A tea cup of iced coke
This is the official database for Lost in Translation: Latent Concept Misalignment in Text-to-Image Diffusion Models(ECCV2024).
We have released our dataset in "version1/" which includes:
173_level5_experiment_data_6patterns.json includes 173 level 5 concept pairs that we used in our experiment, derived from 4 LC-Mis patterns initially summarized by human experts.
90_level5_6_new_patterns(phase3).json includes 90 level 5 concept pairs from newly generated LC-Mis patterns by GPT-3.5 in phase 3, inspired by the 4 initial LC-Mis patterns.
76_level5_6_merged_patterns(phase4).json includes 76 level 5 concept pairs. These are the phase 4 result of GPT-3.5 fusing the newly generated LC-Mis patterns with the initial 4 LC-Mis patterns.
Note: The level ratings of concept pairs in this dataset were generated using the version of the Midjournal image generation system that was available online as of October 1, 2023. The evaluation of these images was conducted through human assessments. It should be noted that the level rating results of concept pairs may vary depending on the specific type and version of the image generation model used. This variation is due to differences in model architectures and updates in the image generation technologies.
In our future work, we are committed to expanding the range of model types and versions used in our dataset and will continuously iterate our data mining algorithms to enhance and enlarge the dataset.