梁洁liae模型精心打造可直接合成 500W底单 200W正训

675966766 · 发表于 2025-8-5 22:00:28

星级打分

1
2
3
4
5

平均分:5 参与人数:1 我的评分:未评

本帖最后由 675966766 于 2025-9-2 22:00 编辑

使用论坛500W次的作为底丹。自己进行200W次训练
后期开GAN训练

ai320_SAEHD_preview_SAEHD src-src.jpg

以下是找网图直接出的效果。
训练素材除了论坛的自己添加了大部分素材表情丰富
闭眼/皱眉等/酒窝/色彩等能有一定效果

22 35B241.JPG

带遮罩训练。试了下闭眼效果还行，表情，颜色和相似度感觉可以的
大角度表现还不够

胸 646.png

============ Model Summary =============
==                                  ==
==          Model name: ai320_SAEHD ==
==                                  ==
==    Current iteration: 8062128    ==
==                                  ==
==---------- Model Options -----------==
==                                  ==
==          resolution: 320       ==
==          face_type: wf       ==
==    models_opt_on_gpu: True       ==
==                archi: liae-ud    ==
==             ae_dims: 256       ==
==             e_dims: 64       ==
==             d_dims: 64       ==
==          d_mask_dims: 22       ==
==    masked_training: True       ==
==    eyes_mouth_prio: True       ==
==          uniform_yaw: False    ==
==       blur_out_mask: True       ==
==          adabelief: True       ==
==          lr_dropout: n          ==
==          random_warp: False    ==
==    true_face_power: 0.0       ==
==    face_style_power: 0.0       ==
==       bg_style_power: 0.0       ==
==             ct_mode: rct       ==
==             clipgrad: True       ==
==             pretrain: False    ==
==    autobackup_hour: 0          ==
== write_preview_history: False    ==
==          target_iter: 0          ==
==    random_src_flip: False    ==
==    random_dst_flip: False    ==
==          batch_size: 4          ==
==          gan_power: 0.1       ==
==       gan_patch_size: 40       ==
==             gan_dims: 16       ==
==    random_hsv_power: 0.05       ==
==                                  ==
==------------ Running On ------------==
==                                  ==
==       Device index: 0          ==
==                Name: NVIDIA L4 ==
==                VRAM: 20.48GB    ==
==                                  ==
=======================================