search

justinpinkney / stable-diffusion

MIT License
1.45k stars 266 forks source link

idream -getting poor results #41

Open orydatadudes opened 2 years ago

orydatadudes commented 2 years ago

HI thank you for idream implementation i used DeepFashion dataset to check if i can use change human pose ( like the paper show ) unfortunately i didn't succeed to get good generated images for example : source image MEN_Denim_id_00000089_3_28_1_front

after trying to change posture (“A person with crossed arms”)

i got bad results ( for different levels of alpha )

0.6: generated_image_MEN_Denim_id_00000089_3_28_1_front jpg_0 6

0.8 generated_image_MEN_Denim_id_00000089_3_28_1_front jpg_0 8

0.9: generated_image_MEN_Denim_id_00000089_3_28_1_front jpg_0 9

i tried with more images and more text prompt and increase alpha range but nothing works for me did someone succeed getting good results ? thanks

justinpinkney commented 1 year ago

Yeah it seems quite sensitive to parameters, I found sometimes that doing the first stage training, (training the embedding) for longer helps.