Made a crappy portal game using Waypoint-1.5-360p from @overworld_ai on my old 6GB RTX3060 as an experiment, the model performed a lot smoother than I expected, super cool stuff! Finally got the chance to execute this idea since this is the only model that’s consumer friendly.
Hanging out in tokyo for the next two weeks.
If vou're a researcher here working on 3D, robotics or World models, would love to meet up and discuss! Just drop me a DM.
While many frontier world models require large-scale GPU infrastructure, Waypoint-1.5 is the only world model designed to run on consumer devices.
The paper includes benchmarks across a wide range of RTX GPUs, demonstrating real-time performance on hardware people already own.
Our CEO, @lcastricato, recently sat down with the @AP to discuss how Overworld came together, where world models are headed, and what we're learning as this category continues to take shape.
Take a look: https://t.co/qv0tOOg1j9
People love to complain about india not having any frontier AI labs, but this is how we treat the CVPR Best paper finalist authors. 🤡
The system is genuinely fucked and I have zero hopes for the future of AI in india, R&D is given no value over here, the system is just stupid.
@arnie_hacker Wake up at 3 am thinking, I'll just make sure the job didn't fail and go back to sleep
Which spirals into "Oh it did fail, but I can just fix and re-run"
Turns into an hour long debugging session
Cool stuff, I mostly see their focus on 3DGS work, which is certainly impactful.
But 3DGS worlds are painfully static, sure it's stable.
But the beauty of AR diffusion world models is in their controllability, editability, and dynamics. Yes, it's very hard to get it right, but once we do, it's going to be revolutionary for 3D understanding and simulation
Can't stop thinking about how close we are to fully generated real-time worlds. People have no idea at how much progress diffusion world models have already made.
The next 12 months are going to be absolutely wild
An idea that I'm really passionate about is actually inspired from the LLM OS by karpathy.
Treating diffusion world models as not just frame predictors, but as complete graphics engines.
With different components handling the physics/collision+ geometry
Lighting and visuals decoupled into a different world model.
Along with a memory component.
As well as LoRAs, controllable NPC/Agents.
Have been so bullish on this!
RL post training on egocentric rollouts.
Would be so interesting to treat the World model itself as the policy and use a GRPO-like optimization, this could be the way forward for autonomous robotics/NPCs in the rollouts
I recorded a playthrough to get a first impression of Overworld's new world model and it's honestly really amazing, first time I've genuinely started having a bit of fun in one of these things
Thrilled to announce Waypoint-1.5, the next step towards diffusion world models running locally. Working on this release was so much fun.
onto the next one now :)