V4 Notes:
Hey, my coat's lapels are black with gold trim like they're supposed to be, how about that?
This is still Noob EPS derived, this will join v3 in the generator for a week and if feedback is good, i'll try to keep it in rotation alongside V3.
Cheers
======================================================
A Noob variant of my MadlyMix using the following recipe:
@bloodsplash 's SpatterXL2DV2
My own style lora at .5 weight
NoobAI-XL Epsilon-Pred 1.0 at .3 weight
and Night XL by Ichxi at .5 weight to bring the saturation into line.
Essentials:
Positive Prompt:
masterpiece, best qualityNegative Prompt:
lowres, worst quality, low quality, bad anatomy, bad hands, 4koma, comic, greyscale, censored, jpeg artifactsAs with Madly Mix, this should give you a healthy dose of clean 2d and sometimes 2.5D images with a noticeable contrast ratio and sharp lines. It should take styles reasonably well. I haven't tested 3D or realistic images intentionally, but if you're into that, by all means.
I recommend making use of Bloodsplash's Spatter ckpt which served as the foundation for this mix.
Cheers!
Description
Updated version of Nightly Noob V2, tweaked to try and fix some of the lighting issues.
FAQ
Comments (39)
My favorite IL checkpoint by far. Its default look is amazing. Will try to new version and see if it is better than v2 ;)
Civitai should definitely add this model to the generator!
it'd be nice. unfortunately it doesn't seem likely, it's why I have it up on Tensor for now.
@MadlyLaughing I know... just wishful thinking, this model seems quite unique
as it happens, v3 is now available to run on the generator.
@MadlyLaughing Amazing!
He keeps posting Bangers. Someone has to stop him
:Clueful:
Just trying for the first time, damn is so good, loved the colors on this one, will post some stuff soon ;)
Is it possible to upload your work and models on the site Shakker.ai
I'll look into it.
I attempted to set something up, but not only does Shakker already have someone who uploaded one of my models, they wouldn't upload the new version, and wouldn't remove the old one when I asked them to. So you'll have to make due using it on Tensor, or locally unfortunately.
not an issue now, you can run v3 in the onsite civitai gen
You can now run this CKPT on Civitai's generator.
wonderful and stable
top checkpoint in ILL
W checkpoint, i keep being a regular.
Great work! Very artistic! Looking forward to seeing future versions
Really like this style!!! It would be awesome if you can train better hands in your next updates
I haven't had any real issues with hands aside from the occasional one-off.
@MadlyLaughing even some of your example images have hand issues wdym. People criticizing hands have some issues doesn't mean all hands are completely and utterly bad. IN COMPARISON hands have issues while in backgrounds,clothes and faces this model is great. You can take constructive criticism without being defensive.
Top tier.
v4 edit: I dont know what you did but v4 is utterly incapable of making NSFW. It especially hates dicks.
This is solid. Only downside i noticed so far is some concept LORA incompability issues here and there
Edit: Its not just LORAs, it has issues with different angles and concepts in general
Edit: Ok after trying to make v3 work consistently with different prompt styles for months i came back to v2. V3 has too many LORA and hand issues.
same
I agree, it won't get "three quarter view" no matter what I do. I also noticed that it struggles a little bit with a few style models, in particular if they're heavily into the anime side. It tries to force a 2.5d style that ends up just looking blurry/noisy.
These issues aside, this is probably the best model I've ever tried lol
@properstyle depends on what styles you're using and how they're cooked has been my experience. I know that if something is particularly burned it probably get a bit grainy. Three quarter view on some noob derived checkpoints (mine included) seems to be a bit of a crapshoot.
@MadlyLaughing That seems to track. There were a few anime style loras I've tried that I've had no problems with, but then there were a few that acted up. In particular, no One Piece anime model I've tried worked lol. Either way, using an extension like loractl allows me to bypass these sorts of issues so it's just mildly annoying.
Goes without saying that this model is amazing either way. I've yet to see a model that is so consistently good at most things, particularly when it comes to character design. Most models only have variety on women, and then there are a few that are male focused, but this one manages to do both really nicely.
By the way, while you can't get photorealistic with this model, it's actually pretty good with oil paintings. If you set it up properly you can make some great oil paintings natively. Just a heads up cause you said you didn't experiment with it.
@Wizzi If you want to put more critiques here, be my guest, but popping a whole new thread up is just extra and unnecessary.
I'll have to look into that, I haven't had too many issues with NSFW in V4 in my testing, but I'll see what's what.
Edit: I posted some of the tests I worked on while checking some basics in a new upload. Dicks seem to work just fine. At this point I want to see your settings, because there's clearly something not working properly. Post some samples with the metadata intact, i'd like to see what settings you're using that you're getting these results.
@MadlyLaughing if you care been playing around and it is the "close-up" prompt. In previous versions that tag was helping get a closer look in bj/paizuri scenes. in v4 it cuts off the dick and sometimes puts a small box of close up at top corner. It is still the same issue with v3, some angles have issues. Not a dick problem.
@Wizzi if I didn't care I wouldn't have gone back to actually see if I could recreate the issue. One thing you may need to use for this model, and I saw it in v3, is you may need to use a subject focus tag when using close ups, that tends to help perspective issues in my experience. I don't know what else you're using, so there could be other factors causing the issues you describe.
@MadlyLaughing nevermind its not close-up. might be from below. or it might simply be when the angle gets too difficult. will upload pictures in a bit so you can check too
idk when it will appear but uploaded. basically if i dont try to make paizuri and bj at the same time there is no issue, but when i try to do both it kinda borks out. it is not always ofc but seen it happen 2 times out of the 6 different prompts i tried to compare with previous versions of yours. and the one i uploaded is me removing prompts to figure out when it stops having trouble. Basically its some sort of anatomy issue, probably can make it work with different prompting but this is still something that happens in comparison to other models.idk if all of them got uploaded. tried to do a mix of v2 with v4 to see if it changes anything but didnt work still. Also checking example images now. i think it might also be lora compatibility issue. the one where i used no lora looks the best.
But ye i spent way too much time on this lol. Hope it is helpful in any way. im out.still love the models style when it works o7
@Wizzi
I'll look at it when they show up in gallery and see what I can figure out if anything.
After looking at the prompt, you've got a couple anomalous tags or tags with low information, and some posing and angle tags that might be conflicting with each other, there are some concepts that need loras at least with my checkpoint. I'm not sure about other check points, you may try using an EPS 1.1 checkpoint, this one's built off 1.0 which doesn't have all the extra e621 training that EPS 1.1 benefits from. Aadditionally the boxes in the corners may be caused by the use of brackets [ ], I found that those caused me problems on several models while I was using them.
The aspect ratio is a little weird, but is likely ultimately a small portion of the issue from what I can tell. This is just my take on it, and as we know, some camera angles are a little dodgy in this line of models. If it worked in V2 you may consider making a merge with V2 and V3 and see if that works for you. V4 is a bit burned in by the way I trained it, so it may not take to being mixed well from what some of my friends have reported.
That's all my take on it just glancing over the samples I saw.
on V3 / beautiful mix, been a bit since a model has really blown me away with its quality.
Why is it unusable all of a sudden?
I'm guessing its the new absolutely ridiculous bid system they have put into place. Such bullshit for them to remove the model unless we pay up... I just put down a bid though. Only reason I bought buzz was to use this model and now ... yeah....
@Mao_of_Saints Yeah, it seems to be the case, I put some buzz into the bidding to bring it back into utilization. I guess it wasn't pushing enough numbers to remain usable anymore by itself.
@MadlyLaughing mannnnn i love this model sm (T_T)