Using the RouWei model as the base model, because it has fewer watermarks and a relatively newer dataset.
E-pred 01
BASE RouWei 0.7 E
Recommended settings:
Steps: 15-30
CFG scale: 5-7
Sampler: Euler a
The VAE is already integrated, please do not ask such questions anymore.
Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.35~0.5
Positive Prompt
masterpiece,best quality,amazing quality,Negative Prompt
bad quality,worst quality,worst detail,sketch,V-pred 01
BASE RouWei 0.7 V
Recommended settings:
Steps: 15-30
CFG scale: 3-5
Sampler: Euler a
The VAE is already integrated, please do not ask such questions anymore.
Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.35~0.5
Positive Prompt
masterpiece,best quality,amazing quality,Negative Prompt
bad quality,worst quality,worst detail,sketch,Used with "by " it's mandatory
"by " is used as meta-token for styles to avoid mixing/misinterpret with tags/characters of similar or close name. This allows to have a better results for styles and at the same time avoid random style fluctuation that you may observe in some other checkpoints.
Multiple give very interesting results, can be controlled with prompt weights.
Characters:
Use full name booru tag and proper formatting, like karin_(blue_archive) -> karin \(blue_archive\), use skin tags for better reproducing, like karin \(bunny\) \(blue archive\)
Description
FAQ
Comments (65)
e-pred? epsilon ? so standard Illustrious, none v-pred?
e-pred is just another way of saying epsilon prediction, so yeah
@sansmia then its basically normal model not worth mentioning that its e-pred? V-pred works different so it should be specified because it does not work on some UI and it has the Advance Model Samling extensions and comfy nodes
@elguachiiii I'd rather my models be specific about it. It's not as if epsilon and vpred are the only noise models, just the main ones.
作者大大 請問有些標籤比如像1234沒有後面1234 \(xxxxx\)的\(xxxxx\),是不是只要輸入1234就好呢? 表達方式可能有點奇怪 請多包涵
So good that RouWei Starts to get more recognition
i prefer normal than v-pred, v-pred are to finicky and need really low cfg and tweaking or everything looks over saturated and not to mention it does not work on a lot of UI's
For the people that have tried it, would you say it's better than WAI v12?
@PIQ100 From my testing, I wonder if the standard WAI v12 is already merged with RouWei as the images I'm getting from both models are very similar.
@QH96 The same thing happens to me, the changes are noticeable when you go beyond the usual tags, for example with monochrome, and other drawing styles.
@PIQ100 That makes sense. Do you have any recommendations for characters, topics, or poses that the standard model struggles with but the new one handles well? I’d like to compare the difference. Much appreciated.
Since Rouwei itself is also trained on the basis of ILL0.1, and this model lacks any guidance, it is difficult to use directly. Therefore, I have modified it using some of my own model's data. As a result, there is a chance that similar compositions may appear.
So, it is an NSFW branch model rather than a completely new model.
i glazed a bit and still think it's really good, especially if you wanna prompt anything recent and are fixated on getting fingies right without using an editor too much, but it does seem a bit desaturated for a vpred to me even with the recommended color fixes for that (+hdr, -sdr, and then some), the wai merge helped somewhat, not tried janku version
what the difference between v-pred and the standard one ?
legend, not getting the contrast issues i had with base rouwei, minimal wai-gloss too, still feels very creative, all the important rouwei benefits too, can prompt some more recent characters/artists with better success without a lora to throw other parts of the image off
我试了一下把这个模型和WAI NSFW模型融合，工具是aki整合包里自带的。结果出来的VAE总是不对。要么是没有VAE要么是鬼图。请问融合时VAE该怎么设置？
vae可以额外加得啊 融合那边不就是有嵌入vae得选项吗
@WAI0731 嗯。谢谢。我选了嵌入任何一个VAE，出图都是鬼图。于是我选不要嵌入，出图就是没有VAE一片苍白，再外挂任何VAE，又变成鬼图了。我在想是不是IL模型和NAI的模型不能融合。我之前融合过1.5的，是正常的。就是IL模型出问题。
Regarding the statement that "by" is mandatory, should I include it in the prompt even when using a LoRA to adjust the art style?
This calls the artists from the checkpoint. For LoRA, please use it according to the provided prompts.
for someone who already tested eps and vpred, which version perfoms better?
From my experience, the v-pred and eps-pred seem to perform very similarly excluding the regular difference in color ranges. I personally prefer the v-pred for most generation due to the more vibrant coloring, but the epsilon model handles some of the NAI controlnets a bit better for me but YMMV.
Maybe it's a placebo effect, but I think this model is better at fine details than wai12. Only tried the E version. despite the fact that it says that a newer dataset is used. Still can't resolve characters from metaphor refantazio, e.g. Hulkenberg. I know you can use Lora, but I always want it to be already in the model. Hopefully future versions will have more characters.
To make a checkpoint remember a character, it usually takes several hundred images. Not every new character can be memorized.
Personally, I think this is a great model that represents a significant step forward. Especially, the recognize of the prompts and the great color of the output image. Less blurry, high quality image.
as an intrepid RouWei evangelist, I'll give this one a go (vpred) and see if I can replace RoiWei with this merge. Have had a good experience with WAI-NSFW, used standalone or as a refiner. Thank you for your hard work!
Edit: looks like there's little reason to use Vpred in this case, especially since this merge does not appear to support RouWei's brighness/gamma control prompting. Epred it is then!
How was your experience with this vs the WAI-NSFW ? In terms of most relevant aspects: prompt understanding, aesthetics, tendency of anatomy errors/slop etc.
@Wojack34 It's fine but I ended up going back to vanilla WAI-NSFW & RouWei. Each have their strengths and I'd rather just use them separately. As stated earlier, losing brightness/gamma control is what makes RouWei special and without that, there's very little point (for me specifically) to use the merge. I still have it in rotation when doing commission work, just to see how it handles a seed I generate with, for example coco or WAI-NSFW.
playing around with latentmodifiers integrated (in forge) helps a bit with this model, not really tried every setting yet, but looking like you can combat the desaturation quite a bit
Currently, SHUFFLE still looks better imho.
Though, since this one is trained on a different "branch," maybe I just haven't found a consistent quality tags for it...
請問這個跟WAI-NSFW-illustrious-SDXL Ｖ12差在哪？ 哪個比較好？
Would be great if you include the brightness control from the original RouWei in future updates.
WAI-Nsfw-Branch-Rouwei and WAI-NSFW-illustrious are two of the best models i have ever used to this date. much prefer the artistic style, lighting and presentation of these models compared to the pony models
feel like epred give more 2d look and vpred show more cartoonish look
This is my favorite so far
Ok wow i am speechless how good this is with its 2d look and prompt flexibility regarding characters. For indoors scenes this is my absolute new favorite WAI model right now. It has some really serious background issues tho.
Top Notch
plis update, rouwei has a new version btw
Still waiting...
Generation never starts. In my environment
Now that both versions of 0.8 are out i can only hope for an update here soon.
Nice merge as always from you WAI, but we need another version of this model because the latest version of Rouwei has more recent dataset and it will be cool
This is so much better than the main illust branch
Best Perturbed Attention Guidance settings for this model? 🦄
When I use SAM to fix the character's fingers, faces and body are drawn on the hands.
When I use USDU, faces are added to unintended places and the vagina is replaced with a mouth.
really like ths model, been using it since it realeased, but need update
從這裡發現了ROUWEI 0.8, 新版本真的很棒...
但是用慣了WAI V14,還是偏愛你的微調版本 (我已經變成了WAI的形狀,EPS VPRED ROUWEI都是WAI)
跪求更新,我什麼都不願意做的(噓)
Please make a RouWei 0.8 version. Your work is great.
Please update this with the latest RouWei dataset!
Way better than most others I've used, pity it is no longer available
Could you give me an example of how to use “by” in prompt?
by [artist name]. It's for artists.
@deitychaser I see, thank you!
hope to see the merged version of RouWei 0.8, which contains more character data for Gakuen Idolmaster
still waiting for update, this is the only model i use rn, cuz its too good
Dear OP,
You should really consider upgrading this model, its way ahead of WAI v15, 14. If you could update it would mean a lot to everyone who's begging here. I don't know why everyone is stuck on illustrious v1.0 as base model. Rouwei is such a big improvement
hey bro,I'm new here.I want to konw which do you think is more user-friendly right now, "Rouwei" or "v16"
the people yearn for another update for this model...
Amazing checkpoint. Any plans to update it?
This checkpoint is truly amazing. Without wanting to repeat what many people already stated, an update of this model to the latest Rouwei version would be deeply appreciated!
In any case: Great work so far, please keep it up!
Any update? 🥺
Love this model..
We need a new version