cs.CVMar 15, 2025
SaveDecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation
Organizations: Arizona State University · Intel Labs · Stability AI
Abstract
Current text-to-3D methods excel at generating single objects but falter on compositional prompts. We argue this failure is fundamental to their optimization schedules, as simultaneous or iterative heuristics predictably collapse under a combinatorial explosion of conflicting gradients, leading to entangled geometry or catastrophic divergence. In this paper, we reframe the core challenge of compositional generation as one of optimization scheduling. We introduce DecompDreamer, a framework built on a novel staged optimization strategy that functions as an implicit curriculum. Our method first establishes a coherent structural scaffold by prioritizing inter-object relationships before shifting to the high-fidelity refinement of individual components. This temporal decoupling of competing objectives provides a robust solution to gradient conflict. Qualitative and quantitative evaluations on diverse compositional prompts demonstrate that DecompDreamer outperforms state-of-the-art methods in fidelity, disentanglement, and spatial coherence.
Figures & tables
Figure 1: Illustration of high-quality compositional 3D assets generated by DecompDreamer for complex text prompts. Existing methods often miss objects or distort spatial relationships, while DecompDreamer accurately captures geometry and preserves inter-object layout.
Figure 2: A visual taxonomy of optimization heuristics. (a) Holistic uses a single global loss. (b) Simultaneous applies all losses concurrently. (c) Iterative applies losses in a sequential loop. (d) Our staged curriculum temporally decouples relational and object losses
Figure 3: Overview of the DecompDreamer pipeline. Given a text prompt, a VLM generates a scene graph to guide a coarse initialization. The core of our method is a composition-aware optimization curriculum that first models joint relationships to build a coherent structure, then refines individual objects to produce high-fidelity, disentangled 3D assets.
Figure 4: Qualitative comparison between the proposed DecompDreamer and state-of-the-art text-to-3D generators. More results can be found in Section A.12 of the appendix.
Figure 5: Empirical validation of optimization heuristics. The plots show the average object, scene, and edge losses for a complex 4-object text prompt ("A knight in shining armor…"). (a) GALA3D converges suboptimally, (b) GraphDreamer diverges, while (c) our staged approach (DecompDreamer) converges stably.
Figure 6: Qualitative comparison of object disentanglement in GALA3D, GraphDreamer, and DecompDreamer. Each method displays the full 3D object (left) and its components (right).
Figure 7: Results of an ablation study on DecompDreamer with the text prompt: “A knight in full shining armor galloping on a brown horse, wearing a blue hat, and holding a sword". Additional ablation studies can be found in Figure 11 of the appendix.
Figure 8: Qualitative comparison of DecompDreamer with different coarse intializers: (a) TripoSR initialization, (b) Point-E initialization.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Additional empirical comparison of optimization heuristics across three complex text prompts (rows). We plot individual loss components (log scale) over iterations. The columns illustrate that GALA3D (Heuristic 2) converges suboptimally with high remaining loss, GraphDreamer (Heuristic 3) exhibits significant instability and divergence, while our staged approach, DecompDreamer (Heuristic 4), achieves stable convergence to lower loss minima.
| Notation | Description |
|---|---|
| Compositional text prompt | |
| VLM | Vision-Language Model |
| Total number of optimization iterations | |
| Generated image representation | |
| Scene graph extracted from VLM | |
| Number of objects in the scene |
Table 4: Summary of Notations Used in the Algorithm
Figure 11: Results of an ablation study on DecompDreamer under different training configurations.
Figure 12: Results of an ablation study on DecompDreamer with (right) and without (left) the azimuth adjustment for object-view dependent text prompts.
Figure 13: Results of an ablation study on DecompDreamer with (right) and without (left) the negative prompts for the objects
Figure 14: Results of an ablation study on DecompDreamer with (right) and without (left) spatial error correction.
Figure 15: More qualitative comparison of object disentanglement in GALA3D, GraphDreamer, and DecompDreamer. Each method displays the full 3D object (left) and its individual components (right).
Figure 16: Qualitative comparison on the prompt: ’A lego man riding a horse, the horse is wearing a blue hat.’ Layout-your-3D merges the hat with the horse’s head and entangles the rider within the body. In contrast, DecompDreamer successfully resolves these physical interactions.
Figure 17: Qualitative comparison with between feed-forward 3D object generation methods such as PartCrafter, Step1X-3D, and Hunyaun3d-2.0, against DecompDreamer.
Figure 18: Example usage of ChatGPT in DecompDreamer for generating compositional 3D objects.
| Prompt # | Prompt Text |
|---|---|
| 1 | A well-furnished bedroom with a Baroque-style bed with a red pillow on it, two wooden nightstands with pots on top of them, a square wooden table holding a decorative table lamp, a wooden wardrobe and a kid infront of the bed and a pet dog standing next to the bed between nightstand and wardrobe. |
| 2 | A well-furnished bedroom with a Baroque-style bed with a red pillow on it, two wooden nightstands, a square wooden table holding a decorative table lamp, a wooden wardrobe. |
| 3 | Three macaw parrots sharing a milkshake with three straws |
| 4 | A bedroom with a Baroque bed, two wooden nightstands, a square wooden table with a potting on it, and a wooden wardrobe |
| 5 | A knight in shining armor gallops on a brown horse, wearing a blue hat, holding a sword in his right hand. A princess sits beside him on the same horse, wearing a flowing gown and a jeweled crown |
| 6 | A bedroom with a bed, two wooden nightstands, a square wooden table with a table lamp on it, and a wooden wardrobe |
Table 7: Full Evaluation Benchmark Prompts
Figure 19: Qualitative comparison of the "Wizard" prompt utilizing proprietary GPT-4o (left) versus open-source Qwen-VL (right) initialization priors. Although the initial spatial estimations exhibit minor numerical deviations (as quantified in Table 8 ), ours spatial correction module successfully absorbs this variance. The final optimized geometry maintains strict relational fidelity, confirming the framework’s robustness without reliance on proprietary VLM supervision.
| Object | Spatial Coords | Angles | Scale | Azimuth |
|---|---|---|---|---|
| Prompt 1: A knight in shining armor gallops on a brown horse… | ||||
| 1. Knight | 5% | 0% | 8.3% | 0% |
| 2. Horse | 5% | 0% | 11.1% | 0% |
| 3. Cap | 5.4% | 0% | 0% | 0% |
| 4. Sword | 5.12% | 11% | 10.0% | 0% |
| 5. Crown | 5.1% | 0% | 0% | 0% |
Table 8: Relative Deviation of Initialization Priors (GPT-4o vs. Qwen-VL). Values show the overall vector percentage change (Euclidean deviation) from GPT-4o to Qwen-VL.
Figure 20: Examples of failure cases. (a) Difficulty in decomposing a single entity into fine-grained subcomponents—such as segmenting human body parts. (b) A case where DecompDreamer fails to model complex actions, such as a bag with its zipper open .
Figure 21: Qualitative comparison of GALA3D and GraphDreamer against DecompDreamer under the same Point-E initialization. Despite using the same low-fidelity coarse initializer, DecompDreamer produces more coherent object geometry and better preserves inter-object relationships in complex compositional prompts, while requiring comparable or lower runtime.
Figure 22: More qualitative comparison of GALA3D, GraphDreamer, MVDream, LucidDreamer, Magic3D and HiFA against DecompDreamer with SD2.1 and SD3.5 as backbones.
Figure 23: More qualitative comparison between the proposed DecompDreamer and state-of-the-art generators, including GALA3D Zhou et al. (2024) , GraphDreamer Gao et al. (2024) , Trellis Xiang et al. (2025) , MVDream Shi et al. (2024) , LucidDreamer Liang et al. (2024) , Magic3D Lin et al. (2023) and HiFA Zhu et al. (2024) .
Figure 24: More qualitative comparison between the proposed DecompDreamer and state-of-the-art generators, including GALA3D Zhou et al. (2024) , GraphDreamer Gao et al. (2024) , Trellis Xiang et al. (2025) , MVDream Shi et al. (2024) , LucidDreamer Liang et al. (2024) , Magic3D Lin et al. (2023) and HiFA Zhu et al. (2024) .
Figure 25: More qualitative comparison between the proposed DecompDreamer and state-of-the-art generators, including GALA3D Zhou et al. (2024) , GraphDreamer Gao et al. (2024) , Trellis Xiang et al. (2025) , MVDream Shi et al. (2024) , LucidDreamer Liang et al. (2024) , Magic3D Lin et al. (2023) and HiFA Zhu et al. (2024) .
More in this topic
3D Asset Generation
3D asset generation creates and edits objects, characters, materials, and scenes from text, images, video, or geometric constraints.
See all 167 papers