@WAI0731 我敲，wai佬秒回。可惜我不会用comfy不然就去comfy再验证一下了，哈哈。还好E预测版本够用了。

@WAI0731 试了一下comfy，感觉应该是forge的问题。comfy随便roll了几张，肢体还挺正常的（脚趾上至少是没问题了）。

虽然有点晚了，但是我估计是你采样选错了，V预测必须用Eular A，默认的DPM不行……目前不管ForgeUI还是ComfyUI都正常

@ShiroNekoAlpha 没，当时我就选的Eular A一直也没改过

@ShiroNekoAlpha V預測用uni_pc + beta 運作良好,不一定要用Eular A

Nice comparison

Why do you think the NAI version is better than the ILXL?

@WAI0731 好的 謝謝!我在試試看 感謝您的回覆!

Based on noobai wich means it is beter finetuned, supports e621 tags, v-pred version is v-pred.

@WAI0731 thanks! I did have webUI. However YT tutorials were outdated and github didn't seem to explain it either. Your method allowed me to access dev branch in an instant. Thanks WAI! Great model btw, just had a test run with it, works flawlessly!

Yeah anatomy and hands are hit or miss when using lora.

It is not recommended to exceed 4 vpred cfg. If a high cfg effect is required, it must be used in conjunction with the dynamic cfg or rescale cfg plug-in,It can greatly improve the quality problems caused by high CFG.

There should be no room for improvement in the compatibility of the vpred model with non-vpred trained LoRa.The vpred lora I trained myself works well on both eps and vpred models (the lora trained by vpred also performs better on eps than the lora trained by eps), but the lora trained by eps is a disaster on the vpred model.

@NTR_BLACK base Noobai model is not visible on CivitAi lora custom model trainer. Any recommended custom model to train it on?

版本已經最新了?
建議還是用ComfyUI, forge 用vpred 有很多奇怪的問題,也要額外的設置

@NTR_BLACK 在启动器里刷新并更新过，也按noobai用户手册里git pull更新过，但画出来的图有很多错误

@D00M5799 cfg 降到4以下看看

@NTR_BLACK 4和3都试了，还是不行

用reForge，这个能正常出图

@zonde306 请问这个reforge有办法弄到秋叶的整合包里吗，我从git上安装到本地了，但是纯净版的我不会用，b站上也没搜到关于reforge的教学视频

你Sampler（采样器）选的是Euler A吗？默认的DPM无法在V-pred类别模型生图

@ShiroNekoAlpha 是euler a，我从来都是用euler a和karras生图

@D00M5799 奇了怪了，我现在就在用秋叶佬的FORGE整合包，用WAI佬这个V预测是没问题的啊。你试试其他模型，如果所有EPS模型都没问题，但所有V预测模型都没法使用，我估计你可能下载错版本了……FORGE版本的秋叶启动器底下会写着“SD-WebUI-FORGE 版本：xxxxxxxxxxxx 的

@ShiroNekoAlpha 我很确定没下错版本，就是forge，但是我每次启动webui后都需要f5刷新网页以后才能正常用，e预测模型都能正常用，但v预测都是出的图里很多错误，尤其是色彩方面

因为C站的服务器是小水管，没法同时支持那么多Checkpoint给所有人生成，所以每周限定200个Checkpoint来生成。如果是热门模型的话会自动被算在这200个模型里面，但是WAI佬的这个V预测模型没EPS版本那么热门，只能在拍卖行里竞价来进每周候选。

@ShiroNekoAlpha 原来是这样😭这个对新人来说真好用，不需要太多技巧就能用了，好可惜

@ShiroNekoAlpha 原来如此，感谢解答

are you using the correct sampler? for V-pred models, the sampler must be Euler A or Euler, the default DPM didn't work.

Yes, besides using Euler A as a sampler, here is what I found to be effective:
At the very end of your positive prompt put this: "soft lighting, dim lighting" or (stronger) "dark, dim lighting". This fixes the overexposure and contrast issues with most artist tags for me.

@ShiroNekoAlpha Yup, I've been using Euler A.

@Noise_Connoisseur I'll give it a shot, thanks!

或许你可以试一下cfg缩放功能，权重0.3~0.7，图像过曝是v预测模型的通病，要注意的是用了它就不要用动态cfg

@Noise_Connoisseur Thank you so much for your advice — your lighting method really helped me as well.
That said, I still dont quite understand why oversaturation happens. Most prompts work fine, but for some reason, only a few end up oversaturated.

At first, I thought specific tags might be the cause (similar to how, in PonyXL, the sweat tag would somehow break the whole image), but after running some tests, that theory didnt hold up…

have you tried using resclaecfg?

This is based on noobAI while that is based on illustrious 1.0. NoodAI was going to be based on pony when illustrious 0.1 or 0.2 was "unofficially" made public and opensource by the dev. Illustrious 1 is a later iteration of the same 0.1 variant Nood is based on.

IF you're asking because you're not seeing much of a difference, Noob and illustrious share a fairly recent common ancestor as it were. BUT they are different models and are diverging with each generation, as it can be seen by WAI himself refusing to "upgrade" to using the latest open weights Illustrious 2.0 model for his checkpoint while this noobai model is based on the latest noob checkpoints.

NO

To give you an actual response in practical terms. Since this is based on noobAI,

1. Larger, more recent dataset. Noob incorporated a more recent danbooru dataset as well as e621

2. Vibrancy. Based on the way v prediction models create the image they're able to create more deep black colors and more vibrant colors overall

你使用他內置的VAE就可以了

Its like voting with buzz for models to be included in the onsite image generator.

WAI-NSFW-illustrious-SDX

因为我在用noob的时候老是出现问题，但illustrious就比较稳定

I'm having doubts this version will ever see an update sadly, which is a real bummer because it is amazing.

@sansmia well. . . surprise i guess

@monicalucci Lol.... Amazing timing.

Different model ancestry. Both models have a very detailed anime style, but this model here is based off NoobAIXL while BranchRouwei is based off Rouwei.