The module already has VAE. Unless you have specific coloring requirements, it is recommended to use the VAE inside the module. Currently, personally recommend NoobAI ( 1.7, 1.9, 2.4, 2.8, 3.3) . If the FP32 upload fails multiple times, stop attempting to upload.Buy a Coffee for me☕️
Suggested: Txt2img 720p+ with Hires fix. Upscaler recommended Remacri setting to 1080p+.
Prompt Prefix:
masterpiece,best quality,perfect quality,absurdres,newest,very aesthetic,vibrant colorsNegative Prompt:
shiny skin, shiny clothing, worst quality, low quality, lowres, signature, username, logo, bad hands, mutated hands
V3.0, V2.9 Variant suggested: forge and use latent modifier integrated
Suggested Resources: Eyeful, adetailer, remacri, Face_finder
Simplified instructions for the standard guidelines of NoobAI
Character and artist tags Prompts achieve reliable performance when applied to NOOBAI CSV files with the Danbooru+e621 dataset exceeding 300 images in the official records. ( NoobAI huggingface csv).
The illustrious checkpoint pre-release beta version for testing must be used in accordance with the illustrious XL standard usage guidelines.This illustrious checkpoint is not stable, and the background sometimes disappears
Description
Fine-tuning and detail adjustments.
FAQ
Comments (19)
哪个版本最好用，谁能推荐一下
V2.6比较稳定，V2.7颜色比较自然
如果你想用的lora比较旧且新版无法支援的话建议1.9或1.7
新手尝试了这个indigo void版本，以及另一个下载量更多的indigo mix版本。在相同的设置下，这个版本的结果更准确（比如写white fur就不会乱出其他毛色，写floppy ears就几乎都是折耳而不是立耳），颜色更自然，并且从没发生过身体穿模的错误。
唯一的问题是（不是缺点），这个void版本好像自带一种画风，如果想用lora模仿某个作者的画风，那么最后得到的就是一种混合画风。而那个mix版本用lora得出的画风就和原作者特别像。
你的问题很难解释，很多lora都是直接使用indigo mix去训练的，理论上lora应该要用根checkpoint(noobai)去进行训练，这样其他noobai系列的checkpoint才会比较好支援，很多Lora你会看到连示范图一张 noobai checkpoint都没有用，还有现在大多 lora 都是用illustrious…，以及就像现实中就算同一棵树有相同的根但有些叶子会比较黄有些叶子会比较大...，情况很复杂...，如果你真的要达到你说的效果的话可能要自己用这模组去训练Lora
@ok_furry_image 谢谢解答，虽然没有特别看得懂。我后来又学了一些提示词方面的技巧，模仿的结果已经可以让自己满意了。你的模型真的很棒，indigo mix可能有别的我不知道的好处，但是void版本对我来说更简单，相同的提示词给的结果会更符合心意。
Who is Shiu Ui? Most of your prompt includes this artist, but I can't find any information about her/him.
我觉得可能是时雨羽衣，但我也不是很确定，可能是NOOBAI本身对提词有做些简化，至少我Shiu Ui与Shigure Ui两个都试过都跑出的东西近似一样
I'm a newbie in AI gen stuff, so this might be a stupid question: does this work with vPred? I'm getting noisy images with it no matter what i do. This model is based on NoobAI so i assume it should work?
This is NoobAI’s e-prediction. The v-prediction has its own dedicated page( I’m not very good at updating the v-prediction, and keeping it updated is a bit tiring.) Your issue is likely caused by using a LoRA that isn’t supported at all. From what I’ve observed, many beginners use LoRAs randomly, and it’s also possible they didn’t read the prompt instructions.
@ok_furry_image Thank you for your thorough reply. I didn't use any LoRAs, and used workflow was pretty basic, so i guess the issue was an incompatible checkpoint.
@bringer35 That could be due to insufficient steps.
@ok_furry_image Everyone recommend and use 26-30 steps on simillar models, i've never set more than 28. How many steps do you suggest?
@bringer35 Basically the same, What about resolution?
@ok_furry_image From my observation the native resolution shouldn't exceed 1920x1080 since the results will have lots of artefacts. So i'm usually setting something close to 1024x1024 depending of the results i want to see, like more height or width.
Top
Perhaps this is peak?
I've been quite busy lately and haven't had time to plan updates.
@ok_furry_image no worries, this model is amazing