Diffusing Differentiable Representations

Authors: Yash Savani, Marc Finzi, J. Zico Kolter

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

Reproducibility Variable Result LLM Response
Research Type Experimental All our experiments used the Hugging Face implementation of Stable Diffusion 1.5 (SDv1.5) from Rombach et al. [2022] with the default hyperparameters as the noise predictor model. We conducted our experiments on a single NVIDIA A6000 GPU. We used the complete Langevin SDE given by the DDIM [Song et al., 2022] procedure for all our experiments, with η = 0.75 as the stochastic interpolation hyperparameter. We interspersed forward and reverse steps to harmonize the diffrep constraints in the sampling procedure. For the suboptimization problem described in Eq. 3, we used the Adam optimizer for 200 steps.
Researcher Affiliation Academia Yash Savani Carnegie Mellon University ysavani@andrew.cmu.edu Marc Finzi Carnegie Mellon University mfinzi@andrew.cmu.edu J. Zico Kolter Carnegie Mellon University zkolter@andrew.cmu.edu
Pseudocode Yes Algorithm 1 DDRep with Re Paint
Open Source Code No The code is not currently in a state ready for distribution. We will release the code after we have some time to clean it up.
Open Datasets Yes We used the first 100 captions (ordered by id) from the MS-COCO 2017 validation dataset [Lin et al.] as prompts.
Dataset Splits No The paper uses the MS-COCO 2017 validation dataset as prompts but does not specify train/validation/test splits of this data for their experiments.
Hardware Specification Yes We conducted our experiments on a single NVIDIA A6000 GPU.
Software Dependencies Yes All our experiments used the Hugging Face implementation of Stable Diffusion 1.5 (SDv1.5) from Rombach et al. [2022] with the default hyperparameters as the noise predictor model.
Experiment Setup Yes For the suboptimization problem described in Eq. 3, we used the Adam optimizer for 200 steps. We used the complete Langevin SDE given by the DDIM [Song et al., 2022] procedure for all our experiments, with η = 0.75 as the stochastic interpolation hyperparameter. We interspersed forward and reverse steps to harmonize the diffrep constraints in the sampling procedure.