ECCV 2026 Project Page

DiTex4D: Direct Text-Driven 4D Generation with Structured Latent Diffusion

Xiaozhe Chen1,2, Mengqi Rong2, Jian Liu3, Shuhan Shen1,2
1School of Artificial Intelligence, University of Chinese Academy of Sciences, 2Institute of Automation, Chinese Academy of Sciences 3School of Space and Earth Sciences, Beihang University
Paper Code

Abstract

Despite recent progress, direct text-driven 4D object generation remains challenging yet highly desirable. In this paper, we introduce DiTex4D, a native text-to-4D generation framework that enables both text-driven 4D generation from scratch and 3D animation from static mesh.

Built upon large-scale pre-trained 3D generation models, our framework DiTex4D avoids intermediate text-to-video pipelines and costly per-object optimization. Specifically, (i) we achieve 4D spatiotemporal consistency via inflating 3D attention with mixed-4D RoPE and tailored correlated noise injection strategy. (ii) To enable 3D animation, we introduce a mask-based diffusion model conditioned on multi-view global context to maintain strict consistency with the initial frame. We further fine-tune the framework for 4D interpolation to synthesize high-frame-rate sequences with smoother motion.

Extensive experiments demonstrate that DiTex4D can achieve higher-quality, semantically align-ed, and spatiotemporally coherent 4D object generation, surpassing most existing state-of-the-art text-to-4D generation methods.

Method

Overview of the DiTex4D method
DiTex4D Overview. Our framework supports two pathways for 4D generation, conditioned on text and static mesh inputs, while utilizing masks for mesh control. The 4D structured latent diffusion model implements a coarse-to-fine two-stage generation of multi-frame SLat, with correlated noise injection for consistency and Decoupled Cross-Attention for multi-condition guidance. The final SLat is decoded into high-quality 3DGS sequences, which are interpolated as the 4D output.

More Generation Results

Text-to-4D Generation

Prompt: “A man in red clothes is running.”

Prompt: “Banana is dancing.”

Prompt: “A cartoon character is raising his hands.”

Prompt: “The whale is swimming.”

Prompt: “A man wearing a blue coat is walking.”

Prompt: “A boy wearing glasses is jogging.”

Prompt: “The dragon is soaring with wings beating.”

Prompt: “Robot is running.”

3D+Text-to-4D Generation

3D input example 1

Prompt: “The 3D airplane is rotating.”

3D input example 2

Prompt: “A man is squatting.”

3D input example 3

Prompt: “The 3D rabbit is walking forward.”

3D input example 4

Prompt: “The spider is crawling.”

Long-Horizon Generation

BibTeX

@article{chen2026ditex4d,
  author    = {Chen, Xiaozhe and Rong, Mengqi and Liu, Jian and Shen, Shuhan},
  title     = {DiTex4D: Direct Text-Driven 4D Generation with Structured Latent Diffusion},
  journal   = {ECCV},
  year      = {2026},
}