AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

XYZFlow: Scaling Multidimensional Shortcut Flows for Efficient Generative Modeling

Jinxiu Liu, Xuanming Liu, Kangfu Mei, Yandong Wen, Weiyang Liu

ICML 2026 · Image generation › Fast diffusion sampling · approach: Rectified flow straightening

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
add full-history (inter-patch trajectory) conditioning (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 3.51 -> 2.02Improved
add local-history (intra-patch temporal) conditioning (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 3.03 -> 2.25Improved
add shortcut (progressive step) schedule (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID and total steps: FID 2.05 -> 2.02; steps 20 -> 14No clear effect
add GAN discriminator loss (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 2.02 -> 1.63Improved
add GAN discriminator loss (ablation)XYZFlow-L · ImageNet 256x256 class-conditionalFID: 1.79 -> 1.25Improved
use 8x8 cell size instead of 1x1 tokens (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 2.85 -> 2.02Improved
treat whole image as one cell (16x16) instead of 8x8 (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 2.02 -> 2.55Hurt
progressive 5->4->3->2 instead of constant 5 steps (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID and total steps: FID 1.63 -> 1.63; steps 20 -> 14No clear effect
aggressive 5->2->2->2 schedule instead of 5->4->3->2 (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalrecall: 0.62 -> 0.58Hurt
8->8->8->8 steps instead of 5->4->3->2 (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 1.63 -> 1.61; steps 14 -> 32No clear effect
use weaker DiT/XL-2 teacher instead of xAR (ablation)XYZFlow-B · ImageNet 256x256 class-conditionalFID: 1.63 -> 1.74 (GAN)Mixed
use stronger xAR teacher (xAR-B to xAR-H) (ablation)XYZFlow-L · ImageNet 256x256 class-conditionalFID (w/ GAN): 1.32 -> 1.25No clear effect

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