Works Perfectly With Most NSFW LoRas
speaking of which.. Checkout my NEW LoRa for old webcam style photography (18+):
https://civarchive.com/models/267954/nostalgicam-18-nude-webcam-girls
Very Real (;
Lazy Mix Plus is a way more NSFW-extensive variation of my photorealistic merge, LazyMix. It has some extra ingredients added to the merge, most significantly, the amazing SubReddit V3 model. Please refer to the model page for trigger words, although they are not necessary to get the model's benefits, they are still optional and can be used.
This model NO LONGER includes DreamLike PhotoReal v2 as of LazyPlus v4.
Note: For optimal results, it's recommended to use this model with a VAE. While it can still function without one, incorporating a VAE will significantly enhance its performance.
VAE IS NOW BAKED IN UNLESS YOU DOWNLOAD THE NV PICKLE FILE
Read: Due to this seed breaking change [ 2023-02-19 - deterministic DPM++ SDE across different batch sizes],
The example images cannot be reproduced unless you toggle the: Do not make DPM++ SDE deterministic across different batch sizes setting, found in the Compatibility tab of A1111's settings.
Description
Updated Lazy Base to v3.0 (Replaced DreamLike with A-zovya, and updated Realistic Vision)
Reworked the entire mix, it still shares a bit from v2 and v3, (also removed urpm in place of pmp)
Updated Subreddit to v7
VAE is now baked in
FAQ
Comments (22)
Glad 4.0 is the Highlight.
What do you mean by highlight?
So whats changed?
Click "about version" on the right side, lists it out in bullet points.
@TheUnpossibleDream I think OP made a typo and meant to say "Updated Lazy Base to v4.0" instead of "Updated Lazy Base to v3.0", confusing some peoples.
(typo since the rest of the description looks okay, "shares a bit from v2 and v3")
@Malessar "Updated Lazy Base to v3.0 (Replaced DreamLike with A-zovya, and updated Realistic Vision)
Reworked the entire mix, it still shares a bit from v2 and v3, (also removed urpm in place of pmp)
Updated Subreddit to v7
VAE is now baked in"
@Iuot no typo. The lazy base is a private model mix that is used as the base for both my Lazy mix and Lazy plus models.
Shares Abit from V2 and V3 refers to lazy plus as the public mix and it's versions.
@kaylazy Ooooh, then I got confused with the names! haha
LazyMix+ got updated to v4 and use the updated Lazy Base v3, got it !
@kaylazy about this -> (also removed urpm in place of pmp) . was urpm checkpoint used in merge or the urpm lora? does that mean you don't need to use urpm LoRa with LazyMix+ ? And what is pmp?
@toshiro_mifune excellent question. Before 4.0 the lazy+ mix starts with a 50/50 merge of my base mix and urpm (the model). So the lora is still optional in all cases but imo it's not gonna be very useful in the most general cases.
And PMP is p*** master pro. A newish model on civitai that I decided to experiment with by replacing urpm. So it does make up a large portion of this mix's "anatomy".
Generally the rule was if it works well with urpm it will work well with lazy+ since they share a lot. The same should still apply as I think PMP and URPM are similar enough to not have any shattering changes.
Thanks for asking, hope I answered your question 😊
Is it able to train LoRA? Subreddit v7 had this problem and it was merged.
most likely not.
I did merges with v7 myself and I never got it to work
@firedelta5 PMP has fixed that issue but idk how the author fixed it. I can always make a version without subreddit if people want.
@kaylazy If you remove it then the model is not the same anymore I guess. I will stick to v3a as all my LoRA were created with this and in my opinion this is the best trainable version.
Maybe you can try to merge with a working version of Subreddit
@Lucifie oh the issue is with v7. I understand now. Yea i can revert back to v6
@kaylazy yeah the issue is v7 only
Was curious about this so i checked my copies of subreddit v4, 5 and 6 and they all use the same clip and vae components. (0000/0000 according to model toolkit, which matches vanilla 1.5 checkpoint.)
I was able to repair v7 by creating a new SD-v1 structure and importing the unet from v7 and clip and vae components from v6. Not sure it would fix the lora training issue but at least model toolkit no longer reports it as broken.
@jmblsmit129 "(0000/0000 according to model toolkit, which matches vanilla 1.5 checkpoint.)" .. hmm everything before v7 was trained on SD 1.4, v7 is the only one trained on SD1.5. Perhaps the average person is using 1.5 in the add difference when merging subreddit into their mix. I know to use SD1.4 and then 1.5 for v7.
Tryed to train with v4 yesterday for a few % to see if it works and it did not fail. Will do a full training test today to compare it with v3.
@Lucifie interesting. I never got it to work with subreddit v7 merges :D
@firedelta5 already trained a LoRA on this model and saw no error.
Sigh I'll come back tomorrow and a half :'(
Details
Files
lazymixRealAmateur_v40.safetensors
Mirrors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40_1.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.ckpt
lazymixRealAmateur_v40.safetensors
LazyMix_Real_Amateur_Nudes_v4.0.safetensors
lazymixRealAmateur_v40.safetensors
lazymixRealAmateur_v40.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.