Research map / 3D vision / 3D generation

DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation

NeurIPS 2026 · 3D vision › 3D generation · approach: Text-to-3D scene generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DecompDreamer, a two-stage composition-aware SDS optimization curriculum (joint relational modeling then targeted object refinement) over VLM-generated scene graphs and 3D Gaussians for compositional text-to-3D generation.

Problem

Text-to-3D methods generate single objects well but fail on compositional prompts with multiple objects and relationships, where simultaneous or iterative optimization leads to entangled geometry or divergence.

Approach

Reframes compositional generation as an optimization scheduling problem: first optimize inter-object relational losses to build a structural scaffold, then switch to targeted and object-level refinement with view-aware prompts and negative prompts, using a staged curriculum instead of simultaneous or iterative losses.

Key results

• User study (20 participants, 30 prompts): DecompDreamer preferred in 61.0% of cases overall and 73.0% for prompts with more than three objects.
• Text-to-3D Alignment score of 73.4% overall (86.6% for >3 objects), versus 63.7% for the next best (GALA3D) overall.
• Table reports CLIP Score 34.5% and Pick-A-Pic 24.0% overall, the highest among compared methods, with a typical gap of about 5%.

Surprising findings

• Automated metrics (CLIP Score, Pick-a-Pic) show a narrower gap than expected, which the authors attribute to metrics ignoring view consistency such as the Janus problem.
• GraphDreamer's Text-to-3D alignment falls from 50.0% on simple prompts to 16.5% on prompts with more than three objects.

Limitations

• Janus problem persists for uncommon viewpoints (e.g., piglet playing piano).
• Struggles with complex actions and decomposing a single entity into sub-parts such as human body parts.
• Relies on VLM (ChatGPT-4o) for scene graph, layout, and azimuth estimation; runtime scales with number of objects (495 min for 11 objects).

Experimental setup

Benchmarks: 30 compositional text prompts (2-11 objects), in-house

Metrics: CLIP Score Pick-A-Pic Text-to-3D Alignment user study preference

Baselines: GALA3D GraphDreamer MVDream LucidDreamer Magic3D Trellis HiFA

Builds on: DreamFusion Gaussian Splatting LucidDreamer GALA3D GraphDreamer

Closest papers by meaning

• Layout-your-3D: Controllable and Precise 3D Generation with 2D Blueprint (ICLR 2025 · similarity 0.88)
• CompGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians (CVPR 2025 · similarity 0.87)
• PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
• Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.86)
• MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation (NeurIPS 2026 · similarity 0.86)
• MIDGArD: Modular Interpretable Diffusion over Graphs for Articulated Designs (NeurIPS 2024 · similarity 0.85)