Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
DreamEdit: Subject-driven Image Editing
Authors: Tianle Li, Max Ku, Cong Wei, Wenhu Chen
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this project, we conduct automatic and human evaluations to understand the performance of our Dream Editor and baselines on Dream Edit Bench. |
| Researcher Affiliation | Academia | University of Waterloo Vector Institute, Toronto EMAIL |
| Pseudocode | Yes | Algorithm 1: Dream Editor Algorithm |
| Open Source Code | No | Our project and benchmark homepage is https://dreameditbenchteam.github.io/. The paper mentions external tools like GLIGEN and Segment-anything with links to their respective implementations, but does not provide specific access to the source code for the Dream Editor methodology itself. |
| Open Datasets | Yes | To this end, we first manually curate a new dataset called Dream Edit Bench containing 22 different types of subjects, and 440 source images, which cover diverse scenarios with different difficulty levels. Our project and benchmark homepage is https://dreameditbenchteam.github.io/. |
| Dataset Splits | Yes | Therefore, we divide the collected source images into 150 easy ones and 70 hard ones. For Subject Addition, ... Accordingly, we divide the 220 backgrounds into 122 easy-typed images and 98 hard ones. |
| Hardware Specification | Yes | All the experiments are run on a single A6000 GPU. |
| Software Dependencies | Yes | We use stable diffusion version 1.41 to fine-tune it with Dream Booth (Ruiz et al., 2023). |
| Experiment Setup | Yes | For all of the 30 subjects involved in Dream Bench (Ruiz et al., 2023), we set iteration number N = 5 and the mask dilation kernel m = 20. The encoding ratio k1/T is set to be 0.8 for the first iteration and decreases linearly as ki/T = k1/T i 0.1. |