noob V3 version released
This is an EPS model with nothing special about it. It just has improved low-light depiction.
Sampling Step 10 ~ 25, CFG scale 1, Euler a, beta
At the positive prompt: masterpiece, best quality, amazing quality, very aesthetic, absurdres
At the negative prompt: none
v-pred model release
Nothing special, just a regular v-pred model. Remember that it is only available in tools that support v-pred.
Sampling Step 10 ~ 25, CFG scale 1, Euler a, beta
At the positive prompt: masterpiece, best quality, amazing quality, very aesthetic, absurdres
At the negative prompt: none
noob V2 version released
This model is a slightly improved version with better lighting. The usage is the same as the previous model, but the effect of the 'dynamic lighting', 'night theme' prompt has been enhanced.
Sampling Step 10 ~ 20, CFG scale 1.5 ~ 3, Euler a, SGM Uniform
At the positive prompt: masterpiece, best quality, amazing quality, very aesthetic, absurdres
At the negative prompt: lowres, low quality, worst quality, bad quality, bad anatomy, sketch, jpeg artifacts, old, oldest, censored, bar censor, copyright name, dialogue, text, speech bubble, Dialogue box, error, fewer, missing, watermark, signature, extra digits, username, scan, abstract, multiple views, censored, logo, bad hands, mutated hand
ilxl1.0+noob version released,
10 minutes ago have merged VAE. Gentlemen who don't have VAE or are having trouble choosing one, don't worry anymore!
To be honest, the data set of ilxl1.0 lags behind that of noob1.1.
But I mixed it up a bit because it helps make the base resolution bigger.
So you can use it just like the existing noob1.1 version.
This model is much closer to noob1.1 than ilxl1.0, but the recognition rate of the dataset and e621-style prompts is damaged, so it is simply named ilxl1.0+noob.
Sampling Step 10 ~ 20, CFG scale 1.5 ~ 3, Euler a, SGM Uniform
At the positive prompt: masterpiece, very awa, best quality, amazing quality, very aesthetic, absurdres, newest, intricate details
At the negative prompt: lowres, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, old, oldest, censored, bar_censor, copyright_name, (dialogue), text, speech bubble, Dialogue box, error, fewer, extra, missing, worst quality, low quality, watermark, signature, extra digits, username, scan, abstract, multiple views, (censored),worst quality, low quality, logo, bad hands, mutated hands,
v1 3D CG released,
It's based on Noob 1.1
This model has built-in DMD.
Since I don't use regular models without DMD, you should also generate images with low CFG and low steps.
I have a condition where I get a heart attack if it takes more than 10 seconds to generate a low-resolution image, so I had no choice.
The reason why it has a different style from the existing PONY model is because I intentionally created it with a different direction, as I believe the PONY model has its own advantages.
However, since my taste hasn't changed, there are many similarities.
Sampling Step 10 ~ 20, CFG scale 1.5 ~ 3, Euler a, SGM Uniform
At the positive prompt: masterpiece, very awa, best quality, amazing quality, very aesthetic, absurdres, newest, intricate details
At the negative prompt: lowres, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, old, oldest, censored, bar_censor, copyright_name, (dialogue), text, speech bubble, Dialogue box, error, fewer, extra, missing, worst quality, low quality, watermark, signature, extra digits, username, scan, abstract, multiple views, (censored),worst quality, low quality, logo, bad hands, mutated hands,
PONY version released,
I'm a lazy person, so I can't take the action of changing the settings to suit the model every time I change the model, so I also made PVC and anime-style models to use in the same settings as the basic V3 model.
At the positive prompt, score_9, score_8_up, score_7_up,
At the negative prompt score_4, score_3, score_2, score_1
For the toon version,
it is recommended to add the following prompt: toon \(style\)
Sampling Step 20, CFG scale 4, Euler a SGM Uniform
I prefer this setting not because the picture quality is good, but because the picture comes out quickly.
Description
FAQ
Comments (38)
any chance you can make it so the next vpred version doesn't look "fried" too early? At 20 it looks way too saturated, and while at 10 it looks good, it doesn't follow the prompts completely since the steps are too low.
I recommend using DMD with minus weights. Something like -0.3 would be good for 20 steps.
vpred models look that way when you use Euler A/Karras. Use Euler/Normal and it won't look fried anymore.
@meden What is DMD and how do I use it? It it a lora?
@slaad0 DMD accelerates image generation and is usually shared in the form of LORA. Various DMDs are available here, so use them as you like.
Not working on Invoke sadly :(
Can this model be used commercially? As per the license it seems like it.
I basically don't care how other people use my models, but I would like to point out that my models are derived from noobai models, so noobai's license will apply regardless of my will.
The model is really fast and looks great, but I dont get to run the v-pred model with the beta schedule. It works fine using SGM uniform, but I was wondering if I am missing something in the generation settings.
This is the best model I've ever used. I LOVE the noob 3d v2!!!!
I think this model is better with DPM2++
Try it with uni_pc_bh2 and sgm_uniform at 1 CFG and 10 steps.
I have an issue: is the model trained only with females? I've tried to generate some characters (not known, for information), but they rule 63'd them.
I was gonna ask about this too. I tried to generate a man (not a specific character, no LoRA) but kept getting girls/femboys, even when I emphasized gender tags and my prompt had "stubble" in it.
Edit: I upped the emphasis some more to get what I expected
It's a very good model, although it may be difficult for those of us who use Forge to understand how to use it, so I'll leave a short guide based on my tests:
- To use it in Forge in the UI, you must select the "All" section.
- Clip Skip: 2 - Sampling Method: Euler A
- The model indicates that we should set the scheduler to beta, but at least in Forge, it doesn't work, so based on my tests, the best one is "Align Your Steps GITS."
- The size I assume is used in the reference images is 1024x1360 with a 2x upscaling.
- It's not necessary to use Adetailer.
- For the High Res fix, I used the "Remacri" upscaler with 30 hi-res steps, a denoising strength of 0.35, a 2x upscale, and only a scale between 1 and 2 in Hires CFG.
I hope this guide helps. Greetings!
Thanks!!!
Yes, I also tried scheduler type beta and it produces nonsense. Any idea why it doesn't work?
PseudoGrafx You need to make the same changes as the latest commit from the dev branch of A1111(2174ce5afea90ca489d222f539988dcef59f1027), then it should work with Forge(the line number will be different for some of the changes). I found that the Beta Schedule Type still produces the best results.
I'm using it with very old Forge (pre Flux) and it's working fine. Did some of my own experimenting and it works very well. I only have problems with backgrounds, but it may be problem with prompting for Illu in my wildcards (Pony optimized) and with ADetailer - the inpainting boxes are very visible.
你好，我是你的粉丝。我怎样才能训练角色Lora以与你的模型兼容？
Ohhh while this model is truly amazing, what I'd really want is this V-Pred model but in a more anime/toon style. It's SO fast and such high quality, but very hard to create something non realistic
This model may not work well with some loras, and may have a style limitation, but it's still the BEST model I've tried. A lot of details, a lot of prompta, fast and uniquely high quality
when i use this, for some reason when using some of the character loras, the face is somewhat has this white or not even skin tone with the rest of the skin only the face is white
the face is very similar to this, https://civitai.com/images/83514566 it has this white tone on the face only which is i don't know how to fix when using other loras
is there a way to avoid same-face-syndrome? All my different anime related generations turn out to be the same character just different clothes+hair, even with lora.
While not a fundamental solution, you can use artist prompts with lower weights to give your character's face some variation and personality.
use combine lora style. . Shexyo. Jizoku, Avanda and others.
Hi, I'm using v-pred model. Do You have some tips for using it with ADetailer? My usual settings are giving highly visible boxes of the ADetailer and different tone skin.
I just use it with the denoising value lowered to 0.1
@meden I decided to not use AD at all. After min-maxing to get rid of those borders it's giving a little better or a lot worse results than pure generated images. I rather correct just eyes or other small elements by inpainting manually than adding new errors that I will have to also inpaint.
What would be the reason this model would stop working with beta schedule? I generated a number of images successfully with beta recently, but now everything is heavily grainy. I've tried with the same prompt and settings that were working a few days ago.
I can't believe that,The best result is not use lora.The result will be worse when I use character's lora.
reducing the influence of lora can improve the result
@Bottomless_Master Thx bro
You should believe it. It's like that with nearly everything. Always start without LoRa and then only use as much strength as is needed to get the intended result if you have to add one. They reduce quality very quickly. Also make sure that the LoRa you're using is for the same base model as the checkpoint you're using for generation.
Is Vpred-V1 merged with a DMD lora or something? If so, can you upload a version that does not?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.