AI Paper Atlas

Research map / Video / Video generation

LayerT2V: A Unified Multi-Layer Video Generation Framework

Guangzhao Li, Kangrui Cen, Baixuan Zhao, Yi Xin, Siqi Luo, Guangtao Zhai, Lei Zhang, Xiaohong Liu

ICML 2026 · Video › Video generation · approach: Compositional text-to-video diffusion

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 LayerAdaLN to vanilla temporal concatenation (ablation)LayerT2V · VidLayer held-out, FG layersubject consistency: 0.931 -> 0.973Improved
add LayeredCrossAttention on top of LayerAdaLN (ablation)LayerT2V · VidLayer held-out, full-video layertext alignment: 0.195 -> 0.214Improved
add LayeredCrossAttention on top of LayerAdaLN (ablation)LayerT2V · VidLayer held-out, FG layersubject consistency: 0.973 -> 0.983Improved
fine-tune Wan VAE decoder with LoRA instead of training a native Mask VAE from scratch (ablation)LayerT2V · VidLayer held-out, FG layeraesthetic quality: 0.4432 -> 0.4971Improved
use VAE LoRA instead of native Mask VAE (ablation)LayerT2V · VidLayer held-out, BG layertext alignment: 0.2102 -> 0.2307Improved
use VAE LoRA instead of native Mask VAE (ablation)LayerT2V · VidLayer held-out, BG layertemporal flickering: 0.9837 vs 0.9830No clear effect
extend 3D RoPE to 4D with a layer axis (ablation)LayerT2V · qualitativevisual qualityHurt
replace LayerFlow with LayerT2VLayerFlow · VidLayer held-out, FG layeraesthetic quality: 0.4591 -> 0.4971Improved
replace LayerFlow with LayerT2VLayerFlow · VidLayer held-out, BG layersubject consistency: 0.9682 -> 0.9838Improved
replace LayerFlow with LayerT2VLayerFlow · User studypreference for aesthetic quality: 0.120 -> 0.724Improved
use native Mask VAE instead of LayerFlowLayerT2V (Native Mask VAE) · VidLayer held-out, full-video layermotion smoothness: 0.9788 -> 0.9776Hurt

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