AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

You Only Look One Step: Accelerating Backpropagation in Diffusion Sampling With Gradient Shortcuts

Hongkun Dou, Zeyu Li, Xingyu Jiang, Hongjue Li, Lijun Yang, Wen Yao, Yue Deng

TPAMI 2025 · Image generation › Fast diffusion sampling · approach: Fast diffusion ODE solvers · 0 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
backpropagate one step (SDO) instead of full backpropagation (DOODL) for gradient computationpixel diffusion model · gradient of x_N, A100time (s): 29.96 -> 1.62 (-91.35%)Improved
backpropagate one step (SDO) instead of AdjointDPMpixel diffusion model · gradient of x_N, A100time (s): 18.73 -> 1.62Improved
backpropagate one step (SDO) instead of DOODLpixel diffusion model · gradient of x_N, A100memory (GB): 11.25 -> 6.12 (-34.68%)Improved
backpropagate one step (SDO) instead of DOODLlatent diffusion model · gradient of x_N, A100time (s): 27.20 -> 1.91 (-88.84%)Improved
backpropagate one step (SDO) instead of DOODLlatent diffusion model · gradient of x_N, A100memory (GB): 18.59 -> 10.33 (-34.95%)Improved
use SDO gradient instead of full backpropagation for adversarial initial noiseclassifier-guided ImageNet diffusion · adversarial sample generationattack success ratio (%): 51.61 -> 96.32Improved
use SDO gradient instead of DOODL for adversarial noiseclassifier-guided ImageNet diffusion · adversarial sample generationFID: 70.4 -> 58.2Improved
use SDO gradient instead of DOODL for text-guided manipulationpixel diffusion on CelebA-HQ/FFHQ · text-guided image manipulationCLIP score: 28.76 -> 32.18Improved
use SDO gradient instead of FlowGrad for text-guided manipulationpixel diffusion on CelebA-HQ/FFHQ · text-guided image manipulationLPIPS: 0.142 -> 0.134Improved
use SDO gradient instead of DOODL for style-guided generationlatent diffusion · style-guided generationstyle loss: 357.2 -> 268.1Improved
use SDO gradient instead of DOODL for style-guided generationlatent diffusion · style-guided generationCLIP score: 31.12 -> 30.78No clear effect
use SDO gradient instead of DOODL for aesthetic enhancementStable Diffusion v1.4 · aesthetic enhancementaesthetic score: 9.02 -> 9.13Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 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