At Kinetix, we combine deep AI expertise with advanced 3D animation technologies to shape the future of character motion. Leveraging extensive proprietary datasets and advanced synthetic data pipelines, we drive cutting-edge AI research to push the boundaries of how motion is created, understood, and applied across digital and physical worlds.
Learn more: https://www.kinetix.tech
We built a pipeline to create the datasets needed to train learning-based retargeting models at scale.
Compatible with any humanoid character, any morphology.
No clipping, no foot sliding, perfect ground contacts preserved.
Ground truth quality at every frame.
Before this, there was only one open-source option: a narrow Mixamo subset, closed-source, artifact-ridden, and prohibited for ML use since 2021.
Learning-based retargeting is critical infrastructure. Humanoid robots need large volumes of contact-correct motion data to learn how to move and interact with their environment. Synthetic data pipelines for gaming and film need physically accurate motion at scale. The models that will power these systems need to be trained on data that actually reflects physical reality.
That data didn't exist, so we built it.
Read full article 👇
https://lnkd.in/e4a3BXCu
Video generation has made extraordinary progress. Today's models produce cinematic sequences with coherent lighting, natural motion, realistic human presence.
But they all share the same structural limit: the moment physical reasoning is required, they break. "A character catches a flying object while avoiding a running dog" is too complex without spatial planning.
This isn't a data problem. It isn't a scaling problem. It's structural.
Current video models learn the appearance of physics. Not its logic. They have no internal representation of where things are in space, how they relate to each other, or why they move the way they do.
The field is converging on this diagnosis. World models are the proposed answer. But world models built on 2D video alone are still missing a layer.
We think 3D is that layer.
We've published a piece today on why physical grounding in video generation requires spatial intelligence, and what it actually takes to build it.
Worth a read if you're thinking seriously about where AI video goes next 👇
https://lnkd.in/ew_6KCHd
Building a state-of-the-art human pose estimation model requires perfectly paired 3D motion and 2D video at massive scale. That kind of dataset does not exist in the wild.
So we built our own synthetic data pipeline from scratch.
Hundreds of hours of MoCap captured with hundreds of actors.
A proprietary mesh-aware retargeting system to transfer motion across radically different body types without breaking physics.
Millions of perfectly annotated video pairs, generated procedurally.
We wrote about how we built it 👇
https://lnkd.in/eNXS3qkD
🎥 Kinetix Kamo-1: when motion control stops being a promise
🧩 A clear premise first
Kamo-1 is still in beta, but Kinetix is one of those companies that don’t sell dreams and then disappear.
I tested this model months ago, back in the alpha phase, and one thing matters: most of the feedback I shared back then has actually been implemented. That’s not something you see often in the AI space.
🧠 This is not just about motion control
The point isn’t simply animating a character.
The real leap is in how character movement and camera movement work together, and this is where Kamo-1 clearly separates itself from many current AI video models.
🧱 The core is 3D, not 2D pretending to be smart
Unlike most models that try to mimic pixels from a reference video, Kamo-1 reconstructs the scene in 3D before generating the final output.
That means you’re not “pasting” motion anymore. You’re directing a scene in space, much closer to real filmmaking logic than prompt-based tricks.
📷 Camera control is the real game changer
Motion control alone is becoming common.
But having coordinated camera movement (pan, tilt, orbit, dolly) that stays coherent with body motion, action, and space is something else entirely. From the tests, it’s clear these are not cosmetic presets, they actually hold up.
🔁 The jump from alpha is very real
More stable camera behavior, fewer breaks in orbit and zoom, better spatial continuity, and movements that feel less “AI-ish”.
Not perfect, but significantly stronger than earlier versions.
🎯 Is it ready for everything? No. Is it just hype? No. It’s real control
Kamo-1 is not “final cinema”, but it’s already usable in productions that need precise motion and camera movement that truly supports the action, exactly where many AI video tools still fail.
🎞️ My personal takeaway
This doesn’t replace filmmakers.
It’s a tool that finally starts thinking like a director, and right now, that’s a very real step forward.
🔔 Follow me for the last AI updates!
#AInews#Kinetix#AIVideo#Filmmaking#CreativeAI
Motion control is great. But it’s not enough.
Most AI video models operating right now are 2D-based. They look at a reference video and try to mimic the pixels.
But I’ve been testing the new Kinetix Open Beta (the Kamo model), and it’s doing something different.
It’s 3D-conditioned.
What does that actually mean you?
It means we aren’t just pasting a movement onto a character. We are directing a scene in a 3D space.
This is currently the only model that gives you two superpowers at the same time:
->Character Motion Control (The acting)
->Camera Control (The directing)
In cinema, you need to show the same scene from different angles, an establishing shot, a close-up, an orbit. 2D models struggle with this.
Kinetix is solving it.
I played around with the new camera presets to see if I could break the orbit and zoom functions.
The Open Beta is live now.
Have you tried Kinetix yet?
Daily Kamo-1 Clip
Today’s clip demonstrates how Kamo-1 uses a simple acting input, one reference frame, and a selected camera movement to produce a controlled, coherent shot.
Controllable generation follows a simple structure:
• your image builds the scene
• your acting video drives the motion
• your camera preset sets the shot
We’ll be sharing more this week. Let us know what we should test next!
This AI model is now in Open Beta.
As a launch partner of Kinetix, I had early access
and have conducted tests since the alpha phase.
Here are the results from my own experiments.
Kamo-1 is a 3D-conditioned AI video model.
Unlike a typical video-to-video model, it transforms
your input video into a 3D file in the background and
generates a new video based on the input image,
prompt, and selected camera motion.
Step 1: Upload or select a short video
Step 2: Upload an image as the first frame
Step 3: Write a prompt & select a camera motion
Step 4: Generate the new video
Compared to the alpha version I tested earlier,
Kamo-1 is a significant improvement.
The camera motion feature makes it unique,
as it makes you feel like both a director
and a hands-on camera operator.
You can "reshoot" a video, take it from another angle,
apply a new motion without traveling back in time.
Keen to try it? Link in comments.
Daily Kamo-1 Clip
Today’s example shows how Kamo-1 takes a basic acting video, one reference image, and a chosen camera preset, and combines them into a fully controlled generation.
This is the core of controllable generation:
• the scene → shaped by your image
• the motion → guided by your performance
• the camera → chosen from 30+ presets
More clips on the way this week. Tell us what you’d like to see next!
Daily Kamo-1 Clip
Today’s clip shows how Kamo-1 takes a simple acting video, a single reference frame, and a chosen camera preset, and turns them into a fully controlled generation.
What you’re seeing here is the essence of controllable generation:
• the scene → built from your image
• the motion → driven by your performance
• the camera → selected from 30+ options
More clips coming this week. Tell us what you want to see next!
Daily Kamo-1 Clip
Today’s clip shows how Kamo-1 takes a simple acting video, a single reference frame, and a chosen camera preset, and turns them into a fully controlled generation.
What you’re seeing here is the essence of controllable generation:
• the scene → built from your image
• the motion → driven by your performance
• the camera → selected from 30+ options
More clips coming this week. Tell us what you want to see next!
