Auto Cherry-Picker: Learning from High-quality Generative Data Driven by Language

Figure 1. Illustration of quality assessment of generated data samples using CLIS. (a) and (c) compare the quality of samples with different CLIS-L and CLIS-I scores, respectively. Samples with low CLIS fail to align accurately with the condition (e.g., contain extraneous objects or exhibit visual flaws). (b) and (d) compare the preferences of CLIS and CLIP score [27]. (e) compares different selection methods for the same volume of synthetic data used in downstream tasks, reporting APr and AP on the LVIS benchmark. See details in Sec. 4.2.