FiT: Flexible Vision Transformer for Diffusion Model

Authors: Zeyu Lu, Zidong Wang, Di Huang, Chengyue Wu, Xihui Liu, Wanli Ouyang, Lei Bai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments demonstrate the exceptional performance of Fi T across a broad range of resolutions. Repository available at https://github.com/whlzy/Fi T.
Researcher Affiliation Collaboration 1Shanghai Artificial Intelligence Laboratory 2Shanghai Jiao Tong University 3Tsinghua University 4Sydney University 5The University of Hong Kong.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Repository available at https://github.com/whlzy/Fi T.
Open Datasets Yes We train class-conditional latent Fi T models under predetermined maximum resolution limitation, HW <= 256^2 (equivalent to token length L <= 256), on the Image Net (Deng et al., 2009) dataset.
Dataset Splits No The paper mentions general training settings (learning rate, batch size, EMA, diffusion hyper-parameters) but does not provide specific dataset split information (percentages, counts) for training, validation, or test sets, nor does it cite predefined splits with specification.
Hardware Specification No The paper mentions 'GPU hardware' as a constraint but does not provide specific details on the GPU models (e.g., NVIDIA A100), CPU models, or cloud computing instance types used for the experiments.
Software Dependencies No The paper mentions software like AdamW, TensorFlow, and Stable Diffusion, but it does not specify exact version numbers for these software components, which is required for reproducible dependency descriptions.
Experiment Setup Yes We use the same training setting as Di T: a constant learning rate of 1e-4 using AdamW (Loshchilov & Hutter, 2017), no weight decay, and a batch size of 256. Following common practice in the generative modeling literature, we adopt an exponential moving average (EMA) of model weights over training with a decay of 0.9999. All results are reported using the EMA model. We retain the same diffusion hyper-parameters as Di T.