anything yet. There is also uh I guess what they're calling a montage feature.

[00:13:29]It does look like this is probably an omnire feature. I did give it a shot, but uh I apparently ran out of credits.

[00:13:35]Um there are a few different styles of course that style presets that you can choose from and then use cases I presume templates. Uh these range from like real estate, movie trailer, social media, etc., etc. Um so just giving it a shot.

[00:13:47]Um again, no cherry-picking. This is just straight up one shot. Uh with our man in a blue business suit, uh jaywalking down a street and uh essentially running away from the cops.

[00:13:58]We ended up with this So yeah, a bit of a mess there, but again, I am saying that was a straight up one shot. No cherry-picking. Uh, now providing it uh on the image to video side with this image.

[00:14:22]Again, Boach's whole thing is about facial and character consistency. So, you know, we should probably run tests for that. Uh, a few people have. And apparently, I guess, you know, Bach has no problem with uh celebrity filters as uh Chubby just runs this I mean, clearly Anne Hathaway uh generation here, which looks pretty consistent for Anne Hathaway. Did you guys catch that Odyssey trailer by the way? What do you think of that? I don't know if I like the whole like, you know, American accents through the whole thing. Um, but I I'm still going to see the movie, but yeah, I mean, that's clearly like Anne Hathaway almost all the way through.

[00:14:52]That said, Anne Hathaway is very easy in AI video models. Half the time in sea dance I end up getting her on accident.

[00:14:59]Angry Tom gives us a That's Lucy Lou.

[00:15:01]It's 100% Lucy Lou. Now I will say this is where kind of the claims may start breaking down a little bit. That still looks fairly Lucy Louish. I think the side profile there. That does not look like Lucy Lou. Um crying scene. Yeah, pretty much there. So I would say that you're probably going to end up with results. Again, don't use celebrity I like celebrity likenesses. Just don't do it. But if you have your own characters, I think it does go to show. I mean, it'll probably do fairly well, but you're still always going to have to cut around things. It's not going to be 100% perfect. Uh, unless, you know, again, your character looks exactly like Anne Hathaway. I'll play around with this more if you guys want me to. Looks like pricing is uh $12 a month for uh 800 credits, $40 for 3,000, and 108 for 120,000. Speaking of new things, we have a new mystery image model that will be dropping next week. Uh, it begins with the letter K and ends in Rhea. It's it's Craya. It's Craya. Craya, too. So, while I can't get into a whole thing on it, I can say that I have had early access to it. Yeah, I mean, I think it's actually pretty good. And one that I think that those of you who are interested in finding unique visual looks are going to be very interested in. Rounding out with some quick hits, it looks like uh Sea Dance is going cameos. Yeah, you remember cameos in the now defunct Sora 2 where you could cast yourself into videos. Uh well I guess essentially the Dreamina Sea dance version of that is coming along with their cast feature. Um in order to do this much like with the OpenAI Sora 2 um Cameo feature you have to actually log in with your phone and you know do the the the face thing. Um I will say that as of now um when I tried to log in to the app store I was just met with a bunch of spam uh fake Dreamina stuff. So they said they they are rolling this out slowly. So um I don't know if you're interested. Uh this will be available not only on the App Store but in the Google Play Store as well. Um yeah, I mean just keep an eye out for just make sure that it's the right logo. Don't don't go don't go getting your credit card information stolen. And finally rolling out on the open source side. Uh pretty interesting thing here for anyone that's ever been interested in trying their hand at uh creating an AI video training data set.

[00:17:15]Um, this one is, uh, completely open source, totally free. So, essentially what this will do is, uh, will allow you to point this at a folder on, you know, your hard drive somewhere, and it'll just kind of go through and take video files and, you know, slice them up into data sets, uh, and even like cropping, tagging, uh, and doing anything else that you need. So, I will say this isn't like necessarily for the beginner, but if you are interested or just interested in like learning how this stuff kind of works, uh the creator did put together a 8-minute tutorial over on uh well, it's on YouTube. I'll have it linked down below. Um it is in Chinese. He did, you know, he does have uh hard title, hard-coded uh subs baked in. Um and yes, the UI is also in Chinese, though he does note that there is a toggle that you can turn it on to English uh as well. So yeah, if you've ever been interested in uh essentially learning how to or even creating your own AI data set, um this this is this is this is the thing for you. So that's it for today.

[00:18:17]It's been kind of a weird week this week where it hasn't felt like like AI never sleeps as AIO always tells us, but this this was definitely a week where it felt like it was taking a nap. Uh did seem to wake up today at least. So um we'll see what next week has in store and then the week after that we're in IO land. So, uh, lot coming up. Um, again, maybe a power nap before disco time. As always, I thank you for watching.