AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

Guiding a Diffusion Transformer with the Internal Dynamics of Itself

Xingyu Zhou, Qifan Li, Xiaobin Hu, Hai Chen, Shuhang Gu

CVPR 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 7 citations (OpenAlex)

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 auxiliary intermediate supervision at block 4 (ablation)SiT-B/2 · ImageNet 256x256 (80 ep)FID: 33.02 -> 30.60Improved
add IG sampling (w=1.5) on top of aux supervision at block 4 (ablation)SiT-B/2 · ImageNet 256x256 (80 ep)FID: 30.60 -> 19.02Improved
place auxiliary supervision in latter half of network (ablation)SiT-B/2 · ImageNet 256x256 (80 ep)FID: 33.02 -> 38.05Hurt
use two auxiliary supervision layers (2 and 6) (ablation)SiT-B/2 · ImageNet 256x256 (80 ep)FID: 33.02 -> 33.34No clear effect
auxiliary supervision instead of self-supervised regularization (SRA) (ablation)SiT-B/2 · ImageNet 256x256 (80 ep)FID: SRA 29.10 vs aux 30.45Mixed
add IG on top of CFG (ablation)SiT-B/2 · ImageNet 256x256FID: 7.30 -> 6.50Improved
use guidance interval [0.3,1) with IG w=2.3 (ablation)SiT-B/2 · ImageNet 256x256FID: 18.53 -> 16.19Improved
restrict IG to high-noise range [0,0.7) (ablation)SiT-B/2 · ImageNet 256x256FID: 18.53 -> 20.69Hurt
raise IG coefficient from 1.9 to 2.3 (no interval) (ablation)SiT-B/2 · ImageNet 256x256FID: 17.38 -> 18.53Hurt
reduce lambda from 1 to 0.25 or 0.5 (ablation)SiT-B/2 · ImageNet 256x256 (400K)FID: 31.24 -> 30.38No clear effect
add IG to SiT-XL/2 at 80 epochsSiT-XL/2 · ImageNet 256x256 (no CFG)FID: 8.61 (vanilla, 1400 ep) -> 5.31 (80 ep)Improved
add IG to LightningDiT-XL/1LightningDiT-XL/1 · ImageNet 256x256 (no CFG)FID: 5.14 -> 2.42Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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