Presentation
DiffPro: Joint Timestep and Layer-Wise Precision Optimization for Efficient Diffusion Inference
DescriptionDiffPro is a simple framework to speed up and shrink diffusion models while preserving image quality. It combines layer-wise quantization, guided by a manifold-based sensitivity check with adaptive timestep selection. Compared with quantization-only or sampling-only baselines, this joint strategy yields better FID–memory trade-offs. We evaluate on MNIST, CIFAR-10, and CelebA using both PTQ and QAT, showing meaningful size reductions (e.g., ~34 MB on a CIFAR-10 setup) with competitive FID. Current limits include occasional mis-ranking of a few blocks at high noise and a focus on unconditioned models. Next, we will extend DiffPro to text/class-conditioned diffusion, replace hand-tuned thresholds with a budgeted optimizer that co-selects per-layer bit-widths and timesteps using per-timestep sensitivity, and incorporate hardware-aware costs.
Event Type
Research and ACM SRC Posters
TimeThursday, 20 November 20258:00am - 5:00pm CST
LocationSecond Floor Atrium
Archive
view