NEW LTX 2.3 Distilled 1.1: Free Kaggle Notebook Setup & Complete Testing

Ajouté : 2026-05-25

[00:00:00]I like things that go deep and leave a lasting impression.

[00:00:03]Like a well-executed crime scene. Good thing I brought a shovel.

[00:00:08]For the landscaping, of course.

[00:00:11]I usually prefer a tight grip, but tonight I think we should just let the body count rise. Remember those LTX 2.3 Kaggle notebooks we were playing with recently? You guys absolutely loved the text-to-video, image-to-video, and audio capabilities. But let's be totally honest, those notebooks were built on the old distilled model, which was painfully slow and honestly delivered pretty weak video and audio quality.

[00:00:37]Well, LTX recently dropped a significant update called distilled 1.1. It's not an entirely new model architecture, but a refined version of the old 22 billion parameter weights specifically optimized to slash generation speeds and dramatically clean up the audio-visual output. I've put together a brand new notebook for you guys that runs much faster than the previous notebook. But before we jump into the setup, I want to take you through every single test I ran showing you the exact prompts and what I was testing for so you can see the difference for yourself. Under the hood, this version runs as an eight-step distilled model on a conditional generation factor of one. Over on their official Hugging Face repository, you can see they've rolled out a dedicated 1.1 distilled file, a separate 1.1 Lora, and even its own standalone 1.1 upscaler. That whole updated ecosystem is exactly why the final spatial detail and synchronized acoustic layers feel so much sharper. Now, a quick heads-up on resolutions before we look at the clips.

[00:01:40]I did all my testing at 480p and 540p because trying to generate 720p on a free T4 GPU in Kaggle takes longer and just isn't practical. Let's look at the prompts and see how they actually performed live. First up, here's the prompt for my very first baseline test.

[00:01:57]Go ahead and pause the video or take a screenshot if you want to copy it down, as I wanted to see how it handled raw generation cleanliness and scene cuts compared to the old model. Check the result. I love how contained this is. No escape. I could do anything to you in this backseat and nobody would hear you scream.

[00:02:16]That sounds less like a promise and more like a threat.

[00:02:20]Either way, I'm already strapped in.

[00:02:23]As you can see, the classic LTE artifacts are gone and the audio clarity has improved. The way the camera handles the cut to scene, shifting from the starting environment directly to a clear focus on the male character, is impressive for a first run. Next, I wanted to try a very short prompt to test high-contrast lighting and camera tracking. You can see the prompt on the screen right now, where I simply asked it to create a sports car with reflections. Check the result.

[00:02:59]>> [music] >> For a 480p output, this looks so much cleaner than the previous version. The camera tracks the car beautifully and the headlights reflect perfectly off the wet road surface.

[00:03:10]The text on the neon signs is still total AI gibberish, but the motion and lighting physics are a step up. For the third test, I wanted to evaluate a dynamic, fast-moving cooking scene with complex environmental elements. Here's the prompt on your screen, which describes a male character cooking intensely inside a kitchen. Check the result.

[00:03:40]Look at how he tosses the chow mein inside that Chinese wok with real flames bursting out of the pan. The motion looks realistic and doesn't give off that muddy AI feel at all. Plus, the clip is completely free of any weird morphing artifacts. Moving on, I wanted to step up the complexity with a slightly longer prompt involving cinematic character placement and alternating dialogue. Take a look at the prompt on screen and feel free to pause if you need to read it as I'm testing a dialogue sequence between a male and a female character positioned near a swimming pool. Check the result. You look tense. You really should relax. I don't usually bite.

[00:04:17]Unless you ask me nicely. It's not the bite I'm worried about. It's the venom that comes after.

[00:04:22]>> This one came out in 540p and the overall generation is acceptable. The speech quality and lip sync hold up remarkably well as the female delivers her dialogue line followed cleanly by the male character. Next up is another dialogue test and I gave it this prompt on the screen for a conversation between a man and a woman inside a coffee shop.

[00:04:42]Check the result.

[00:04:43]>> my coffee how I like my internet.

[00:04:46]Fast, strong, and completely [music] unshared.

[00:04:50]>> [laughter] >> The female character is telling the guy about the kind of coffee she likes. I did notice some minor glitchy artifacts around her eyes here and when it cuts to the male character laughing, he accidentally duplicates one coffee cup into two. But hey, since the notebook gives you unlimited free runs, you can just regenerate it until it lands perfectly. For this next test, I wanted to see how it manages a multi-character environment with complex spatial elements. Here's the prompt on your screen setting up a poker game table with a male, a female, and a few other people sitting around. Check the result.

[00:05:25]You're playing a dangerous game holding on to something that big without showing it. Trust me, when I put it on the table, everyone loses.

[00:05:35]This 540p poker scene looks way better and much more refined. The female speaks her dialogue, the camera cuts over to the male as he speaks his, and you can even see realistic smoke drifting through the air. The man is hilariously holding his cards completely inside out, but overall, it's a very solid and usable generation. Next, I wanted to move away from human dialogue for a second and check how the model handles pure 3D generation. You can see the short prompt on screen where I simply requested a 3D dragon with slight camera movement. Check the result.

[00:06:09]>> [music] [music] >> This turned out to be an incredibly impressive generation. Even though I didn't specify any audio elements, the model automatically engineered matching background music and the deep sound of the dragon breathing. The camera tracking is also spot-on, moving smoothly from a distance into a crisp close-up of its face. After that, I decided to test character anatomy consistency and specific physical traits. Here's the prompt on your screen for a fitness center scene featuring a dialogue between a male and a female character. Check the result.

[00:06:49]You know, staring at me counts as a cardio workout, but you still have to pay for the gym membership.

[00:06:55]>> [laughter] >> Visually, the quality is stellar and I don't see any eye anomalies. The model gave the female a highly accurate muscular gym type build. The only hitch is that when the second cut to scene happens, both characters suddenly look exactly identical, which is a classic AI consistency bug. But a quick re-roll fixes that right up. Let's try a kitchen dialogue setup to see if it handles action consistency any better. Here's the prompt on your screen where a man and a woman are interacting in a kitchen environment. Check the result. It's all about how you handle the blade. One wrong slip and someone bleeds out. Don't worry, I always slice precisely where it hurts the most.

[00:07:37]This clip is acceptable and looks great.

[00:07:39]The man is actively cutting up pieces of meat, the female delivers her line, and the male responds.

[00:07:45]If you look closely at the close-up cut to scene, the model accidentally removes the guy's glasses, even though the character model stays the same. Those are the small details you have to watch out for, but it's still an impressive generation. Next, I wanted to look at a different environment with male and female characters. Take a look at this prompt on screen, detailing a dialogue sequence inside a library setting. Check the result. We have rules here, strict ones. If you want to handle the rare materials, you need a very gentle touch.

[00:08:13]Don't worry, I always wear gloves. I wouldn't want to leave any marks before the official check. The dialogue audio quality here is fantastic, and the visual grading is very clean. The camera transition cutting over to the male angle is incredibly smooth, too. The only minor flaw is right at the start, where the female's top seems a bit strangely, but it's acceptable and an easy fix with a regeneration. For the next run, I used this prompt on screen to specifically test complex camera framing and tracking during a spoken scene. Check the result. If we don't [music] survive this mission, I am definitely deleting your search history.

[00:08:51]>> [music] >> Overall, this generation is super impressive with zero noticeable visual artifacts. The camera execution is great, pulling into a close-up and moving toward the male character. The only quirk is right at the beginning, where the female is looking completely in the wrong direction while delivering her initial dialogue before finally turning to face the guy. But the audio video upgrade is undeniable. Next up, I wanted to see how it performs with moody, high-contrast environment lighting. Here's the prompt on your screen for an intimate red-lit bar booth with character dialogue. Check the result.

[00:09:25]>> Most people can't swallow something this bitter, but I think you can choke it down.

[00:09:31]I've survived worse poisons, but this one might actually kill me. We got an absolutely beautiful generation here on the very first try. I didn't do any retries on this one, and both the rich red lighting and the clear audio synchronization turned out incredibly well, exactly as I got it. Let's check another object interaction test in a kitchen setting. Here is the prompt on screen where I asked for a professional chef taking a freshly baked item out of an oven.

[00:09:57]Check the result.

[00:10:00]Fresh from the oven.

[00:10:08]The visual fidelity here looks pretty solid. I didn't see any artifact. The camera tracks the movement perfectly, and the focus stays entirely locked onto the baked product throughout the shot.

[00:10:20]Next, I wanted to throw a stylized sci-fi prompt at it to test voice matching for non-human characters.

[00:10:27]Take a look at the prompt on screen for a friendly little robot paired with specific camera movements and a line of text. Check the result.

[00:10:40]I'm almost fixed.

[00:10:44]This turned out beautifully. The camera glides from a distance into a crisp close-up, and the robot moves its arms after perfectly speaking the exact dialogue I specified, saying, "I am almost fixed."

[00:10:58]Now, for my absolute favorite generation of the entire batch. Here is the prompt on screen for a luxury penthouse lounge scene with a male and female character where I actually wrote some subtle double-meaning dialogue into the text.

[00:11:12]Check the result. I like things that go deep and leave a lasting impression.

[00:11:17]Like a well-executed crime scene.

[00:11:19]Good thing I brought a shovel.

[00:11:21]For the landscaping, of course.

[00:11:23]This one came out absolutely flawless.

[00:11:26]There are zero IR effects. The characters deliver their lines with great pacing, and the over-the-shoulder cut to scene transitions are seamless.

[00:11:34]This is definitely the generation I liked the most. Next, I wanted to see how the model pairs environmental sounds with moving machinery. Here is the prompt on screen for a train passing by a station accompanied by a PA announcement. Check the result.

[00:11:50]Train arriving on platform two.

[00:11:59]The train rolls by the platform perfectly right as the audio announcement states that the train is arriving at platform two.

[00:12:07]Moving on, I wanted to see what happens when you prompt an intense situation, but completely omit the actual text dialogue. Check out this prompt on screen for a roadside argument between a man and a woman. Check the result.

[00:12:19]>> Horace, just nice.

[00:12:23]That was high.

[00:12:25]Straight.

[00:12:26]It's okay for us to talk about this thing. No, that's a >> [music] >> This one turned out pretty bizarre because I didn't give it explicit lines to say, the model forced them to speak completely incomprehensible garbled gibberish. Even crazier, instead of a verbal argument, the characters got physical and started swatting at each other with their hands. It just proves that the model outputs exactly what you prime it for. So, without explicit script lines, things get chaotic. Now, let's jump over to the image-to-video testing, specifically looking at the first and last frame transition feature to see if it improved. Here is the prompt on screen. The starting image is a close-up of an eye pupil, and the ending frame is a deep space galaxy.

[00:13:09]Check the result.

[00:13:25]I ran this for 15 at 540p and you can see the pupil's black hole expands. The camera zooms directly inside it and then it reveals the galaxy scene. The transition itself could be a little smoother, but it's still acceptable.

[00:13:42]Let's try another transition with drastically different textures. Here is the prompt on your screen where the first frame is a classic marble statue and the last frame is a sleek futuristic robot. Check the result.

[00:13:59]>> [music] >> This turned out incredibly clean. Lines appear across the marble statue and it slowly and smoothly morphs into the exact robotic design from the final frame with great structural continuity.

[00:14:14]For this next transition test, I wanted to check environmental shifting. Take a look at the prompt on screen. The starting frame is a frozen snowy oak tree and the end frame is the exact same tree but in golden autumn colors. Check the result.

[00:14:39]The model nailed it. The video starts in the dead of winter. A massive explosion clears the frame and it beautifully shifts into the golden autumn tree precisely as requested.

[00:14:50]Next up is a highly artistic transition test that completely blew me away. Here is the prompt on screen where the initial frame is a bird sculpture sitting on a pile of ashes and the final frame is a bird made of pure fire. Check the result.

[00:15:07]>> [music] >> This is incredibly spectacular. The ash bird opens up and a gorgeous phoenix engulfed in brilliant flames transforms and flies straight out of the frame. I loved that result so much that I decided to run the exact same prompt and the exact same input images a second time just to see the variance under identical conditions. Check the result.

[00:15:47]This second output is just as fantastic, but this time the model introduced much more dynamic camera work circling around the bird before pulling off the fire transition flawlessly. Both generations are incredibly impressive. So, that is the full breakdown of my personal testing and I hope it gives you a solid idea of what to expect. Overall, the visual fine-tuning and audio synchronization are significantly improved. However, if you try to throw tough prompts involving sports, complex dancing, or intense motion physics at it, the results are still not that impressive and there hasn't been much improvement there. It's essentially the same core model with polished visual and acoustic layers.

[00:16:29]Let's jump into the notebook setup so you can run it yourself. Head over to my GitHub repository where you'll find all my builds and open the notebook labeled LTX 2.3 22B distilled 1.1 Q4 for text-to-video.

[00:16:45]Once you're inside the Kaggle interface, the absolute first thing you need to do is open your settings, go to accelerator, and select the dual T4 GPU option. After that, just click run all.

[00:16:57]It takes about 7 to 8 minutes to completely initialize the environment and generate your Gradio link. The script will automatically handle the dependency installation and start pulling the files. You'll see the main GGUF model downloading, which is about 18 GB, followed by the 8 GB distilled Laura and the 13 gigabyte Gemma 3 LLM.

[00:17:20]Because the total size climbs to almost 60 gigabytes and Kaggle's main directory caps you at 20 gigabytes, I wrote a custom workaround that routes all the excess model data directly into the temp folder to prevent any disk space crashes. Just ignore any storage warnings that pop up. Once the logs read all downloads complete, the launch cell will take about 2 minutes to offload roughly 20 gigabytes of data into your system RAM before utilizing the GPUs during actual generation. Once the model's loaded into the RAM, your public Gradio link will appear. Click that link and it will open the familiar Gradio interface. You can input your text prompts directly or switch over to the image to video tab to upload your starting and ending frames. You can set the duration up to 15 seconds, but I highly recommend sticking to 480p or 540p resolutions for the faster speeds on Kaggle since 720 and 1080p will slow down drastically. Leave the seed at -1 for random generations, adjust your prompt strength as needed, and hit generate. I spent a massive amount of optimization testing this notebook to make it significantly faster and cleaner than the old one. When you use it, you will immediately realize how much faster it is compared to the old one. The repository link is waiting for you in the description below. If you appreciate the effort that goes into building these free tools, please hit that like button, drop a comment with your results, and subscribe for more deep dives. I'll see you guys in the next video. Bye. Yo, check the intellect. My mental state's architect. I cracked the code of text.

[00:18:54]Synthesizing data streams, living out the future tense. It's logic on a grand scale. It just makes perfect sense. THIS IS THAT AI QUEST. THE MENTAL POWER FLOWS, unlocking all the secrets that the digital world holds. THIS IS THAT AI QUEST. YEAH, THE KNOWLEDGE is a weapon.

[00:19:07]We building up the future while the rest of y'all are stepping. So, hit subscribe and join the mission. Peace to the whole cypher. Air Quest. We bringing fire.

[00:19:15]>> Air Quest. Yeah, the knowledge is fire.

[00:19:17]>> It's fire.