# Kling 4! Gemini 4! ElevenLabs v4! Ideogram 4.5! Everything Goes Four!

Channel: Theoretically Media
Video: https://www.youtube.com/watch?v=U9tFP81PSbQ
Duration: 14 min
Language: English
Words: 2454
Transcript page: https://viewrankai.com/tools/youtube-transcript/U9tFP81PSbQ

---

[0:00] Well, today four is your lucky number with big AI updates across all the modalities. Today I'm taking a first look at Cling 4 with some tests to see where it lands. Speaking of four, Google is back with Gemini 4, Return of the King. We'll see. Plus, more four themed updates from Ideogram and 11 Labs. Yeah,

[0:21] four. It's a bit of a theme today. Kicking off. Well, it has been a minute. Cling. Uh the last big version jump for from the big K was back in February with the release of Clling 3.0. Now we did get 3.0 Turbo in June, but yeah, as far as a whole new generation goes, yeah, I mean it has been a while. Obviously, there has been a lot of anticipation for

[0:44] the V4 model and as of now, we have it. Well, kind of because yes, Cling 4 has entered early access, although the wider rollout is still yet to come. And yes, sadly, uh, this is the Flash model. Flash's initial roll out is for Ultra subscribers on the annual plan. To be fair, this is not uncommon for Cling. And, uh, we'll see a full roll out with the, you know, the larger Big Daddy

[1:10] model uh, planned sometime in October. But in the meantime, well, let's take a look at at 4 as it stands today. Kicking off on the texttovideo side with our man in the blue business suit about to get a jaywalking ticket.

[1:42] So, overall, I mean, for text to video, not not so bad. Um, there is some like weird actions in there, like uh the police officer with this motion. I don't I don't know what that is. I definitely didn't prompt for it. I think he's like trying to ID him from a photo. Again, I didn't prompt for it, but uh yeah, I mean, overall, pretty solid. Uh let's move on to our other standard image to video test of our uh FBI agent uh having

[2:07] coffee in a Pacific Northwest diner.

[2:14] That's a damn fine cup of Joe. But how's the pie? It's really good. It's an old recipe from the log lady who got it from the owls. Overall, not too bad. The reason like the benchmarks for this test are really two things. One, can it handle all of that dialogue, which in this case it definitely did with room to spare. Uh and secondly, that in the initial image, obviously the waitress's back is turned to us. So, you know, when she turns around, the model essentially has to invent her. Uh, and I think it did a pretty good job here and more than that managed to hold stylistic consistency,

[2:55] you know, throughout the scene. Now, I'm not going to discount the fact that you're probably going to see some decoherence in these, but remember, we are looking at the fast model. Uh, we'll talk about the, you know, the big the big daddy model in a little bit, but in the meantime, moving on to the Omni side of things. You know, each of the models all have sort of different takes on a character in their omni and reference modes. I do have to say when it comes to flamethrower girl, uh, Cling's like omni version of her has always been one of my favorites. So, uh, you know, I figure

[3:24] let's let's give her a roll and bring her home in version 4. Oh, you guys didn't forget my birthday, but I got to blow out the candles for my wish.

[3:42] Nailed it. So, biggest issues with this generation are that when she, you know, when she first walks in, obviously the candle is lit on the birthday candle. Additionally, I'm sure that someone out there is upset that the candle itself didn't melt. Uh, you know, when she basically like flames over the entire tabletop. At the same time, of course, I want to point out that it's a pretty silly prompt. Uh, the other thing I did want to note with this particular generation is the uh, you know, the somewhat improvement in spatial awareness. uh when she walks in uh our kitchen sink is here and then later on

[4:12] in the generation kitchen sink is in the same location. So I we pretty well done. Speaking of locations, I figured this would be a good chance to check in with uh Tuesday Sinclair uh Mallaloy who I did cast as Tom once again. Something kind of interesting happened here in of course like in Mallaloy's office uh for

[4:29] our cinematic noir scene.

[4:37] Mr. Mr. Mallaloy, I hear you're a detective and I'm in need of one. You heard wrong. I'm done taking cases. I think you'd be interested in this particular one. It's a murder, but the thing is it hasn't happened yet. Oh, yeah. Who's getting iced? Well, that's just it. You are. Well, you have my attention. So, a few things to note about this particular generation specially. I mean, I think it holds up. um you know it's a it's pretty simple in terms of like camera setups but um you know uh ey lines match so there's that do want to note that interestingly for whatever reason uh Kling's Omni model decided to

[5:17] put um Tom into a gray jacket uh that was actually that was not provided uh nor was it called out for in the prompt. I will say that video quality albeit it is only 720 right now is pretty good personally I mean I do find the performances to be a little bit on the flat side. That said, the big win here is that Cling 4.0 does seem to have solved that lip sync drift problem that

[5:42] was like just plagued Cling 3.0. It's really good. It's an old recipe from the log lady who got it from the owls. Now, on the performance side, and this is something that I do have to dig more into, but I think you can improve it by utilizing a method that I've been using a lot more, which is uh, you know, kind of combining key frames with an omni reference. you're kind of getting the best of both worlds. Uh, and for uh, you know, key frame referencing, uh, 4.0

[6:10] does allow up to 10 references. You're standing in my light. I'm looking for a monster that took a little girl. You're not the first to ask tonight, but I think you're the one that will bring her home. Here, you'll need this. How much? You'll pay me on the way back. Where do I find it? Down the alley to the left. Whatever you

[6:41] do, don't take off the cross. So, aside from some edit timing issues at the beginning of the clip, I mean, I do think that this one holds together pretty well. Uh, you know, my my running theory is that by providing it key frames uh for your characters, it kind of gives a the model a better grip on what the performance should, you know, should should feel like. So yeah, I mean

[7:03] overall an early look at Cling 4.0. Maybe not the Sea Dance Killer that some had hoped for, but I'll definitely roll back with a deeper look at everything once, you know, essentially the big daddy model drops. Meanwhile, and staying on video for just kind of a quick one. Uh Heyen have released Hey Gen Video. This is their first uh general purpose video model. It is another one that is built off of Miniax H3. We are definitely seeing a lot of those. But hey Jen's pitch here is that it is cheap and fast. And I do mean really cheap. Pricing starts at 1 cent a second. I do want to point out that this

[7:37] is just a 50% off launch sale going through the end of October. So, uh yes, the the the regular price is 2 cents a second. I do want to point out that this 1 cent offer uh is only 480. So, you're like, "Ah, I got you there." Um but the 720 version or well 7.68 um is actually a cent and a half during the sale. So, uh, you know, 3 cents normally. So, I

[8:02] mean, I really can't get mad at that. Now, I haven't tested it out, but you know, it is a Miniax H3 fine-tune, so you know, you can kind of figure where it's going to land. Uh, it is currently available via their API and, uh, comfy UI and Hey Gen's web interface will be coming later. Moving on, the Giant Awakens. Yes, Google has returned with Gemini 4 sort of. Google have announced Gemini 4 Argon, which I'm not going to lie, initially when I saw it, I thought it it was uh I thought it was Aragon and I just kept thinking to myself, and that is bold. I mean, very bold, but no, it

[8:37] is Argon. And you know, to note, once again, this is more on the announced side as opposed to the widely released side. Uh and at the moment, you know, what we're looking at is benchmarks. Uh the benchmarks are impressive, but also they are benchmarks. Now, what do those benchmarks say? Well, of course they say that Gemini 4 is ahead of OpenAI and Anthropic. What I mean, what else do you think it's going to say? Now, I do have to point out that when it comes to all of these benchmarks, that one of the secret powers of, you know, Gemini in general is its video understanding. And

[9:06] in this case, uh, 4 does look to be pretty impressive, scoring a 91.7 on LVB, which is the test for long- form video comprehension. Now, as to when this rolls out more broadly, well, we we do not know, but I mean, it's no secret that Google has been kind of stuck in flash land, at least since Google IO. So, it's nice to see a flagship model finally on the verge of being widely released. And I mean to me the bigger thing is what what's behind it, what's going to follow it because the big guy model will create some momentum behind you know the creative stuff that we use

[9:40] stuff like you know Nano Banana 2 Pro possibly a Genie update and if we're lucky Google decides to get back into the video race and we'll see more of a major push with the like Omni model or VO or omni VO or whatever it's I mean who knows with Google naming conventions. Other spot that I'm curious to see what Google has in store is to match this new class of like aentic assistants rolling out like Muse and Open AAI's dot. Uh they had previously released Gemini Spark, but I mean let's face it, that wasn't it. In the meantime, rolling out of the Google section. And speaking of Nano Banana,

[10:13] poor one out for the OG. Yeah, Google has scheduled the shutdown of Gemini 2.0 2.5, I'm sorry, Flash Image, uh which was like the our first taste of the banana. This is like Nano Banana 1. Um, so that is going away in the Gemini API on October 2nd. Moving on, another quick hit on the image side. Uh, Aiogram 4.5 has released and the focus here is on more precise editing. To be honest, I haven't gotten a ton of time to play around uh, with 4.5 as of yet. But, I mean, Aiogram has been it's always been a really solid model for me, but I definitely wanted to give the editing

[10:50] capabilities a shot here. So, I pulled up this image uh kind of kind of funny. Uh this was generated in audiogram about 10 months ago and clearly around the uh the time of Gemini uh 3's release. Um you know, overall a lot of the well 10-month-old uh the AI image problems, a lot of like garble text here. Um it did get almost almost whatever that prompt was correctly uh minus the fact that it's we're so ocoped. So, making some text edit corrections on this uh and uh throwing in a image of flamethrower girl because why not? Just updating that uh that user profile. Um yeah, I mean

[11:30] pretty solid. Yeah, well done audiogram. I'll definitely keep writing some tests on this one. Um Adio, they've always been super solid. I definitely don't think that you should sleep on this one. So, um yeah, I'll let you know uh as it works its way into my workflow. Finally, rounding out on the audio side, 11 Labs have released 114 along with a turbo

[11:50] version for real-time voice and agents. Dialogue and sound have obviously been a bit of a bane uh at this point for AI video. So, it is nice to see 11 Labs, I mean to be honest, kind of catching up with seed audio 1.0. So, with 11 Labs V4, you can put in well natural language

[12:07] direction uh directly into the script. And 11 Labs does say that the new model is better at following those instructions, including things like laughter and sound effects. So, I figured this would be a good opportunity to return to a test we ran uh a while back. I think that we did this with Seed uh 1.0. Um, essentially the opening to a

[12:27] high fantasy uh audio novel. At the northern watchtower, snow whispered through the arrow slits. The old watchman kept vigil beside a green flame. God, it's freezing. Why are we guarding a fire that gives no heat? That's not what the flames for, lad. Then what's it? Don't stare into the flames. The boy looked away. Something in the

[12:57] fire kept watching. Overall, this is another one that I'll be digging more into. I think that audio references in Omni Video models are going to be a pretty big thing as it does seem to be a while for video models to catch up on the audio front. 11 Labs V4 is available now, including on the free tier. So, yeah, head over, go check it out. So, that's it for today, but overall, I mean, impressive updates across the board. Uh, you know, starting to get a little bit toasty here in AI Fall. And the leaves haven't even started changing color yet. Although, I guess once Zuck has Metav Vision

[13:30] strapped to all of our faces, yeah, we could probably just edit them in. As always, I thank you for watching.

---

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