FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly

Authors: Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, Bingbing Ni

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have highlighted the superior editing capabilities of Focal Dreamer in both quantitative and qualitative evaluations.
Researcher Affiliation Collaboration Yuhan Li1, Yishun Dou2, Yue Shi 1, Yu Lei 1, Xuanhong Chen 1, Yi Zhang 1, Peng Zhou 1, Bingbing Ni1 1Shanghai Jiao Tong University, Shanghai 200240, China 2Huawei
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. It describes the method in prose and through diagrams.
Open Source Code No The paper does not provide an explicit statement or link for the release of its source code.
Open Datasets No We assemble the dataset with 15 high-quality meshes found on the internet. The paper describes assembling its own dataset but does not provide concrete access information (link, DOI, citation) for public availability.
Dataset Splits No The paper mentions using a 'Synthetic Object Dataset' of '15 high-quality meshes' and conducting user studies with '65 participants' but does not specify any training, validation, or test dataset splits or cross-validation setup.
Hardware Specification Yes Focal Dreamer usually takes less than 30 minutes (3000 steps) for geometry and 20 minutes (2000 steps) for texture to converge on 4 Nvidia RTX 3090 GPUs
Software Dependencies No We use the Stable Diffusion implementation by Hugging Face Diffusers for SDS, and adopt DMTet to learn geometry and texture separately with NVDiff Rast as a differentiable renderer. The paper lists software components but does not provide specific version numbers for them.
Experiment Setup Yes We use Adam W optimizer with the respective learning rates of 1 10 3 and 1 10 2. Focal Dreamer usually takes less than 30 minutes (3000 steps) for geometry and 20 minutes (2000 steps) for texture to converge. The hyperparameter ξ is a small positive threshold to prevent topological structures from minor positive SDF values. Moreover, the closer pi is to the target region, the less the penalty, for this distance-aware setting permits geometry to overrun beyond the rough focal region slightly. σ1 = 0.05 and σ2 = 0.01 control how sensitive the loss is. k determines the extent of the soft merge and is set to 0.15 by default.