Publications
This page highlights all 14 publications in reverse chronological order. Each entry links to its detail page and external resources when available. You can also browse citations on Google Scholar.
Conference (full paper)
Recovering Diversity in Distilled Diffusion Models via Low-Rank Policy-Optimized Merging
Recovering diversity in few-step diffusion models through low-rank, policy-optimized teacher-student merging.
Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs
Transferring alignment from strong vision models to audio and video MLLMs without further fine-tuning.
Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems: Optimizing Guidance and Stochasticity Schedules
Coordinating guidance and stochasticity schedules for diffusion-based inverse problems.
Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers
Region-adaptive latent upsampling accelerates diffusion transformers with minimal quality degradation.
Perceptually Consistent Low-Resolution Previews with High-Resolution Image for Efficient Workflows of Diffusion Models
Training-free, low-resolution previews that stay perceptually consistent with high-resolution outputs.
Erasing Thousands of Concepts: Towards Scalable and Practical Concept Erasure for Text-to-Image Diffusion Model
Scalable concept erasure that removes thousands of concepts while preserving generation quality.
On Epistemic Uncertainty of Visual Tokens for Object Hallucinations in Large Vision-Language Models
Identifying and masking uncertain visual tokens to reduce object hallucinations without training.
Skrr: Skip and Re-use Text Encoder Layers for Memory Efficient Text-to-Image Generation
Skipping and reusing text encoder layers for memory-efficient text-to-image generation.
Efficient Personalization of Quantized Diffusion Model without Backpropagation
Personalizing quantized diffusion models without backpropagation.
INTRA: Interaction Relationship-aware Weakly Supervised Affordance Grounding
Learning interaction-aware affordance grounding from weak supervision.
BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion
Generating detailed, human-centric scenes beyond 8K with pretrained diffusion models.
Conference (short paper, abstracts or workshop papers)
Harmonization for a black-box deep learning model
Harmonizing MR images for black-box models without access to model parameters.
Technical Report
On Geometrical Properties of Text Token Embeddings for Strong Semantic Binding in Text-to-Image Generation
Using text embedding geometry to improve semantic binding in text-to-image generation.
DITTO-NeRF: Diffusion-based Iterative Text To Omni-directional 3D Model
Reconstructing diverse 3D objects from text or a single image with 2D diffusion priors.