Associate Professor and Canada CIFAR AI Chair @SFU @AmiiThinks. Ph.D. from @Berkeley_EECS and Bachelor's from @UofTCompSci. Formerly @GoogleAI and @the_IAS.
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
@bayesiankalki@kianhk3 Yes, @kianhk3's talk was great!
I discussed the relationship between IMLE and explorative modelling last month in a Twitter thread.
VLA policies are slow and jerky. Introducing IMLE-VLA, which swaps out flow matching for conditional IMLE. IMLE-VLA generates action chunks at 55 Hz (3.67x faster than π0.5), produces >2.2x less jerk and finishes tasks >3.9x faster.
Details at https://t.co/axlJ1JcaWg
If you are at #IROS2026, don't miss @kianhk3's talk on Monday at 9am in Room 329.
Joint work w/ @kianhk3, Qinhe Peng, @hekoshii, @mehranag, @jianing_qian, @tristanengst, @Moazeni_Alireza and @dineshjayaraman
Thread 👇
1/8
Ah yes, we focused mostly on speed in the paper, but in retrospect it would have been a good idea to explain the non-invertibility point a bit more.
IMLE can map multiple noise vectors to the same action chunk, so it can learn a non-invertible map. I was referring to both - the action chunk generated by IMLE is itself less noisy than that generated by flow matching, and because it's less noisy, the variance across action chunks is lower.
Website: https://t.co/KPiGQybArL
Paper: https://t.co/9F1NrUQ8vn
If you are at #IROS2026, check out @kianhk3's talk on Monday at 9am in Room 329 and/or poster 4624 on Monday at 11am in Pod D8 of Exhibit Hall B.
Add to your #IROS2026 schedule: https://t.co/PCfYSqS0yz
8/8
Conditional IMLE (cIMLE) handles multimodal distributions by giving the model the opportunity to generate samples from multiple modes and avoiding penalizing samples from the modes that are different from the mode corresponding to the current observation.
7/8
Editing appearance properties (e.g., colour, shading, albedo, etc.) of 3DGS scenes can yield unexpected results - after editing, the colour may be very off.
We found it was caused by misattribution of ray colour to primitives, a side effect of volume rendering.
PAPR gets around this by attending to all nearby points, which encourages points to move to the surface.
Joint work w/ @Moazeni_Alireza, @PengShichong, @yszhang170 and @researchirag
If you are attending #ECCV2026, stop by poster 187 this morning to learn more!
Project Website: https://t.co/J9fviDl2Kg
Paper: https://t.co/HMlTHMYc1Q
Code: https://t.co/KnZimT2eIq
Video: https://t.co/ZL2lzYaFax
Excited to share another work at ECCV 2026!
I’ll be presenting PointGT at #ECCV2026 tomorrow — on Thursday, Sep 10 at 10:30 AM – 12:30 AM poster # 206! Come by if you’re interested in texture or geometry editing in 3D. Would love to chat and hear your thoughts!