224WF-全角度万能模型-金晨

qoungyoung · 发表于 2022-4-16 20:02:05

星级打分

1
2
3
4
5

平均分:NAN 参与人数:0 我的评分:未评

本帖最后由 qoungyoung 于 2022-4-16 20:11 编辑

各位好久不见，这次带来的是使用猫大224WF新丹（https://dfldata.xyz/forum.php?mod=viewthread&tid=4290）训练的全角度金晨万能，使用方法依旧是无需训练直接合成（想要继续训练的朋友可以去论坛内搜索金晨的src数据集，此贴只提供模型）

合成的参考参数：
W：15-20
E：90
X：xseg-dst(推荐遮罩：镇坛之宝或者独孤求败)
模型参数如下：
================== Model Summary ===================
==                                              ==
==          Model name: 224df-d 4G_SAEHD       ==
==                                              ==
==    Current iteration: 719329                ==
==                                              ==
==---------------- Model Options -----------------==
==                                              ==
==                archi: df-d                   ==
==          resolution: 224                   ==
==             ae_dims: 192                   ==
==             e_dims: 48                   ==
==             d_dims: 48                   ==
==          d_mask_dims: 16                   ==
==          face_type: wf                   ==
==       session_name:                      ==
==    autobackup_hour: 2                      ==
==    maximum_n_backups: 24                   ==
== write_preview_history: False                ==
==             pretrain: False                ==
==             use_fp16: False                ==
==          target_iter: 0                      ==
==          batch_size: 8                      ==
==    models_opt_on_gpu: True                   ==
==    masked_training: True                   ==
==          eyes_prio: False                ==
==          mouth_prio: False                ==
==          uniform_yaw: True                   ==
==       blur_out_mask: True                   ==
==          adabelief: True                   ==
==          lr_dropout: n                      ==
==       loss_function: SSIM                   ==
==                   lr: 5e-05                ==
==             clipgrad: True                   ==
== retraining_samples: True                   ==
==    background_power: 0.0                   ==
==    true_face_power: 0.01                   ==
==    face_style_power: 0.0                   ==
==       bg_style_power: 0.0                   ==
==    eyes_mouth_prio: False                ==
==             cpu_cap: 8                      ==
==    preview_samples: 4                      ==
== force_full_preview: False                ==
==          gan_version: 2                      ==
==          gan_power: 0.001                ==
==       gan_patch_size: 36                   ==
==             gan_dims: 16                   ==
==       gan_smoothing: 0.1                   ==
==          gan_noise: 0.0                   ==
==          random_warp: False                ==
==    random_src_flip: False                ==
==    random_dst_flip: True                   ==
==    random_downsample: False                ==
==       random_noise: False                ==
==          random_blur: False                ==
==          random_jpeg: False                ==
==    random_shadow_src: False                ==
==    random_shadow_dst: False                ==
==    random_hsv_power: 0.0                   ==
==             ct_mode: rct                   ==
==       random_color: False                ==
==       src_pak_name: faceset                ==
==       dst_pak_name: faceset                ==
==       random_shadow: none                   ==
==                                              ==
==------------------ Running On ------------------==
==                                              ==
==       Device index: 0                      ==
==                Name: NVIDIA GeForce RTX 2060 ==
==                VRAM: 3.87GB                ==
==                                              ==
==          DeepFaceLab Me 2022 Sp1          ==
==---- RG 2.41 version was edited by kingboy! ----==
合成效果：[mq_bili]https://www.bilibili.com/video/bv1Fa411v7ZD[/mq_bili]

一些简单QA：
Q:这个模型的妆好浓啊，有办法消除吗
A:准备好你的src，关闭GAN打开眼嘴扔进去训练到妆容消失，然后GAN0.01小跑一下

Q:我觉得有些角度还是不全，我想补充一些
A:准备好你的src，关闭GAN扔进去训练到角度清晰，然后GAN0.01小跑一下

Q:可以在这个模型上继续跑其他src吗？
A:不建议，不如直接使用猫大的丹重新跑

另外宋轶模型也打算使用224WF重新训练，会在宋轶原贴继续更新，预计两周后发布，同时也在跑384WF的模型

kkk1kkk · 发表于 2022-4-17 22:41:51

还不错，挺牛逼的！

michaelleon · 发表于 2022-4-16 20:12:56

感谢分享

大财神 · 发表于 2022-4-16 20:20:16

支持神丹分享

随遇而安 · 发表于 2022-4-16 20:44:25

感谢分享

1234qwer111 · 发表于 2022-4-16 20:58:29

路过看看

daiandy · 发表于 2022-4-16 21:13:18

正脸不错，侧脸问题很明显，那些中远景的镜头切出来不模糊吗？

1420699 · 发表于 2022-4-16 21:37:29

感谢分享

longsxp · 发表于 2022-4-16 22:35:47

看参数不象是猫大的丹，象是正在输入的

deepy123 · 发表于 2022-4-16 22:44:44

感谢分享

yingguanchana · 发表于 2022-4-16 22:57:19

你发了个视频就被B站封号了？？？

		自动登录	找回密码
密码			立即注册（仅限QQ邮箱）

224WF-全角度万能模型-金晨

万事如意节日勋章