for moss在moss-base-bf16上进行训练,延续anima版本的标注倾向(懒),以1024的分辨率进行训练.Trained on moss-base-bf16, following the annotation tendency (lazzy of the anima version, with training at a resolution of 1024.风格无触发词,但额外设计了标注如下painterly, faux traditional media, detailed face, realistic, no lineart.召回效果符合预期No trigger words for styles, with additional designed tags as follows: painterly, faux traditional media, detailed face, realistic, no lineart. The recall performance meets expectations. 标注时detailed face某种程度上和realistic,no lineart是对立的. detailed face is to some extent opposed to realistic, no lineart.例图都是CFG =3.0 STEP=15~20,实际上高低步数(12~40),高低CFG(2.0~6.0)都能还原(?)不同的风格 All sample images use CFG =3.0 and STEP=15~20. In fact, both low and high step counts (12~40) and low and high CFG values (2.0~6.0) can reproduce different feeling.已知存在双人或者多人交互时退化的问题.单人的训练效果割人认为较anima base的版本更好.Degradation issues are known to occur for scenes with two or more interacting characters. The training performance for single-character scenes is considered better than that of the anima base version.目前只能在BASE上用,想要在SFT上用要在SFT上炼,否则压制不过GPT味.It can only be used on BASE for now.To use it on SFT, it needs to be trained on SFT; otherwise, it cannot suppress the GPT-style tendency.<hr />for base v010v2版本在v3依旧可以正常工作,但在basev1上存在偏色问题,因此重新在base v1版本以1280px分辨率训练The v2 version still works properly on v3, but has color deviation issues on base v1. Therefore, retraining is conducted on the base v1 version with a resolution of 1280px.实际上个人判断训练结果并不如v2上的表现,非常遗憾.In fact, I personally think the training results are not as good as those shown in v2, which is quite regrettable.另外,旨在为风格混合时提供特殊的质感,可以考虑使用低lora权重以及高触发词权重In addition, to create a special texture when blending styles, you can use low LoRA weights and high trigger word weights.<hr />for preview v2 and v3触发词@Nyte,推荐与@dino共同使用,封面2的ninomae ina'nis为混合风格,封面1的platinum \(shimmering dew\) \(arknights\)为该lora在1权重时的效果Trigger word @Nyte,Recommended for use with @dino. Ninomae Ina'nis on cover 2 features a mixed style, while Platinum (Shimmering Dew) (Arknights) on cover 1 shows the effect of this LoRA at a weight of 1.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1805424/nyte-tyde-style-or-illustrious">https://civarchive.com/models/1805424/nyte-tyde-style-or-illustrious</a>从该LORA模型蒸馏,在anima preview v2上以lora算法进行训练,纯tag标签,参杂了带勾线的作品,对打光做了标注Distilled from this LoRA model, trained using the LoRA algorithm on Anima Preview V2 with tag labels only. incorporate works with line art, and annotate the lighting especially.存在问题:Known issues:1.ninomae ina'nis的tentacle hair会受到影响不能正常生成Ninomae Ina'nis's tentacle hair may be affected and fail to generate properly.2.受训练素材印象,手指可能会比较暧昧Due to the influence of training materials, fingers may appear ambiguous.3.模型的构图似乎受到一定程度影响,会倾向于生成近景的特写.这与我在v1上训练类似素材的经验有出入.但可以通过约束lora跳过前面60%阶段生成解决这个以上3个问题he model's composition is somewhat affected, tending to generate close-up shots. This differs from my experience training similar materials on V1. However, all three issues above can be resolved by constraining the LoRA to skip the first 60% of the generation phase, especially when used with @Dino.

Nyte tyde style

在moss-base-bf16上进行训练,延续anima版本的标注倾向(懒),以1024的分辨率进行训练.Trained on moss-base-bf16, following the annotation tendency (lazzy of the anima version, with training at a resolution of 1024.已知存在双人或者多人交互时退化的问题.单人的训练效果各人认为较anima base的版本更好.Degradation issues are known to occur for scenes with two or more interacting characters. The training performance for single-character scenes is considered better than that of the anima base version.

Files with SHA256: cde0985674c7fc3aaef757ecc0fbf8126e5fa38759ce9a875a25b891f83d26ec