I’m the Co-Head of Visual Research at Stability AI with research interests in the intersection of machine learning and computer graphics
Germany
Joined July 2009
- Tweets322
- Following264
- Followers1.5K
- Likes242
Mark Boss retweeted
What if making video-based world models smaller isn't just about better compression, but about making their representations easier to predict?
Recent approaches to video generation explore building scenes from coarse to fine. The first few tokens capture the big picture, like a person playing a guitar, while later tokens progressively add visual details.
Our Interactive Research team just published SemanTok, which takes this idea further. By making those early tokens more semantically meaningful, we give the model a clearer understanding of what's happening in a scene, making the representation easier to predict and video generation more efficient.
The result: a model using SemanTok matches or beats the performance of a model more than three times its size.
Read the full paper: stability.ai/research/semant…
At the start of the project we found that flexible token schemes break in low token counts. This started this project and we discovered that semantic alignment of tokens is essential to achieve true flexible tokenization. Amazing project Mike!
Excited to share SemanTok 🎬, from my summer project at @StabilityAI!
Predictable semantic tokens for autoregressive video generation: our 49M AR model beats SOTA VideoFlexTok up to 47× its size in semantics, and at 201M it wins a 3.4× larger one in fidelity
Absolutely amazing work by Wei. Check out the website. It has super cool examples on what is possible with the technique. I really like that it constantly create plausible motion and even creates snowboard trails.
Being a director is hard. 🎬
📷 Move the camera.
🧍 Stage the subjects.
⏱️ Get the timing right.
✨ With 4Director, you place the camera and every object in 3D and draw their paths. The model generates the shot.
🌐 stability-ai.github.io/4dire…
Mark Boss retweeted
Being a director is hard. 🎬
📷 Move the camera.
🧍 Stage the subjects.
⏱️ Get the timing right.
✨ With 4Director, you place the camera and every object in 3D and draw their paths. The model generates the shot.
🌐 stability-ai.github.io/4dire…
Mark Boss retweeted
ICLR 2027 has received more submissions than all previous years (2013–2026) combined
Mark Boss retweeted
Color consistency is a persistent friction point in production. Our interactive research team just returned from The 19th European Conference on Computer Vision (@eccvconf), where they presented new work focused on solving just this: ensuring colors match reliably across shots and AI-generated reference photos.
We’re also hiring beyond 3D, video, and world models:
🎨 Researcher, Creative Workflows — combine ideas across research areas to create new creative tools.
stability.ai/careers?gh_jid=…
🤝 Forward Deployed Engineer — help bring our research to partners.
stability.ai/careers?gh_jid=…
Mark Boss retweeted
Today marks a new milestone for Stability AI. We've closed our Series B, bringing total funding to $232M under new leadership. This moment is more than the number, it's about who's in the room.
Mark Boss retweeted
I used Fable 5 to port Command & Conquer: Generals Zero Hour to the iPhone and iPad!
This is the actual 2003 engine compiled for ARM64 natively, no emulator.
Campaign, skirmish, Generals Challenge all work with touch controls built for an RTS.
Open sourcing it all below!
Mark Boss retweeted
Arbor: Control 3D Generation with Explicit Geometry
Text-to-3D makes plausible assets, but spatial control is still hard. Arbor adds typed 3D constraint: hull, avoidance, and touch.
Paper w/ @markb_boss @AjlEngelhardt @Simondonne @CG_Tuebingen Intern @StabilityAI
More 👇
Mark Boss retweeted
We're excited to share Stable-Layers!
We train Qwen-Image-Layered further with RL for improved layerization,
using only feedback from a VLM — no paired supervision required!
Paper: arxiv.org/abs/2605.30257
Project Page: stability-ai.github.io/stabl…
We released OCTOPUS 🐙
KV-cache compression using octahedral parametrization: rotate keys, group them into 3D triplets, encode each direction, and decode directly inside attention.
Paper: arxiv.org/abs/2605.21226
Project Page: octopus-quant.github.io/
We initially cared about local LLMs, but KV caches appear in more than text.
So we also investigated OCTOPUS for autoregressive video and audio transformers.
Joined work with @VikramVoleti, Simon Donné, @esx2ve
