Publications

A diffusion-based autoregressive framework that anticipates future keyframes to generate temporally coherent, step-by-step instructional videos in a streaming fashion.

The first dynamic data-free quantization: synthesizes plausibly difficult images with soft labels so a network learns to adapt its bit-width to each input.

Quantization-aware training that reduces feature distribution mismatch in SR networks through gradient-aligned regularization, without test-time dynamic adaptation.

On-the-fly adaptive bit allocation for SR, calibrated with only a few images, cutting quantization processing time from hours to seconds (×2400 faster).

Face inpainting that adaptively blends context from the visible image with learned latent priors, guided by attention between missing and visible regions.

Allocates bit-widths per image patch and per layer based on local content, enabling SR quantization below 8 bits with little accuracy loss.

Learns a layer-wise pruning ratio for N:M structured sparsity, improving the efficiency–accuracy trade-off for super-resolution and deblurring.

Channel-wise, distribution-aware quantization that brings SR networks to ultra-low precision without a significant drop in restoration quality.

Data-free filter pruning that scores each filter's importance from the batch-normalization parameters of a pre-trained CNN.