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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| remove metric-preconditioned geometric correction (use identity metric) (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 3.02 -> 2.31 | Improved |
| use second-order RK2 instead of Euler integration (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 2.87 -> 2.31 | Improved |
| add geometry-aware components over Euclidean RK2 baseline (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 3.41 -> 2.31 | Improved |
| add temporal averaging (EMA) of the metric (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 3.45 -> 2.31 | Improved |
| increase truncation rank r from 32 to 64 (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 2.89 -> 2.31 | Improved |
| increase truncation rank r from 64 to 128 (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | FID: 2.31 -> 2.28 | No clear effect |
| damping rho outside [0.6, 0.9] and correction step lambda outside [0.05, 0.15] (ablation) | GeoRK2 · ImageNet-256 DiT-XL/2 | stability / FID | Hurt |
| use GeoRK2 instead of TaylorSeer at matched ~2.7x speedup | feature-cache acceleration · ImageNet-256 DiT-XL/2 | FID: 2.73 -> 2.67 | Improved |
| use GeoRK2 instead of TaylorSeer at ~4.5-4.9x speedup | feature-cache acceleration · ImageNet-256 DiT-XL/2 | FID: 4.31 -> 3.32 | Improved |
| use GeoRK2 instead of TeaCache/DBCache for text-to-image | caching acceleration · FLUX.1-dev DrawBench | ImageReward: 0.8784 / 0.6162 -> 0.9889 | Improved |
| use GeoRK2 (N = 8) instead of TaylorSeer (N = 7) at ~4.4x speedup | feature-cache acceleration · FLUX.1-dev DrawBench | CLIP score: 33.798 -> 33.476 | Hurt |
| use GeoRK2 instead of TaylorSeer for text-to-video | feature-cache acceleration · HunyuanVideo VBench | VBench score (%): 79.99 -> 80.73 | Improved |
Closest papers by meaning
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.88)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- GeoSPRINT: Geometric Redundancy-Aware Step Pruning for Inference in Diffusion Trajectories (NeurIPS 2026 · similarity 0.87)
- GeoAda: Efficiently Finetune Geometric Diffusion Models with Equivariant Adapters (NeurIPS 2025 · similarity 0.87)
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.
