If you want to use more my checkpoint online generation, please visit here.
https://tensor.art/u/762555264535746522
V-pred-04
Data balancing and adjustment have been made, and some background logic has been optimized
Recommended settings:
Steps: 30
CFG scale: 5-7
Sampler: Euler a
Positive Prompt
,masterpiece,best quality,newest,absurdres,highres,very awa, Negative Prompt
low quality,worst quality,normal quality,text,jpeg artifacts,bad anatomy,old,early,copyright name,watermark,artist name,signature,V-pred-02
The stability of the limbs, with the overall color leaning towards a warmer tone.
v-prediction versions are experimental models.
You need to use the webui that support v-prediction.
ComfyUI
reForge
Forge
AUTOMATIC1111 (dev branch)
Use Hires. fix (The situations where ADetailer is needed are not many)
Style and Character reference NOOB
All example images are generated at 1024x1360,and Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.5.
Recommended settings:
Steps: 30
CFG scale: 5-7
Sampler: Euler a
Positive Prompt
,masterpiece,best quality,newest,absurdres,highres,very awa, Negative Prompt
low quality,worst quality,normal quality,text,signature,jpeg artifacts,bad anatomy,old,early,copyright name,watermark,artist name,signature,V-pred-01
BASE:NOOB V-pred 1.0
This is a test model, attempting to use some unconventional methods to achieve merging.
The merging materials used are not purely from the v-prediction model, so there may be issues that I'm unaware of. If there are any, please leave me a message.
v-prediction versions are experimental models.
You need to use the webui that support v-prediction.
ComfyUI
reForge
Forge
AUTOMATIC1111 (dev branch)
Use Hires. fix (The situations where ADetailer is needed are not many)
Style and Character reference NOOB
All example images are generated at 1024x1360,and Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.5.
Recommended settings:
Steps: 30
CFG scale: 7
Sampler: Euler a
Positive Prompt
,masterpiece,best quality,newest,absurdres,highres,very awa, Negative Prompt
low quality,worst quality,normal quality,text,signature,jpeg artifacts,bad anatomy,old,early,copyright name,watermark,artist name,signature,V2
BASE:NOOB eps 1.1
Try to ensure image quality without using quality keywords, in order to create a model that can be easily used by beginners.
Not very necessary to use Positive and Negative quality Prompt
Use Hires. fix (The situations where ADetailer is needed are not many)
Style and Character reference NOOB
All example images are generated at 1024x1360,and Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.5.
NO Positive and Negative quality TAG
Recommended settings:
Steps: 30
CFG scale: 5.5
Sampler: Euler a
v1
Recommended settings:
Steps: 28-35
CFG scale: 5-7
Sampler: Euler a
Positive Prompt
,masterpiece,best quality,newest,Negative Prompt
low quality,worst quality,normal quality,text,signature,jpeg artifacts,bad anatomy,old,early,copyright name,watermark,artist name,signature,Description
FAQ
Comments (41)
WAI, which models of yours do YOU enjoy the most? You have quite a lot of versions now!
im so happy to see updates in this one
I have reForge installed using the stability matrix, but I get a black screen.
Finally an update to my favorite model!
性能太强了，3和2比完全不是一个东西了，能生成很好的背景，手也比2好
Is there any reason for the tag "signature" to appear twice in the negative prompt, or is it just an oversight ?
請問negative prompt中出現2次signature是單純多打還是有其必要 ?
just Input miss
fucking finally
OH WOW IT'S FINALLY HERE!
Thanks so much for the new version! It'd be great if you could update the dataset sometime in the future.
I'm kinda surprised,i tested and until now i think this is more stable than wai illu in prompt control,and quality of eyes
EDIT: I fixed the issue, just had to redownload. Loving the results so far!
(old comment) Pardon the probably silly question, but I can't get this to load in v-pred mode in reForge despite all of the right parameters and setup (low cfg, special schedulers etc). Thus, I get noise. It seems like it needs a config yaml file? But idk how to go about making that. Can anyone help? My reForge is up to date afaik.
Nvm guys I got it working!! 🥰
how you fixed?
@Ibukinho I had to redownload the model completely from CivitAI, for whatever reason my original download just wasn't set up right or something. I don't really understand it myself lol
Forge跑V-pred-03，同样的参数有时候可以正常出图，有时候会出鬼图，是什么原因呢？
It generates very saturated images for me. This only happens with this model. How can I fix it??????????????
Try lowering CFG
this is vpred model, tweak your configuration accordingly
You can also rescale CFG to 0.7 I find that fixes most blowouts for me.
For those who are having problems,i got a solution for me,euler a+ allign your steps CUSTOM,idk why but alling your steps normal isn't working
my align steps is normal, not custom, how do i get custom?
@MokkoRino Schedule type
I know a model that do needs another version: WAI-CheapFast-ANI-NoobAI.
Perhaps making another version of that model will eventually make it again elected to generate images in CivitAI.
Models are no longer "selected" by Civit staff. You have to pay for it to have it on the generator. Mostly a minimum of 10.000 yellow buzz per WEEK. So if its not making more than that, the creator is losing buzz (money)
用这个模型配合regional prompter画多人图时，如果开了高清修复就会在放大过程出现tag互相污染，不知道是sd webui的问题还是v模型不是很适配regional prompte？
This is a crazy good model. Not even "designed" for furry stuff but it does everything I throw at it so well and it's very responsive to all sorts of loras.
I look forward to see its further development.
Is there any difference in the output between this and WAI-Illustrious? What's better for just regular anime?
Not sure what I did to deserve this early Christmas present but thank you! Been playing with v4 and it's been giving me solid results as all models in this line tend to do.
Every time you update it, I will reiterate how much I love this model.
Any trick to make eyes better? Somehow is the only thing I'm missing
The eye thing seems to be setup dependent, I don't see it happening after upgrading to v4.
Here is my setup, maybe it will help: 30 steps Euler A, hi-res fix 1.3 with 0.5 denoise, standard NAI dimensions.
Two helper loras in the negative prompt, they fix many NAI problems like wonky contrast (pay attention to the values): <lora:NOOB_vp1_detailer_by_volnovik_v1:1>, <lora:NOOBVP1_stoopid_colorfix_modified:-1>
For detailer it's [v-pred1_v1]: https://civitai.com/models/987736/noobai-xl-detailer
For colorfix it's [v.1.0]: https://civitai.com/models/1555532?modelVersionId=1760239
ADetailer makes the eyes a lot better, but even without it I don't see any degradation from v3. I mostly generate with western artists. Hope this fixes it for you.
@Noise_Connoisseur I think it had to do more with the data quality, as for example I was testing this with Nekomusume, and she might not have enough data on the eyes as official art data might be limited. I tested this later on Nakahara Komugi and Konata Izumi and the eye quality actually was good
FaceDetailer can help you fix a eyes
Eyes looks better on v3...on v4 looks kinda blurry
Are you gonna make a Z Image finetune on this when it drops?
With or without the V-PRED model, the quality of the LoRA dataset used for merging is the best in my setting.
But the character often sweats
yes so many sweatdrops