Research map / Image generation / Text-to-image diffusion
Apply Hierarchical-Chain-of-Generation to Complex Attributes Text-to-3D Generation
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 HCoG, an automated LLM-driven pipeline that decomposes long complex-attribute text prompts into occlusion-ordered blocks and generates 3D Gaussian assets block by block using part segmentation, ControlNet/MVDream SDS, and Gaussian Extension with Label Elimination.
Problem
Text-to-3D generation struggles with long prompts containing many attributes, causing wrong attribute binding, and objects with occlusion between parts need a generation order that prior methods required users to supply manually.
Approach
Use an LLM to split the prompt into parts and order them in-n-out by occlusion, generate each block coarsely then refine attributes precisely per part via SAM-supervised Gaussian label optimization, and add new parts via Gaussian extension plus label elimination so no manual boxes or orders are needed.
Key results
- Ours achieves BLIP-VQA 0.7295 and CLIP-Score 31.998, beating Progressive3D (0.6553, 30.451) on the complex-attribute prompts.
- With Stable Diffusion v3 backend, HCoG reaches BLIP-VQA 0.8055 and CLIP-Score 33.189, the best in the table.
- Inverse generation order drops BLIP-VQA from 0.7295 to 0.5023 versus the proposed order (ablation on HCoG).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HCoG instead of Progressive3D* | Complex-attributes 3D generation · Complex-attributes 3D generation | BLIP-VQA: 0.6553 -> 0.7295 | Helped |
| Use HCoG instead of MVDream | Complex-attributes 3D generation · Complex-attributes 3D generation | CLIP-Score: 31.132 -> 31.998 | Helped |
| Use HCoG + SD3 instead of GaussianDreamer | Complex-attributes 3D generation · Complex-attributes 3D generation | BLIP-VQA: 0.4919 -> 0.8055 | Helped |
| Generate from more-occluded parts to less-occluded parts rather than in inverse order (ablation) | HCoG · Complex-attributes 3D generation | BLIP-VQA: 0.5023 -> 0.7295 | Helped |
| Generate from more-occluded parts to less-occluded parts rather than in random order (ablation) | HCoG · Complex-attributes 3D generation | BLIP-VQA: 0.5961 -> 0.7295 | Helped |
| Generate from more-occluded parts to less-occluded parts rather than in inverse order (ablation) | HCoG + SD3 · Complex-attributes 3D generation | BLIP-VQA: 0.6363 -> 0.8055 | Helped |
| Generate from more-occluded parts to less-occluded parts rather than in random order (ablation) | HCoG + SD3 · Complex-attributes 3D generation | BLIP-VQA: 0.7033 -> 0.8055 | Helped |
| Add ControlNet shape guidance during fine-grained part optimization (ablation) | HCoG · Part optimization | shape consistency | Helped |
| Add Label Elimination after Gaussian Extension to remove redundant kernels (ablation) | HCoG · Generating new parts between blocks | attribute preservation | Helped |
Surprising findings
- Generation order alone changes BLIP-VQA substantially even with the stronger SD3 backbone (0.6363 inverse vs 0.8055 proposed).
Limitations
- Evaluation relies on a small set of qualitative examples and two automatic metrics; BLIP-VQA scores were modified from the T2I-CompBench multiplicative definition to an average.
- Depends on the LLM (GPT-4o) for correct occlusion analysis and decomposition, and on Lang-SAM segmentation quality.
Experimental setup
Benchmarks: T2I-CompBench BLIP-VQA CLIP-Score
Metrics: BLIP-VQA CLIP-Score
Baselines: GaussianDreamer MVDream LucidDreamer Progressive3D GaussianDreamer + SD3
Builds on: DreamFusion GaussianDreamer GALA3D MVDream Progressive3D
Closest papers by meaning
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.86)
- CG-MLLM: Captioning and Generating 3D content via Multi-modal Large Language Models (ICML 2026 · similarity 0.86)
- CompGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians (CVPR 2025 · similarity 0.86)
- Large-Vocabulary 3D Diffusion Model with Transformer (ICLR 2024 · similarity 0.86)
- Towards Text-guided 3D Scene Composition (CVPR 2024 · similarity 0.86)
- HyperSDFusion: Bridging Hierarchical Structures in Language and Geometry for Enhanced 3D Text2Shape Generation (CVPR 2024 · similarity 0.85)
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.