From brief tests, it's gotten quite a bit better at anatomy (My main gripe was eyes during the Epsilon versions, which were messy half the time. Still can happen like an image model typically does but much less..exaggerated.)

Artist styles I'm also still trying, especially ones under E6 that might've flown under the radar and it seems to be decently nailing the ones I kinda had issues with. Watermarks still exist, but I'm learning to live with those until someone with a custom tune or the devs focus to zap those out in future Noob models.

Overall, since its like minutes new as of posting for now, I'd just go ahead and download it and compare your past Epsilon results to this version. It's based on 1.0 as the description says, so no extra data should be lost. Then again, I'm only just now testing it and so far don't need to do a whole surgery on my prompt to fix the art or eyes lol

(Edit: I did experience a bit of the black screens/blacked-out characters issue. So far I've just did the same as a few comments instructed (renaming/deleting the config files on Forge and running it so they update etc) and it seems to be working well again. But ofc, I hope this won't be a widespread issue or at least the devs update the instructions on its use and/or makes the next vpred avoid that more easily.)

@Veryhelpfulfur thanks for sharing!

From what I've tested also (using comfyui) you don't need "model sampling discrete" node and have to set v-pred to 'true'. It works right out of the box. If you are still using that node, that could be causing issues with your gens, so I'd suggest bypassing it or just deleting it.

artists aren't overbaked, colors are better.

@Enigmata It seems you didn't read the tutorial carefully. This tutorial not only teaches you to turn on the v prediction parameter, but before that, you also need to switch sd-scripts to the dev branch. As for lora, after actual testing, lora for Noob1.0 can be used on vpred0.5.

@Wenaka_ sorry, I don't use sd scripts. I have lora easy training scripts client. So tutorial is not so relevant for my training. But is v parametrization which is a option for SD2.1 the same as V parametrization from tutorial or not?

@Wenaka_ Scale V pred loss on or off?

I am using the Forge one click installer, is there something specific I need to do? Seems there is no longer a dev2 branch, all I can find on github is sd35 and Main. Still unable to get this model to work. Still getting completely white background images.

Edit: Going to try reForge.

The model should know senko perfectly well with senko \(sewayaki kitsune no senko-san\) as well as shiro and possibly the other characters

senko works perfectly fine. any char with 80+ quality artworks on danbooru works to some degree, most perfectly.

a1111 dev branch

velocity_prediction

https://rentry.org/wtfvpred

You might need to switch to dev version of whatever SD-gui you're using

@dmddmdd for comfy too?

You use v-pred model? In ComfyUI you need to use "ModelSamplingDiscrete" node and set it to v_prediction, zsnr might need to be "true" too

@munchkin That helped a lot. Thanks!

hi what do you use for A111?

pic 100+ diverses images from the show and train a Lora yourself then.
or pay someone to do it on civitai.

arcane characters and style is already on there for anything in danbooru. otherwise train lora yourself or offer a bounty.

@fizalpher oh I tried many pony Loras, but they look terrible with this checkpoint...

@reyjandplay380 DO NOT USE PONY LORAS FOR ILLUS for ANY XL-BASED MODEL. this model ALREADY has the most popular arcane chars, you DO NOT NEED lora.

This model is very different from il and pony, i don't use comfy but looks like there's a comfy workflow on their huggingface you can check there https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5/tree/main. Also keep in mind that they are still training, so for now they recommend using only euler sampler(not euler a) and cfg 4-5.

@Darkwen Thank you! I'm going to give that workflow a few tries!

switch to dev branch

What error are you getting?

Yep. Pony stands no chance, it can't even hope to stand against this. Don't know why they don't just give up, pony was pretty much dead a month after it came out anyways and now they wanna obfuscate artists & such.

@fizalpher Pony is pretty decent and very easy to prompt with, but yeah this model seems incredible. My download is scuffed though, produces static.

@Imdrekaiser set v-prediction and ztsnr on if using v-prediction model. also use built in vae. much simpler to do with ComfyUI than any other UI available.

@fizalpher Thank you so much for the tips, those worked and I am blown away by how good this model is. Also, on reForge it is super easy to change v-prediction and zero SNR (it's in the Advanced Model Sampling tab).

Have you tried Regional Prompter?

comfyui有一个节点叫“attention couple”；或者试试create frame；这些都是我的猜测xD

The vae built into it, just drag vae from load checkpoint to the vae decode on comfy.

use the recommended settings. Euler and cfg 4-5. did you not read the instructions? This is still training, so very important to follow instructions.

not the problem, cg rescale needs to be at 0.7 and your SCHEDULER needs to be on normal. I gen perfect pics with dpm++ 2s ancestral and normal scheduler.

yes, use newest forge. just update it

简介里的NoobAI-XL Short Test Report里有不少测试用例

@realhjt 參考不了，底層邏輯都不一樣了，那是早期測試的版本，放在v版根本沒用

很中肯的建议，感谢反馈，后续版本会优化。

团队成员大多是兼职开发模型，有主业在身，自愿投入时间、精力和金钱，而且零回报，还望谅解。

如果有安装、使用等等问题，也可以加QQ群或者DC群，随时提问~

Laxhar/noobai-XL-Vpred-0.5 ·拥抱脸

@Euge_  Such a generous response is admirable. Hope you enjoy creating and sharing.

@Euge_ 高情商

wtf?

Are you using the recommended settings from the "About this version" section? Euler (not Euler a), 28-35 steps, CFG 4-5

In fact all correctly setup stable diffusion checkpoint will have a baked-in VAE, whenever you have doubts, try the baked-in one first. If you are not satisfied with it, then maybe try other VAEs.

Training or running Flux isn't easy. A non-distilled model is much easier to train. The best option is they decide on SD 3.5 Medium which is close to SDXL but has an improved architecture, or wait until a better model with 4 billion or 5 billion parameters comes along.
A model with 12 billion parameters is simply too big to conveniently train and run.

v_pred is now supported in main branch of ReForge, which is better than vanilla A1111.

@mewtsy Is it compatible with A111 plugins?

@burnera679889 I don't know about its compatibility with all plugins, but it uses same plugin repo as A1111. It's basically same webui as A1111, since it is a fork of A1111, except it has performance optimizations under the hood and it is actively being developed.

@burnera679889 I clearly typed: that ReForge is "actively being developed", I don't see how that implies that A1111 is no longer being developed, nor did I say that. I am not the information source depot for A1111, simply telling you what works for v_pred at the moment. ReForge main auto recognizes models that use v_pred, and it's a good overall webui due to it having optimizations not default in vanilla A1111.

No it isn't I have been using it just fine. You need to READ and follow the usage guides for this model https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5?not-for-all-audiences=true#please-read-the-guide-carefully

Especially this is the part you are looking for:
Parameters

CFG: 4 ~ 5

Steps: 28 ~ 35

Sampling Method: Euler (⚠️ Other samplers will not work properly)

Resolution: Total area around 1024x1024. Best to choose from: 768x1344, 832x1216, 896x1152, 1024x1024, 1152x896, 1216x832, 1344x768

@mewtsy ok, already checked

@Y_X Well, I am telling you, it's not broken. Temperamental, yes, but that's mainly because this iteration relies heavily on danbooru and e621 tags. I am using A1111 ReForge with the recommended settings and it's working great for me. The OP listed artist and character tags on their Offficial Discord, but you can also just open Danbooru and search for a concept then use those tags. I think that the next or final release will be easier to use, but as of right now, ComfyUI and ReForge (main) are the easiest ways to use this model. Also, ReForge needs to be updated, the automatic v_pred recognition on model load was a recent addition (some time earlier today I think), so you may just need to update.

@mewtsy even following instructions with dev reforge with advanced generation options enabled, it gives really unstable outputs. It's genuinely just a bit too bleeding edge.

@chuunikaede I've been posting images on this models gallery including metadata, using ReForge (main) on a 3060ti, did not have to set anything, v_pred model was automatically recognized on load, I know because I checked the terminal. The images look fine to me.

@mewtsy when did you update your reforge?, basically a few days ago the dev removed/added new branches,added/removed commits and everything. That may be changing the reforge in others

@quamainedaxten713 I'm on ReForge Main branch. I updated it yesterday night when I saw the notice on NoobXL Discord. I did another git pull about 3 hours ago, no changes.

Version: f1.0.5-v1.10.1RC-latest-763-g91cb30c1

Commit hash: 91cb30c177a3179bb3846a47ede2701649634380

@mewtsy I updated reforge to dev as of yesterday morning, generation only works with advanced generation settings enabled and even then is still unstable, even with models trained on vpred. I'm gennjng on a 3080ti, with correct settings.

@chuunikaede Well, I'm not using Forge Dev. I am using ReForge Main. It recognizes the prediction type on load automatically and has no issues as long as Euler sampler is used, as is stated in their recommended usage guide. If you check the About this Model section, it even recommends ReForge and ComfyUI. It's a bit outdated though, because it states ReForge Dev, but that's because yesterday is when ReForge Main was updated ahead of the Dev branch. So, ReForge Main, is the way to go for now if you don't want to mess with ComfyUI.

Version: f1.0.5-v1.10.1RC-latest-763-g91cb30c1

Commit hash: 91cb30c177a3179bb3846a47ede2701649634380

@chuunikaede Here is the github for ReForge if you ever want to try it out: https://github.com/Panchovix/stable-diffusion-webui-reForge

@mewtsy yeah im using that. im sorry i meant to type reforge dev. you can see where my previous posts reflected that. reforge main at the time that i updated did not have compatibility and it required being on the dev branch to be able to gen. there also wasnt autodetection and required enabling advanced generation settings.

100% agreed. Very annoying to even get the model to work. I've seen most success in ComfyUI. It's just very picky with settings at this point. EDIT: Also, you really HAVE to add a style to make any generation look decent. Not necessarily a Lora, just an artist name. Otherwise you will get a western style cartoon look with bad anatomy. Have to follow the recommended settings on their HF too

https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5?not-for-all-audiences=true#please-read-the-guide-carefully

I ran an Illustrious LoRA of mine on noobaiXLNAIXL_epsilonPred10Version.safetensors in Fooocus.
It works, which is ok, I guess.

no

@Euge_ thank you!

不知道为什么我也是,但没人提这个

这是v预测版本，需要在comfyui或者reforge下使用

目前WEBUI的主分支不能使用VPRED版本的模型

hi，可以跟随 “About this version” 的引导使用~

该版本是v预测模型，需要一些特殊操作才能用。

我们稍后会更新更全面的指示。

@zeseren 好的，蟹蟹

@BakA164 好的好的，蟹蟹

@Euge_ok 

@lemonjing 我看评论有人解释了

中文版使用指南请 点此此处。

@Euge_ 你好，git怎么切换分支呢，就是找不到QAQ

@chenjt26 非常建议 在切换分支之前备份原本的 webui！
方法：在 webui 的安装目录下打开控制台，然后输入 "git checkout dev"，就可以了。

ui \(blue archive\)

Problems with character might be a matter of undertrained tags, low post-count characters, too many characters on the scene or even artist styles or concepts making the fidelity harder

You may refer to the "trigger" column of this csv table file:
https://huggingface.co/datasets/Laxhar/noob-wiki/blob/main/danbooru_character_webui.csv

If you'd like to generate a character, just copy and paste the "trigger" of that character.

If the fidelity is not satisfactory, then just add the "core_tags" of that character.

Epsilon

- Easy to use;

- Larger community base;

- Defects in color, lighting and shadow, etc.

V

- Full color gamut;

- Higher dataset fidelity;

- Better prompt following;

- Higher ceiling;

- Sensitive to generation parameters.

@Daida add me on discord "mimizukari" i can send you the char wildcard. i have both full sheet & one trimmed down to 75 tags (so the ones that should work, 18.5k/200k+). used a python script to turn these into wildcards day 1, Lol. (only Danbooru ones though, i don't particularly care for e621... same script would probably work for e621 but...)

@fizalpher can you post it on civitai? it would be very helpfull, or you can send a link to download
maybe from anonfiles or catbox