Factorized Diffusion Architectures for Unsupervised Image Generation and Segmentation

Authors: Xin Yuan, Michael Maire

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that our model achieves accurate unsupervised image segmentation and high-quality synthetic image generation across multiple datasets.
Researcher Affiliation Collaboration Google yuanxzzz@google.com Michael Maire University of Chicago mmaire@uchicago.edu
Pseudocode Yes Algorithm 1 Training Masked Diffusion ... Algorithm 2 Image and Mask Generation
Open Source Code No Answer: [No] Justification: We will release the code upon acceptance.
Open Datasets Yes We evaluate on: (1) real image segmentation, (2) image and region mask generation, using Flower [45], CUB [62], FFHQ [31], Celeb AMask-HQ [36], and Image Net [53].
Dataset Splits Yes For unsupervised segmentation on Flower and CUB, we follow the data splitting in IEM [54] and evaluate predicted mask quality under three commonly used metrics, denoted as Acc., IOU and DICE score [54, 9].
Hardware Specification Yes For all datasets except Image Net, we train 64 64 and 128 128 models on 8 and 32 Nvidia V100 32GB GPUs, respectively.
Software Dependencies No The paper mentions the use of U-Net and Adam optimizer but does not specify software dependencies like Python, PyTorch/TensorFlow, or CUDA versions.
Experiment Setup Yes We use Adam to train all the models with a learning rate of 10 4 and an exponential moving average (EMA) over model parameters with rate 0.9999. ... For Flower, CUB and FFHQ, we train the models for 50K, 50K, 500K iterations with batch size of 128, respectively. For Image Net, we train 500K iterations on 32 Nvidia V100 GPUs with batch size 512. We adopt the linear noise scheduler as in Ho et al. [27] with T = 1000 timesteps.