|
|
星级打分
平均分:1 参与人数:1 我的评分:未评
本帖最后由 夢生若汐顔 于 2024-1-7 19:59 编辑
本人在花200元买的(现在想想确实买贵了，初入坑时买的）400万次底丹基础上，正向训练145万次的刘诗诗专丹。刘诗诗scr素材选用本站 脸包 的2558张和602387193c 的15000张（精简后有删减）素材，加上本人提取的步步惊心 步步惊情 轩辕剑 采访视频 直播视频等素材共1万9千多张训练。此丹是使用预训练的丹底（丹底并非我训练，DST素材未知，但本人试过训练几小时即可使用）。专丹是使用上述SCR素材， 因为是个人自用， DST素材根据自己喜好添加，每次使用不一样，训练步骤也全由个人喜好，喷子及只会嘴嗨的请绕道，并不是专门为某些人训练的他们想象中的万能神丹，介意勿购，谢谢！
20220925更新，已迭代到280W，之前以购买过的可在原链接下载。
另附赠训练步骤详解及总结（在外网2022年9月12日更新的，DFL2.0 使用指南和问题答复上翻译总结），及其中一训练详解章节的原文翻译。
================= Model Summary ===================
== ==
== Model name: WF256DFUD_SAEHD ==
== ==
== Current iteration: 2834707 ==
== ==
==---------------- Model Options -----------------==
== ==
== resolution: 256 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: df-ud ==
== ae_dims: 256 ==
== e_dims: 64 ==
== d_dims: 64 ==
== d_mask_dims: 22 ==
== masked_training: True ==
== uniform_yaw: False ==
== lr_dropout: y ==
== random_warp: False ==
== gan_power: 0.1 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: rct ==
== clipgrad: True ==
== pretrain: False ==
== autobackup_hour: 2 ==
== write_preview_history: False ==
== target_iter: 0 ==
== random_flip: True ==
== batch_size: 12 ==
== eyes_mouth_prio: False ==
== blur_out_mask: False ==
== adabelief: True ==
== random_hsv_power: 0.0 ==
== random_src_flip: False ==
== random_dst_flip: True ==
== gan_patch_size: 32 ==
== gan_dims: 16 ==
== ==
==------------------ Running On ------------------==
== ==
== Device index: 0 ==
== Name: NVIDIA GeForce RTX 3080 ==
== VRAM: 9.25GB ==
== ==
====================================================
10月24日更新：
赠送两颗娜娜专丹，原来购买过的可直接在原链接下载，无需重新购买。
1. 在底丹的基础上加157万迭代。因为是自用丹，之前的是根据几段目标视频，每次换的DST素材不一样。 最后几十万迭代，DST使用的是坛主的FFHQ素材。
scr使用的是本站xinmi501505的素材。素材的光影和仰脸角度有所欠缺。外加B站上自己切的采访和自拍视频，但清晰度个人感觉不够清晰。
2.
在底丹的基础上，DST使用坛主的FFHQ素材加52万迭代。
娜娜的素材真是不好找，B站上用会员下的采访和自拍视频，切出来后感觉清晰度不够，所以在这没有用。
所以scr使用的全部是本站xinmi501505的素材。素材的光影和仰脸角度个人觉得还是有所欠缺。
================== Model Summary ===================
== ==
== Model name: WF256DFUD_SAEHD ==
== ==
== Current iteration: 1577520 ==
== ==
==---------------- Model Options -----------------==
== ==
== resolution: 256 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: df-ud ==
== ae_dims: 256 ==
== e_dims: 64 ==
== d_dims: 64 ==
== d_mask_dims: 22 ==
== masked_training: True ==
== uniform_yaw: False ==
== lr_dropout: y ==
== random_warp: False ==
== gan_power: 0.1 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: rct ==
== clipgrad: True ==
== pretrain: False ==
== autobackup_hour: 2 ==
== write_preview_history: False ==
== target_iter: 0 ==
== random_flip: True ==
== batch_size: 12 ==
== eyes_mouth_prio: False ==
== blur_out_mask: False ==
== adabelief: True ==
== random_hsv_power: 0.0 ==
== random_src_flip: False ==
== random_dst_flip: True ==
== gan_patch_size: 32 ==
== gan_dims: 16 ==
== ==
==------------------ Running On ------------------==
== ==
== Device index: 0 ==
== Name: NVIDIA GeForce RTX 3080 ==
== VRAM: 9.25GB ==
== ==
====================================================
严正声明：本模型仅供学习研究，严禁用于任何非法行为，如作他用所承受的法律责任一概与本作者无关。
|
|
[复制链接]
|Archiver|手机版|deepfacelab中文网
|网站地图
楼主
发表于 2022-9-11 19:05:25
