你加个ugly face试试

我也这么认为，脸部还原较之v6差了点，但是整体更好了

唉，国内模型都是微博美图秀秀厚滤镜的图

@Merjic 先用v7随机生成一张图片，正向提示词 1girl，获取这张图片的种子，再切换v2=v25模型，使用上一张图片的种子，再生成图片，生成的图片一模一样。。。。

你这不开玩笑嘛，不一样啊

@zl1417896412 你可以试试，固定种子后其他设置都一样，就换个模型，出来的图完全一样

我是同样的参数，生成的图片和例图不一样，复制参数的时候看到作者用的VAE模型是vae-ft-mse-840000-ema-pruned.safetensors。hash是: 2453b80bc1。但是我的vae-ft-mse-840000-ema-pruned.safetensors出图的Hash是735e4c3a44。在秋叶一键启动的模型管理里看到hash: 2453b80bc1是另外一个VAE的，叫autoencoder_fix_kl-f8-trinart_characters.ckpt。这影响出图的吗？

@marvinyorng808  你是不是遇到显存污染了。退出SD重启一下。

@cjs_young 确实是被污染了，开了两个web页面，页面上显示用的模型是v2=v25，但做出来图的详细信息里引用的是v7模型

光影效果完全不同了

@Merjic 谢谢你的建议（〜^∇^)〜，我尝试更换模型，在不变动众多标签（可能也许大概过多了吧(~‾▿‾)~）的情况下，仍然如此，最后尝试将迭代步数从30左右增加到40-60多就好了(更多没有尝试)，也许是标签过多了？或者显卡问题？谁知道呢？不想去探究了，最后感谢您的模型，太棒了

ヽ(*≧ω≦)ﾉ

@Merjic 好的，感谢

negative 也不是越多越好！还是建议要筛选适当的使用

CFG太高，调低。

@Merjic thank you very much!!

实际上是一个大型人物lora😏

并没有完全一样吧

没有，那使用插件实现的

@543485276463 
請問是甚麼插件? 我看都不會閃爍,好厲害,也想來做看看

@plokijuh0000  https://github.com/continue-revolution/sd-webui-animatediff.git

@Merjic hey..sorry I didnt get can u explain whats that?

最右上角左边算起第2个按键，有箭头指下的那个。

就靠18+顶上去的

You can try different LORAs

I'll talk to you as a beginner, a month ago I started using SD(stable diffusion acronym), so started from here : https://stable-diffusion-art.com/prompt-guide/

But, each model has its own prompting methods, like anime (hentai) based models (a.k.a checkpoints in civitai), they use danbooru tags, there are two types of prompting, long phrases with everything in them, each component separated by "," (e.g : woman, glasses, blue eyes, blonde hair, long hair...), then comes the WEIGHTS, this is extremely important to know in Stable diffusion because it's the atomic way to get the best results using a model, since every model contains its own weights, read this to understand more (https://dev.dezgo.com/guides/prompt-weighting/), from that point, you play with the numbers, yes, there's no magic way to know, you have to do a lot of trial and error, that's why you start with a low resolution, take the seed number, and play with it, it's the fun part actually, need anything, hit me up

Hey, to put it simply, your prompts is basically what you want and you make them kinda how you would make hashtags for example a female portrait would be like "1girl, solo, portrait, looking at viewer, flower garden" etc and sometimes adding "masterpiece, 4k, 8k" can help along with other finer details such as "god rays, soft shadows" but its not that necessary, also add specifics if you want her blonde, put blonde in the tags, short hair? put short blonde hair.

if you also want to make the AI focus more on a certain tag, lets say you really want it to focus on her being blonde you can add the tag in brackets instead and add a number like 1.3 like this "(short blonde hair:1.3)"

for negative tags, its really things you dont want to see like "bad fingers, disfigured face, bad quality, low quality, disfigured arms" etc

hope that helps, im no expert though but i get good results

我也曾經遇過 要不就是PROMPT 有敏感字眼, 有時候因為使用的 LORA...換個LORA,試一下,

@suihou827 我用同样的prompt，同样的Lora生出的图，部分能上传，部分不能，也有全部不能上传的时候，就是换了个大模型而已，好奇怪

@CHINGEL@CHINGEL 要看生成的图是否像未成年。。这裡好像不能上傳未成年人像照片 當然 系統有時候会誤判

@suihou827 没有指定年龄段，默认都是成年人图像

@CHINGEL 看 terms and service, 是有的

Content depicting or intended to depict photorealistic minors, regardless of context. For a more detailed explanation, please refer to our guidelines on depicting photorealistic minors.

@CHINGEL 系統誤判是小孩臉也有这可能

@suihou827 不太可能，但如果是系统出问题那也没办法，我最近出的一组图，同一个prompt的结果，上传上来2张，其他传不上来的脸跟传上来的也相差不大，肉眼可见的成熟脸，原因不明，另外看到有一个上传模型的哥们也遇到这个问题，他自己的模型出的例图也不能全上传上来，尴尬

@CHINGEL 檢測了有沒有放行? 我發圖這麼久也檢測了好多次 極大部份是誤判 只有兩次給刪了 其實刪了的圖也沒什麼....不明白他的邏輯

@suihou827 我有2批被删过，改掉MD5值重新上传还是被删，都来就放弃了，不懂他们的判定逻辑

@CHINGEL 我近兩個月經常TOS. 你upload時當有TOS系統會有提醒pop up, 當你click ok, 系統會刪除你upload的圖, 但你一路重覆再upload, 系統最後會給你過的, 還有,我有點"Mature?"同delete prompt, 好似都有幫助

@jasonccc 重复上传我试过没用，点 “mature”这个我去试试看，delete prompt会影响出图效果吧？

@CHINGEL delete prompt 指上傳時顯示的資訊

@jasonccc 刚刚试了下，点“mature”没用

@jasonccc 把图片另存为JPG，丢掉生图信息，能上传了，，，不错不错，管用，谢谢

inpaint and use adetailer to repaint the hands

Of course, the model was trained on asians since the owner is Chinese. It is his "normal" ethnicity.

ya like what you downloaded a checkpoint thats been trained on asian girls XD