JinLong Yang

Bio

I am a fourth-year undergraduate student in Software Engineering at Sichuan University. My research focuses on efficient generation with diffusion models and long video generation.

I am particularly interested in understanding and accelerating the feature dynamics of Diffusion Transformers, with the goal of making high-quality image and video generation substantially more efficient.

Email GitHub

Publications

BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast DiT Inference

JinLong Yang et al.

ACM Multimedia (ACM MM), 2026

A training-free acceleration framework that combines local trajectory modeling, barycentric rational extrapolation, and adaptive Chebyshev weighting for efficient DiT inference.

Project page

FADE: Fractional Anomalous Dynamics Extrapolation for Diffusion Transformers Acceleration

JinLong Yang et al.

NeurIPS 2026

Models DiT latent trajectories with fractional dynamics and uses a lightweight lookup strategy for adaptive, zero-overhead extrapolation.

Under Review

LaBACK: Laurent-Guided Backtracking for Training-Free Flow Matching Acceleration Under review

JinLong Yang et al.

A training-free framework that contracts cache-driven updates and applies Laurent-guided forecasting with negative-power correction, delivering consistent speed–quality gains for text-to-image, distilled, and text-to-video generation without extra model evaluations.

HDS: Heterogeneity-Driven Safeguarding for Structure-Aware DiT Acceleration Under review

JinLong Yang et al.

Profiles inter-layer sensitivity online to safeguard vulnerable layers while accelerating robust ones in Diffusion Transformers.

GRAD-RIR: Topology-Aware Graph-Conditioned Diffusion Model for Sparse RIR Prediction Under review

JinLong Yang et al.

A graph-conditioned diffusion approach for reconstructing room impulse responses from sparse measurements under geometric occlusion and strong spatial variation.