| Sep 25, 2026 |
Two papers are accepted to NeurIPS 2026: FlashControl, a one-step controllable generator distilled from single-stream teachers, and Deltafugue, on algorithmic length generalization.
|
| Jul 20, 2026 |
Three papers are accepted to ECCV 2026: VARIN (noise inversion for visual autoregressive editing), Cross-Space Distillation, and Spectral-Aware Analytic Class-Incremental Learning.
|
| Jun 1, 2026 |
Started a research internship at Adobe Research, working on perceiver transformers for pixel-space video diffusion.
|
| Feb 20, 2026 |
Two papers are accepted to CVPR 2026: MPDiT, a multi-patch global-to-local transformer for efficient diffusion and flow matching, and an optimal-transport approach to online incremental learning.
|
| Nov 1, 2025 |
DICE is accepted to WACV 2026. It is the first method for precise inversion and controllable editing in discrete diffusion and masked generative models.
|
| Sep 28, 2025 |
AutoEdit is accepted to NeurIPS 2025. It uses reinforcement learning to tune the hyperparameters of diffusion-based image editing automatically.
|
| Jun 1, 2025 |
Started a research internship at Apple Machine Learning Research, working on efficient diffusion training in the DCT frequency domain.
|
| Jan 22, 2025 |
Improved Training Technique for Latent Consistency Models is accepted to ICLR 2025. We identify impulsive outliers in latent space as the main cause of unstable consistency training, and address them with a Cauchy loss, OT coupling, and an adaptive robust scale scheduler.
|
| Dec 10, 2024 |
SCFlow is accepted to AAAI 2025. It distills flow-matching models into one-step and few-step generators that give consistent outputs for the same noise, whatever the number of sampling steps.
|
| Sep 23, 2024 |
DiMSUM is accepted to NeurIPS 2024. It is a hybrid Mamba–transformer architecture that uses spatial and frequency information for faster-converging, state-of-the-art diffusion and flow-matching models.
|
| Jul 21, 2024 |
RDUOT is accepted to ECCV 2024. It combines unbalanced optimal transport with diffusion to give the first diffusion framework robust to corrupted training data.
|
| Jul 13, 2023 |
Anti-DreamBooth is accepted to ICCV 2023. It protects users’ photos from malicious personalization by adding imperceptible adversarial perturbations that disrupt DreamBooth fine-tuning.
|
| Feb 26, 2023 |
Our paper WaveDiff is accepted to CVPR 2023. WaveDiff introduces a wavelet-based diffusion framework that brings diffusion models close to real-time sampling.
|