Gradient Guidance for Diffusion Models: An Optimization Perspective

Authors: Yingqing Guo, Hui Yuan, Yukang Yang, Minshuo Chen, Mengdi Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Simulation and image experiments are provided in Section 7 to support theoretical findings on latent structure-preserving and optimization convergence.
Researcher Affiliation Academia Yingqing Guo Hui Yuan Yukang Yang Minshuo Chen Mengdi Wang Princeton University Department of Electrical and Computer Engineering, Princeton University. Authors emails are: {yg6736, huiyuan, yy1325, minshuochen, mengdiw}@princeton.edu.
Pseudocode Yes Algorithm 1 Gradient-Guided Diffusion for Generative Optimization and Algorithm 2 Gradient-Guided Diffusion with Adaptive Fine-tuning.
Open Source Code Yes Our code is released at https://github.com/yukang123/GGDMOptim.git.
Open Datasets Yes We independently sample a total of 65536 data points as our pre-training data set. We employ the Stable Diffusion v1.5 model [53] as the pre-trained model. For the reward model...Image Net [21].
Dataset Splits No The paper mentions a 'pre-training data set' and 'batch size' but does not provide specific training, validation, or test dataset splits or percentages.
Hardware Specification Yes We summarize the time cost of our experiments on one NVIDIA A100 GPU in Table 1.
Software Dependencies No The paper mentions using U-Net, Adam optimizer, and DDPM implementation, but does not provide specific version numbers for software dependencies like PyTorch, TensorFlow, or Python.
Experiment Setup Yes For linear data structure, we set the data s ambient dimension as D = 64 and the linear subspace dimension as d = 16... We discretize the backward process to have 200 time steps... the U-Net is trained using our generated data set for 20 epochs. We use Adam as the optimizer, set the batch size as 32, and set the learning rate to be 10 4.