AI Paper Atlas

Research map / Video / Video generation

Diff4Splat: Repurposing Video Diffusion Models for Dynamic Scene Generation

Panwang Pan, Chenguo Lin, Chenxin Li, Jingjing Zhao, Yuchen Lin, Haopeng Li, Yunlong Lin, Kairun Wen, Yixuan Yuan, Yadong MU

CVPR 2026 · Video › Video generation · approach: Efficient video diffusion models

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 motion loss (ablation)Diff4Splat · single-image dynamic scene generation (160-sample benchmark)FVD: 351.382 -> 210.153Improved
add motion loss (ablation)Diff4Splat · single-image dynamic scene generation (160-sample benchmark)QA-Quality: 2.145 -> 2.813Improved
add motion loss (ablation)Diff4Splat · single-image dynamic scene generation (160-sample benchmark)Avg. Matches: 4821.56 -> 5114.22Improved
add deformation Gaussian field module (ablation)Diff4Splat · qualitativevisual qualityImproved
add static pre-training before dynamic training (progressive training) (ablation)Diff4Splat · qualitative, 100K iterationsvisual quality and training time: 21 vs 7 days for comparable baselineImproved
use explicit 3D Gaussian field instead of implicit video generationAC3D · pose fidelityavg RPE translation: 3.001 -> 0.012Improved
use explicit 3D Gaussian field instead of implicit video generationAC3D · pose fidelityavg RPE rotation: 0.810 -> 0.008Improved
use feed-forward Diff4Splat instead of AC3D + Moscatwo-stage optimization pipeline · single-image dynamic scene generation (160-sample benchmark)FVD: 235.961 -> 210.153Improved
use feed-forward Diff4Splat instead of AC3D + Moscatwo-stage optimization pipeline · single-image dynamic scene generation (160-sample benchmark)KVD: 2.012 -> 2.316Hurt
use feed-forward Diff4Splat instead of AC3D + Moscatwo-stage optimization pipeline · single-image dynamic scene generation (160-sample benchmark)QA-Quality: 2.842 -> 2.813Mixed

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