Diffusion/flow matching is a go-to generative AI paradigm, but it's slow, sample inefficient and complicated. Do we really need all this complexity?
We design a small and simple one-step model, ROMS-IMLE, which attains an FID of 2.56 on ImageNet.
Joint w/ @researchirag
1/6
Was just as excited as everyone else to read this cool new paper, and it felt like a trip down memory lane!
XM seems to be the same as IMLE, and the motivation and insights are also similar. Having worked on IMLE for years, I'll provide some context for my followers ๐.
1/n
Diffusion and flow matching-based robot planners are slow and generate noisy and jerky trajectories.
Delighted to share our ICRA 2026 paper, which leverages IMLE to improve planning frequency 19-fold from 4.3 Hz to 83 Hz and reduces jerk by 38% relative to flow matching.
Joint work w/ Grayson Lee, Minh Bui, Shuzi Zhou, Yankai Li and Mo Chen.
(1/7)
Introducing WIMLE, a model-based RL method that substantially improves sample efficiency and asymptotic performance on hard tasks. Rather assuming a Gaussian world model, WIMLE trains a world model with IMLE.
Joint w/ @mehranag, @Moazeni_Alireza, @yszhang170.
See ๐ for links.
๐ฅณ 3DV2026 is almost there!
๐ฃ We are at ๐ฃ๐ผ๐๐๐ฒ๐ฟ ๐ฆ๐ฒ๐๐๐ถ๐ผ๐ป ๐ฎ ๐๐ผ๐บ๐ผ๐ฟ๐ฟ๐ผ๐ ๐ฎ๐ณ๐๐ฒ๐ฟ๐ป๐ผ๐ผ๐ป!
๐ค Come discuss the opportunities for applying interactive digital twins in robotics!
#3DV2026#AI#ComputerVision#DigitalTwin#Robotics
If you are at #ICCV2025, check out our work on interpolating between two states of a 3D scene with large motion at poster 269 on Tuesday afternoon. It proposes a general-purpose method that can disambiguate points with similar appearance.
Website: https://t.co/Kcx0JFEL3U
1/5
Are Diffusion and Flow Matching the best generative modelling algorithms for behaviour cloning in robotics?
โ Multimodality
โFast, Single-Step Inference
โSample Efficient
๐ก We introduce IMLE Policy, a novel behaviour cloning approach that can satisfy all the above.
๐งต๐
Diffusion models turn the data into a mixture of isotropic Gaussians, and so struggle to capture the underlying structure when trained on small datasets. In our new #ECCV2024 paper, we introduce RS-IMLE, a generative model that gets around this issue.
Website: https://t.co/eVg8oiKhAP
Code: https://t.co/Mf0WlNYBrK
Joint work w/ @researchirag and @PengShichong
If you are at #ECCV2024, come and check out poster 279 on Thursday afternoon from 4:30pm-6:30pm.
(1/6) Thread ๐