|
|
星级打分
平均分:5 参与人数:3 我的评分:未评
本帖最后由 zyl521 于 2024-1-6 02:11 编辑
凡是说链接失效的，你给我放截图，我每次有人留言链接失效我都检查一遍，结果哪次都没失效？？？
自从搬了四个懂得都懂XSeg后听各位大佬说不错，于是今日闲的皮疼没事做找了8个模型。
至于为啥挑这八个。众所不周知，在一些比较奇怪的情况下，人的脸会被拉长，而且有的时候会有酒窝之类的。
而且有的时候遮罩是比较奇怪的，练出来的也不那么到位（可能是我懒得找素材加懒的训练）
于是乎既然有了奇怪的遮罩，怎么能没有奇怪的模型呢，这八个是现有能找到表情比较夸张的模型，
由于预览模型表情比较6，我就在文件开头用6形容我看到预览图的表情，6越多，代表我看到的模型越夸张。
和XSeg一样，自行体会，预览图放不了。
依据我个人的理解，我个人是这样想的，当然我虽是计算机专业但主攻方向不是换脸，如果有理解误区请不吝指正。
我大概总结了以下几点：（下面的表情代指两点：此表情以及出现此表情的角度，我懒得打字。）
①如果src有这个表情，丹学习过这个表情，dst也有这个表情，那么可能此时训练最快。
②如果src有这个表情，丹没学习过这个表情，dst有这个表情，那么丹可能会根据表情训练并丰富其学习能力。
③如果src没有这个表情，丹学习过这个表情，dst有这个表情，那么会根据丹学习过的表情和扭曲来想src是dst表情时长啥样。
④如果src没有这个表情，丹也没学过这个表情，dst有这个表情，那么软件只能通过扭曲来想像当src是dst表情时长啥样，五官该放哪（类似于纯意淫）。
⑤如果src有这个表情，丹有没有学习过不重要，dst没有这个表情，那么软件想干嘛干嘛，估计可能也学学或者来拿放在表情不是那么匹配的图上用于意淫。
综上所述：src是地基，丹是基础建筑，丹不是很重要，丹跟学习速度挂钩。而最终出片要看src的质量。
当然以上也不是绝对的，如果缺表情，或者缺这个角度，即使对不上也会拿来学的，扭曲我觉得就是意淫。
所以我个人粗浅的认为丹这个东西，只是看其见“世面”的多少（看片的多少），以及其本身的智力（参数），还有有没有走火入魔（学习了错误的东西，很难纠正）。
如果各位下载是用于那啥的，那么这些标明的专丹也好，非专丹也好。就像个人，看过有些丹没见过的东西，不管是学习还是意淫可能都会快一些。
如果是正常用用的话，这些专丹本身素质应该都差不了多少，都是国外大佬练的。
以下这些为专人专丹：
DF-UD - res: 384, AB, dims: 352/88/88/16 人名：Maddie Ziegler
DF-UD - res: 416, AB, dims: 336/84/84/28 人名：Daisy Ridley
DF-UD - res: 320, AB, dims: 416/104/104/34 人名：Jessica Alba
DF-UD - res: 416, AB, dims: 416/80/80/16 人名：Dorothea Wierer
DF-UD - res: 416, AB, dims: 400/80/80/26 人名：Jennifer Love Hewitt
由于所有的模型我都下载，解压，删敏感图，补充基本参数，肉眼评级，看看与论坛已发模型是否有重复并重新使用极限压缩。
并且检查了一遍，由于某歌云盘现在下载大文件有时会分开压缩，需要解压合包，如遇个别不管用请留言反馈。
一个一个都打包好了，极限压缩完共18.3G，小模型大模型都有。
随便找了三个模型贴参数：
=============== Model Summary ===============
== ==
== Model name: DF-UD320WF_SAEHD ==
== ==
== Current iteration: 2309387 ==
== ==
==------------- Model Options -------------==
== ==
== resolution: 320 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: df-ud ==
== ae_dims: 416 ==
== e_dims: 104 ==
== d_dims: 104 ==
== d_mask_dims: 34 ==
== masked_training: True ==
== eyes_mouth_prio: False ==
== uniform_yaw: False ==
== blur_out_mask: False ==
== adabelief: True ==
== lr_dropout: y ==
== random_warp: False ==
== random_hsv_power: 0.0 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: none ==
== 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: 0.01 ==
== gan_patch_size: 60 ==
== gan_dims: 30 ==
== ==
==-------------- Running On ---------------==
== ==
== Device index: 0 ==
== Name: Tesla T4 ==
== VRAM: 13.43GB ==
== ==
=============================================
=============== Model Summary ===============
== ==
== Model name: DF-UD416WF_SAEHD ==
== ==
== Current iteration: 1601351 ==
== ==
==------------- Model Options -------------==
== ==
== resolution: 416 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: df-ud ==
== ae_dims: 400 ==
== e_dims: 80 ==
== d_dims: 80 ==
== d_mask_dims: 26 ==
== masked_training: True ==
== eyes_mouth_prio: False ==
== uniform_yaw: False ==
== blur_out_mask: False ==
== adabelief: True ==
== lr_dropout: y ==
== random_warp: False ==
== random_hsv_power: 0.0 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: none ==
== 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: 0.01 ==
== gan_patch_size: 60 ==
== gan_dims: 30 ==
== ==
==-------------- Running On ---------------==
== ==
== Device index: 0 ==
== Name: A100-SXM4-40GB ==
== VRAM: 37.47GB ==
== ==
=============================================
==================== Model Summary ====================
== ==
== Model name: DF-UDT-288_SAEHD ==
== ==
== Current iteration: 419999 ==
== ==
==------------------ Model Options ------------------==
== ==
== resolution: 288 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: df-udt ==
== ae_dims: 300 ==
== e_dims: 80 ==
== d_dims: 80 ==
== d_mask_dims: 22 ==
== masked_training: True ==
== eyes_mouth_prio: False ==
== uniform_yaw: True ==
== blur_out_mask: False ==
== adabelief: True ==
== lr_dropout: n ==
== random_warp: False ==
== random_hsv_power: 0.0 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: none ==
== clipgrad: False ==
== pretrain: False ==
== autobackup_hour: 0 ==
== write_preview_history: False ==
== target_iter: 0 ==
== random_src_flip: False ==
== random_dst_flip: False ==
== batch_size: 6 ==
== gan_power: 0.0 ==
== gan_patch_size: 48 ==
== gan_dims: 32 ==
== ==
==------------------- Running On --------------------==
== ==
== Device index: 0 ==
== Name: NVIDIA GeForce GTX 1070 Ti ==
== VRAM: 6.63GB ==
== ==
=======================================================
|
评分
-
查看全部评分
|
[复制链接]
|Archiver|手机版|deepfacelab中文网
|网站地图
楼主
发表于 2023-4-2 21:19:04
