PlacidDreamer

Shuo Huang, Shikun Sun, Zixuan Wang, Xiaoyu Qin, Yanmin Xiong, Yuan Zhang, Pengfei Wan, Di Zhang, Jia Jia (Corresponding author). Tsinghua University & Kuaishou.

The official implementation of ACM Multimedia 2024 paper "PlacidDreamer: Advancing Harmony in Text-to-3D Generation".

Examples of text-to-3D results of PlacidDreamer, within **~25mins** on A100.

2D Score Distillation Experiments

Examples of 2D generation results of SDS (left) and BSD (right).

Sharing the same computational costs as SDS, BSD is able to reconstruct images that are much more similar to those produced by diffusion sampling.

Try:

cd Balanced_Score_Distillation
sh run.sh

Installation

git clone https://github.com/HansenHuang0823/PlacidDreamer.git --recursive
conda create -n PlacidDreamer python=3.10.13 cudatoolkit=11.8
conda activate PlacidDreamer
pip install -r requirements.txt
pip install submodules/diff-gaussian-rasterization/
pip install submodules/simple-knn/

Inference

### Image to multi-view images + mesh reconstruction
### GPU requirements: 30GB
### cactus.png ---> multiview_images/cactus/*.png + mesh/cacuts.obj
python imagetomesh.py --elev 10 --im_path cactus.png

### Finetune text-to-image diffusion models
### multiview_images/cactus/*.png ---> lora_checkpoints/cactus
sh train_dreambooth_lora.sh

### Balanced Score Distillation
# Standard BSD with guidance ratio that classifier : smoothing = lambda_ : 1. You can control the saturation.
python train.py --opt configs/cactus.yaml --lambda_ 18.0 --name cactus --lora_path lora_checkpoints/cactus

# A variation of BSD using SDS decomposition. This is equivalent to standard BSD with an automatically increasing lambda. While you cannot control saturation anymore, it mostly achieves satisfactory saturation.
# See explanation below.
python train.py --opt configs/cactus.yaml --auto_BSD true --name cactus --lora_path lora_checkpoints/cactus

Evaluation on T3Bench

We show the training configs and scripts on T3Bench at configs and scripts folders. Due to version updates, it is necessary to modify the storage path of the program output to ensure the pipeline runs smoothly.

### Evaluation of quality.
python T3Bench_Evaluation/eval_quality.py
### Evaluation of text alignment.
python T3Bench_Evaluation/eval_blip.py
# Replace the placeholder with your own GPT-4 api-key.
python T3Bench_Evaluation/eval_alignment.py

Credits

This codebase is built upon the LucidDreamer. Thanks to the authors for their great codebase and contribution to the community.