Infoblock alpha
cs-cv/vision IB-arxiv-2607.28611-dense-backbone-compute-efficiency measured

Measured by pretraining diffusion loss, Chimera's dense hybrid backbone was 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

compute_efficiency_multiple 1.7 · baseline_params_b 2

01

Primary source

  1. arxiv · 2607.28611 · v1 · abstract
Title
Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers
Authors
Chongjian Ge, Hanwen Jiang, Tianyu Wang, Jiuxiang Gu, Yiran Xu, Ziwen Chen, Shaoteng Liu, Jing Shi, Yicong Hong, Zefan Cai, Hailin Jin, Hao Tan
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.