This Open Model Makes Video AND Sound. It Claims To Beat LTX 2.5. I Tested It.

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: This Open Model Makes Video AND Sound. It Claims To Beat LTX 2.5. I Tested It. · Channel: Prompt Engineer 48 · 07:23 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI content generation, Local and open-source models

Summary

The video tests Kandinsky 6.0, an open model that generates video and audio together, against LTX 2.5 using seven prompts. Kandinsky produced some strong results, but LTX won four comparisons, Kandinsky won one, and two were close; LTX was also more than five times faster in the test. The creator notes that the test was small and that the paper’s reported comparison used different resolutions. Kandinsky’s MIT license may make it preferable for some users, while LTX’s community license has a revenue threshold.

Key points

Kandinsky 6.0 has Light and Pro versions with 3 billion and 29 billion parameters, respectively, and generates five-second clips with audio.
Its video and audio streams use cross-attention to keep sound and picture aligned; a separate super-resolution model produces full HD output.
The creator used the distilled Pro version and rented an H100 for testing.
The 16 GB GPU preset still requires about 58 GB of system RAM for Pro, and the model can be slow.
Kandinsky’s built-in prompt rewriter altered the requested speech and added unrequested flute music, so the creator disabled it for the comparison.
The seven-prompt test gave LTX four wins, Kandinsky one, and two close results.
In the creator’s test, Kandinsky’s median generation time was 289 seconds, compared with 54 seconds for LTX.
The creator cautions that the test used only seven prompts and one seed each, and that the paper compared different resolutions.
Kandinsky is released under an MIT license; LTX uses a community license that is free only below $10 million in revenue.

Chapters

Start	Chapter	Summary
00:00	Introducing Kandinsky 6.0	The creator introduces Kandinsky’s joint video-and-audio generation, its model variants, and how its video and audio streams work together.
01:35	Setup and comparison method	The creator describes running Kandinsky on an H100, resolving a driver issue, disabling its prompt rewriter, and setting up a comparison with LTX.
02:37	Seven prompt tests	The creator compares Kandinsky and LTX on seven prompts covering action, emotion, camera movement, text, and speech.
05:50	Results, caveats, and licensing	The creator summarizes the scores and speed results, discusses limitations in the test and paper comparison, and weighs the models’ licenses.

Quotes

the results did not match. — 00:00

For the real test, I turned this off. — 02:06

LTX took four, Kandinsky took one, and two were close. — 05:50

To be fair, this is seven prompts and one seed each. — 06:21

If you want speed or the best look from my seven, I would go with LTX. — 06:52

Concepts and entities

Kandinsky 6.0, LTX 2.5, RunPod, Hugging Face, Litrix, Whisper, MIT License, community license, MiniMax H3, Seedance 2.0

Full transcript (original language)

00:00 All right. In this video, we will see Kandinsky 6.0, which makes the video and the sound together in one open model. Now, mind you, the paper says it beats LTX 2.5. So, I rented one H100 and tested that claim myself, and on my seven prompts, the results did not match. Let’s go ahead and listen to this first. >> Good morning. This video and its sound were generated together by one open model. >> So, that was Kandinsky speaking, and the lips are moving with the words. This is all one model in one pass. So, what is it? It’s a family of two models, a light

00:31 one with 3 billion parameters, and a pro one with 29 billion parameters. Both make 5-second clips with 44 kHz audio, and that includes lip sync. You can go from text or from an image. Now, the way it works is there are two streams, one for the video, and one for the audio, and they talk to each other with cross-attention, so the sound and the picture stay lined up. The video is made at a smaller size first, and then a separate super-resolution model takes it to full HD. And there are distilled versions that need only 10 steps. In their own comparison, the distilled pro

01:02 is basically tied with the full pro 51 to 49, so that is the one I’m using. Now, the big question, can you run this on your own GPU? They give three presets for 32, 24, and 16 GB cards, and it keeps only two blocks of the model on the GPU at a time. Mind you, on the 16 GB preset, the pro model needs about 58 GB of system RAM. So, that is the real requirement. And speed is where it hurts. For a 5-second full HD clip with pro, their table says about 1,247 seconds on a 4090, and about 402 seconds

01:35 on an H100. That is almost 21 minutes on a 4090 for 5 seconds of video. So, for this, I rented one H100 on RunPod at $3.49 an hour. The setup is just UV, and just you run just setup, then just download pro distill, and then just generate with your prompt. Really simple. Now, there was one thing, the pod driver was too old for the CUDA 13 build of PyTorch. It installs and it said driver too old. The way out is to install the CUDA combat package and point the library path to it and then it

02:06 works. And here’s another thing you need to know. It has a built-in prompt rewriter. I asked for a woman saying, “Good morning. This video and its sound were generated together by one open model.” And what it wrote was just good morning and it added a flute playing in the background that I never asked for. So, you can see that it changed my words. For the real test, I turned this off. Now, on the other side we have LTX 2.5 from Litrix. This is a 22 billion parameter model and it also makes audio natively. You do need to accept the terms on Hugging Face to download it.

02:37 Same prompts, same seed and both get 5 seconds. Kandinsky comes out at 864 by 480 and LTX at 960 by 512. So, this is close to an equal comparison. First prompt, a fight. Shaky camera, one man kicks the other into a wall. On the left is Kandinsky and on the right is LTX. >> [groaning] >> So, Kandinsky gives us two men arguing

03:09 and shoving and there is no kick at all. LTX actually has the hit and pushes him into the wall. This one goes to LTX. Second, a woman opens a letter and a single tear slips free. Now, both of them get the tear. Kandinsky’s face is sharper and cleaner. LTX is softer, but a bit more intense. I would call this a tie. Third one,

03:39 continuous zoom. Earth, then New York, then an office, then a person on a phone. So, Kandinsky goes from Earth straight to an office and skips New York completely. LTX tries every step, but the Earth turns into a strange blob of buildings. Neither is right, but LTX made the real attempt, so I give it to LTX. Fourth, camera control, a dolly down a dark hallway, then an orbit, then

04:11 a crane shot. In the frames I checked, Kandinsky gives a clean dolly, but I do not see the orbit or the crane. LTX is darker and the end of the shot is tilted. So, this one goes to Kandinsky for being cleaner. Fifth, text, a couple kisses on a beach and the sky says, “And they lived happily ever after.”

04:46 >> [music] >> Kandinsky gives a nice kiss, but there is no text. LTX has the text, but it is spelled h a p i y and I can’t see a clear kiss. So, LTX gets the text with a typo. Sixth, my favorite, a professor explaining the Pythagorean theorem on a whiteboard and saying it out loud. >> So, the square of the hypotenuse equals the sum of the squares of the other two sides. >> So, the square of the hypotenuse equals the sum of the squares of the other two sides. >> So, Kandinsky’s whiteboard is complete

05:18 gibberish. LTX has a believable teacher and the formula is close, but not right. And both of them said the sentence word for word. I checked with Whisper. So, this one goes to LTX. And last, the claim that Kandinsky is strongest at, which is speech, a woman in a kitchen saying the same sentence. I ran both through Whisper. >> Good morning. This video and its sound were generated together by one open model. >> Good morning. This video and its sound were generated together by one open model. >> Kandinsky says a sound instead of it

05:50 sound. LTX is exact. It is one word in one clip. So, do not over read it, but it was not the wind the paper suggests. So, the score on my seven prompts, LTX took four, Kandinsky took one, and two were close. And on speed, the median was 289 seconds for Kandinsky and 54 seconds for LTX. That is more than five times faster. Now, here’s the thing. Everything in the paper is self-reported by the Kandinsky team. They say Pro is preferred over LTX on visual quality and speech quality, and those are

06:21 significant. They also say LTX is slightly ahead on audio sync and overall audio, and better on word error rate. And they compared Kandinsky at 512 by 768 against LTX in its full HD setup. So, it is not an equal comparison. To be fair, this is seven prompts and one seed each. So, it is a small test, and I only measured speech with a transcript, not a listening test. Against the closed models, the paper itself says it is mixed. Minimax H3 and Seed ends 2.0 are still stronger on visuals, and clips are

06:52 limited to 5 seconds. So, what would I use? If you want speed or the best look from my seven, I would go with LTX. But Kandinsky is the one with the MIT license, and LTX is under a community license that is free only under $10 million of revenue. And the Kandinsky weights are already out on Hugging Face with the instructions on the GitHub page. And the whole test setup both models 14 clips cost me about $3 on RunPod. If you’re interested, you can go ahead and try this on your own system, and let me know how it goes. I’ll see

07:23 you in another one.