AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

BézierFlow: Learning Bézier Stochastic Interpolant Schedulers for Few-Step Generation

Yunhong Min, Juil Koo, Seungwoo Yoo, Minhyuk Sung

ICLR 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 1 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
train Bespoke with the BezierFlow global LPIPS loss (ablation)Bespoke Solver · CIFAR-10 ReFlow RK1 NFE=4FID: 45.31 -> 38.34Improved
replace Bespoke discrete parameterization with Bezier (same loss) (ablation)Bespoke* · CIFAR-10 ReFlow RK1 NFE=4FID: 38.34 -> 20.64Improved
jointly optimize LD3 timesteps with the Bezier scheduler (ablation)BezierFlow · CIFAR-10 ReFlow RK2 NFE=10FID: 3.74 -> 3.17Improved
jointly optimize LD3 timesteps with the Bezier scheduler (ablation)BezierFlow · ImageNet FlowDCN RK2 NFE=10FID: 7.56 -> 10.88Hurt
jointly optimize LD3 timesteps with the Bezier scheduler (ablation)BezierFlow · CIFAR-10 EDM UniPC NFE=10FID: 2.09 -> 2.71Hurt
increase Bezier control points from 4 to 32 (ablation)BezierFlow · CIFAR-10FID: n=4 to n=32Improved
increase Bezier degree from 16 to 32 (ablation)BezierFlow · CIFAR-10FIDNo clear effect
apply BezierFlow trained at NFE=10 to unseen NFE=8 (vs LD3 trained at NFE=8)BezierFlow · CIFAR-10 ReFlow RK1 NFE=8FID: 12.54 -> 9.02Improved
BezierFlow instead of LD3 at NFE=4LD3 · ImageNet FlowDCN RK1 NFE=4FID: 11.62 -> 15.60Hurt
BezierFlow instead of LD3 on FFHQ at NFE=10 (UniPC)LD3 · FFHQ EDM UniPC NFE=10FID: 2.92 -> 3.13Hurt
BezierFlow instead of LD3 on FFHQ at NFE=4 (UniPC)LD3 · FFHQ EDM UniPC NFE=4FID: 22.48 -> 17.05Improved
BezierFlow (15 min training, NFE=10) vs consistency distillation (8 days, NFE=2)Consistency Distillation · CIFAR-10 EDMFID: 2.93 -> 2.09Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 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