Optimized View and Geometry Distillation from Multi-view Diffuser
Youjia Zhang, Zikai Song, Junqing Yu, Yawei Luo, Wei Yang.
IJCAI 2025
Our technique produces multi-view images and geometries that are comparable, sometimes superior particularly for irregular camera poses, when benchmarked against concurrent methodologies such as SyncDreamer and Wonder3D, without training on large-scale data. To reconstruct 3D geometry from the 2D representations, our method is built on the instant-NGP based SDF reconstruction instant-nsr-pl.
Our proposed rectification method essentially combines the [unconditional noise] prediction from the base model and the [conditional noise] prediction from the fine-tuned model. This can be further interpreted through the formulation provided in Appendix A.
Concurrent methods, like SyncDreamer and Wonder3D impose limitations on the viewing angles of the input image.
Where setting λ = 1, we get Formula SDS. We observed that setting λ = 0 can significantly improve the details of the 3D results generated using SDS.
# USD image-to-3D
python launch.py --config configs/usd-patch.yaml --train --gpu 0
text.to.3D.mp4
# --------- Stage 1 (NeRF, SDS guidance, lambda=0) --------- #
python launch.py --config configs/usd-text-to-3D-patch.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple"
# --------- Stage 2 (Geometry Refinement, SDS guidanc) --------- #
# refine geometry with 512x512 rasterization
python launch.py --config configs/usd-text-to-3D-geometry.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" system.geometry_convert_from=path/to/stage1/trial/dir/ckpts/last.ckpt
# --------- Stage 3 (Texturing, SDS guidance, lambda=0) --------- #
# texturing with 512x512 rasterization
python launch.py --config configs/usd-text-to-3D-texture.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" system.geometry_convert_from=path/to/stage2/trial/dir/ckpts/last.ckpt
We have intensively borrow codes from the following repositories. Many thanks to the authors for sharing their codes.
@article{zhang2023optimized,
title={Optimized View and Geometry Distillation from Multi-view Diffuser},
author={Zhang, Youjia and Yu, Junqing and Song, Zikai and Yang, Wei},
journal={arXiv preprint arXiv:2312.06198},
year={2023}
}