文本生成图像研究综述

曹寅,秦俊平,马千里,孙昊,闫凯,王磊,任家琪

Survey of text-to-image synthesis

Yin CAO,Junping QIN,Qianli MA,Hao SUN,Kai YAN,Lei WANG,Jiaqi REN

表 1 基于GAN的文本生成图像方法在不同数据集上的各类评价指标对比

Tab.1 Comparison of various metrics for GAN-based text-to-image generation methods on different datasets

方法 CUB鸟类数据集 Oxford-120花卉数据集 COCO数据集 IS FID R-precision IS FID IS FID R-precision GAN-INT-CLS[2] 2.32 68.79 — 2.66 79.55 7.95 60.62 — StackGAN[10] 3.70 51.89 — 3.20 55.28 8.45 74.65 — StackGAN++[25] 4.04 15.30 — 3.26 48.68 8.30 — — AttnGAN[11] 4.36 23.98 67.83 — — 25.89 35.49 85.47 DM-GAN[53] 4.75 16.09 72.31 — — 30.49 32.62 88.56 ControlGAN[36] 4.58 — 39.33 — — 24.06 — 82.43 SEGAN[34] 4.67 18.16 — — — 27.86 32.28 — MirrorGAN[41] 4.56 — 57.67 — — 26.47 — 74.52 XMC-GAN[60] — — — — — 30.45 9.33 71.00 CI-GAN[66] 5.72 9.78 — — — — — — DF-GAN[57] 5.10 14.81 44.83 — — — 21.42 67.97 DiverGAN[37] 4.98 15.63 — 3.99 — — 20.52 — SSA-GAN[58] 5.17 15.61 75.9 — — — 19.37 90.60 Adam-GAN[59] 5.28 8.57 55.94 — — 29.07 12.39 88.74 TVBi-GAN[71] 5.03 11.83 — — — 31.01 31.97 — 文献[67]方法 4.23 11.17 — 3.71 16.47 — — — Bridge-GAN[92] 4.74 — — — — 16.40 — — textStyleGAN[63] 4.78 — 49.56 — — 33.00 — 88.23 文献[45]方法 3.58 18.14 — 2.90 34.97 8.94 27.07 — SD-GAN[38] 4.67 — — — — 35.69 — — PPAN[29] 4.38 — — 3.52 — — — — HfGAN[27] 4.48 — — 3.57 — 27.53 — — HDGAN[26] 4.15 — — 3.45 — 11.86 — — DSE-GAN[61] 5.13 13.23 53.25 — — 26.71 15.30 76.31