In this work, we present 3DGen-Bench, the first comprehensive human preference dataset for 3D models. For efficient data collection, we build the 3DGen-Arena, a public voting platform in a pairwise battle manner. Then, we perform a comprehensive evaluation for numerous 3D generative models, and train an automatic scoring model 3DGen-Score, which aligns well with human judgment.

• We carefully design 1,020 prompts(510 for text and 510 for image), select 19 generative models(9 for text-to-3D and 13 for image-to-3D), and obtain 11,220 3D assets finally.
• We build 3DGen-Arena, a public benckmark platform for 3D generative models. You can contribute votes and check learderboard from here
• We released full prompts, 3D assets and human annotations on Huggingface, you are free to download from here.

The environment requires Python>=3.9, Cuda>=11.3

    conda create -n 3dgen python==3.9
    conda activate 3dgen
    pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
    pip install -r requirements.txt

make ./checkpoint dir and download our pretrained weights from here

python demo.py

• Step1 Download data

  make ./data dir and download preference data from here.

  • data/gallery: unzip images_prompts.zip and move into data/gallery/rgba
  • data/preference_annotation: human preference data in json format, splited into ["train", "valid", "test"]
  • data/surrounding_views: 360 degree surrounding videos of 3D assets in 3 formats
  • data/objects: .ply files of 3D assets(Unused here)

• Step2 Data preprocess

  • split .gif into frames (replace <xx> as needed)

      python preprocess/split_gif.py --mode <xx> --task <xx> --method <xx>
  

  • concate 4 views

      python preprocess/con_image.py --mode <xx> --task <xx> --method <xx>
  

  Parameters:

  • mode: rendering mode, choices = ["rgb", "normal", "geo"]
  • task: generative task, choices = ['text2shape', 'image2shape']
  • method: supported 3D generative model

• Step3 "Two-stage" training strategy

  • Stage1 Contrastive Loss

      accelerate launch --dynamo_backend no --gpu_ids all --num_processes 1  --num_machines 1 --use_deepspeed trainer/scripts/train.py +experiment=clip_h_neg
  

  • Stage2 Evaluation Loss

      accelerate launch --dynamo_backend no --gpu_ids all --num_processes 1  --num_machines 1 --use_deepspeed trainer/scripts/train.py +experiment=clip_h_2