DriftWorld world model achieves 30 FPS with low-resource training
It cuts training costs sevenfold compared to prior methods.
63102822729.7K
Original post unavailable.
Sentiment
Positive users celebrate DriftWorld's real-time world modeling on 1-2 GPUs for making advanced AI feasible outside big labs, while negative users worry collapsed diversity could limit post-training usefulness.
Pos
100.0%
Neg
0.0%
4 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
LIKES1REPLIES1
@du_yilun Very cool ! I like the idea of using just a single positive sample since as the dynamics should be (near-)deterministic. Totally make sence for a WM
RETWEETS27
Excited to share DriftWorld -- a world model built on drifting! https://susie-lu.github.io/driftworld/ Existing world models are both slow and very expensive to train. DriftWorld is much more efficient -- natively runs at 30+ FPS, and also only requires a 1-2 GPU to train!
Excited to share our work on DriftWorld, a fast 1-step action-conditioned world model! •A single forward pass to generate future frames from current frame + actions •30+ fps real-time generation •Efficient inference-time action search & policy evaluation •Cheap to train (1/7)
Have replicated on my private dataset. It’s amazing I can train a i2v model(without action) with such low resources from scratch!
Excited to share DriftWorld -- a world model built on drifting! https://susie-lu.github.io/driftworld/ Existing world models are both slow and very expensive to train. DriftWorld is much more efficient -- natively runs at 30+ FPS, and also only requires a 1-2 GPU to train!
@du_yilun I was expecting another kind of drifting
@du_yilun However I suppose it mean no post-training possible on the WM since the diversity will be totally collapse. Maybe not great in practice
@du_yilun Training on 1-2 GPUs is the real headline. World models stop being a big-lab monoculture the moment a grad student can train one. My bet: the most interesting interactive-WM ideas of 2027 come from labs that couldn't afford Genie-scale compute.
@du_yilun Amazing work!