AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

GeoRK2: Geometry-Guided Runge-Kutta Integration for Diffusion Transformer Acceleration

Chaoqun Sun, Zongjing Fu, Powei Chang, Jinpeng Zhang, Jianxiang Xiang, Yukang Gao, Chenyu Wang

CVPR 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
remove metric-preconditioned geometric correction (use identity metric) (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 3.02 -> 2.31Improved
use second-order RK2 instead of Euler integration (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 2.87 -> 2.31Improved
add geometry-aware components over Euclidean RK2 baseline (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 3.41 -> 2.31Improved
add temporal averaging (EMA) of the metric (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 3.45 -> 2.31Improved
increase truncation rank r from 32 to 64 (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 2.89 -> 2.31Improved
increase truncation rank r from 64 to 128 (ablation)GeoRK2 · ImageNet-256 DiT-XL/2FID: 2.31 -> 2.28No clear effect
damping rho outside [0.6, 0.9] and correction step lambda outside [0.05, 0.15] (ablation)GeoRK2 · ImageNet-256 DiT-XL/2stability / FIDHurt
use GeoRK2 instead of TaylorSeer at matched ~2.7x speedupfeature-cache acceleration · ImageNet-256 DiT-XL/2FID: 2.73 -> 2.67Improved
use GeoRK2 instead of TaylorSeer at ~4.5-4.9x speedupfeature-cache acceleration · ImageNet-256 DiT-XL/2FID: 4.31 -> 3.32Improved
use GeoRK2 instead of TeaCache/DBCache for text-to-imagecaching acceleration · FLUX.1-dev DrawBenchImageReward: 0.8784 / 0.6162 -> 0.9889Improved
use GeoRK2 (N = 8) instead of TaylorSeer (N = 7) at ~4.4x speedupfeature-cache acceleration · FLUX.1-dev DrawBenchCLIP score: 33.798 -> 33.476Hurt
use GeoRK2 instead of TaylorSeer for text-to-videofeature-cache acceleration · HunyuanVideo VBenchVBench score (%): 79.99 -> 80.73Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy