search

senmaoy / Easy-Text-to-Image

4 stars 0 forks source link

SSA-GAN #3

Open Heavenhjs opened 2 years ago

Heavenhjs commented 2 years ago

paper:SSA-GAN, results:IS:4.97±0.10,FID:15.699683021293254 contributor:hjs details: batchsize=24 nf=64 on CUB comments:The mask predictor improves the semantic consistency and assigns more text information to the main body of image

Heavenhjs commented 2 years ago

model name: old DF-GAN code（2020） results:IS:4.67±0.13 FID:22.974645958490953 contributor:hjs details: batch_size=32, nf = 32, on CUB comment: A simple and efficient baseline,it's good for beginners.

Heavenhjs commented 2 years ago

model name: new DF-GAN code（2022） results:IS:4.52±0.12 FID:14.81 contributor:hjs details: batch_size=32, nf = 32, on CUB.It is noteworthy that this code change the way to caculate FID. Horizontal flip and small-scale random offset are used to expand the pictures of the test set. comment: A simple and efficient baseline. It can calculate FID while training the model！