This started as a purely experimental merge, separate from my GonzaLomo models, but has now become my main SDXL source of images. I think the name Mixture of Perverts says it all.
My primary purpose in creating this model is to generate images that I can use for WAN I2V. I wanted something that had a more amateur feel while still being able to produce good color, composition and lighting.
The v5.1 models are among my favorites and offer a wide variety in speed and usage while giving the best photorealistic results.. See the version details for the recommended settings for each of those.
v6.1 and v7.1 are closely related and offer a decent variety of styles and facial types for DMD models.
Most of these versions are DMD models, intended to produce images very rapidly. I recommend using my SDXL DMD workflow or you can choose to use the workflow embedded in the showcase images.
The recommended settings for the DMD models are:
Sampler: LCM
Scheduler: karras, exponential or beta
Steps: 6-14
CFG: 1.0-1.3
Clip Skip: 1-5 (1 doesn't work with some of the newer models)
Fooocus Setup Instructions
1) Click the Advanced checkbox.
2) On the Settings tab, leave the Preset as 'initial' and the Performance as 'Speed'.
3) On the Models tab, set both the Base and Refiner models to MoP.
4) Change the Refiner Switch At setting to anywhere from 0.75 to 0.90. I like 0.85.
5) On the Advanced tab, change the Guidance Scale to 1.
6) Change the Positive ADM Guidance Scaler to 1.
7) Change the Sampler to 'lcm'. You can leave the Scheduler as Karras or Exponential.
8) Set the Forced Overwrite of Sampling Step to 7 to 11. I like 9.
Description
This is now the third version of v5.1 and this is a bit of an experimental oddball. It's a block-level hybrid between v5.1 with its DMD2 textures and v5.1 Non-DMD with its regular SDXL textures. The benefits of this v5.1 Flexible version are:
You can use a wide variety of samplers, schedulers and CFG just like the Non-DMD version.
It only needs as little as 6 steps to generate an image but can go as high as 20.
So basically you get the regular SDXL textures with DMD2 speed.
I strongly recommend using a good refining ComfyUI workflow such as one of mine, or adding the refining option if you use another tool.
Please experiment with whatever settings you like but the settings I use are:
Sampler: dpmpp_sde
Scheduler: exponential
Steps: 8
CFG: 2.0
CFG is a bit more limited than with the non-DMD version. I haven't had good success with any value higher than 3.0 and 1.5-2.5 seems to work best depending on the sampler.
FAQ
Comments (34)
Do you have any free time? :-) THANK YOU for your work, it's always nice to see something new from you. I'll be posting more today.
Thanks for all the great images you post. I really appreciate it! And no, I don't have any free time :)
@GBRX Thank you very much, always a pleasure, it's fun to work with, have a nice weekend and thanks again for the buzz.
@GBRX Thank you very much, always a pleasure, it's fun to work with, have a nice weekend and thanks again for the buzz.
Tried the new one a bit. Its not that flexible:) You can forget LCM and Euler a, other samplers need sgm uniform or other smooth scheduler. SDE is slow, not worth it.
I will test it more. Thanx.
I'm not sure what you mean. So far I get very good results with euler a, dpmpp_2s_a, dpmpp_sde, dpmpp_2m_sde, and ddpm. I normally use karras or exponential. If you want to use LCM then you'll need to apply the DMD lora at about 0.50 strength or just use the DMD checkpoint.
I merged MoPV5.1 with MoP-ILV1 65/35%, very nice... I love how MoP-IL kind of AMPs everything up a bit... and V5.1 rocks! I'll post a couple "re-dos" of your slutty college co-ed!
Awesome, thanks! I really like the Illustrious model. I created it originally by merging MoP with a couple of Illustrious models (Dixar and Perfection Cinematic) so your merge is going to add some more photorealism.
Posting again since I’ve only recently managed to properly set up v5.1 Flexible and started getting great results (using your workflow).
Now I’ve installed v5.1 DMD, but with the same settings I’m getting somewhat blurry, “soft” images.
How should I adjust the settings for Fooocus? Do I need to add any extra node or component?
Thanks a lot for your work! 🙏
I am adding an example, perhaps it will be useful
https://imgpx.com/en/QkUQNgCMf4t9
https://imgpx.com/en/yn9eiiZjUHYv
@acc0nfans921 Hey, thanks for trying out my models. In the model description I have some instructions for setting up Fooocus to work with DMD models. Please try those settings and see if it fixes the problem for you.
@acc0nfans921 Just looked at your ComfyUI screenshot. It looks like your using the Flexible checkpoint but the sampler settings are for the DMD checkpoint. I always put the recommended settings in the "About this version" section. For Flexible you should try the dpmpp samplers and a CFG of between 1.5 and 2.5. For DMD you should use the LCM sampler with a CFG of 1.0. The steps for each of them is about the same, between 6 and 14 is usually good.
@acc0nfans921 You can also try using ddpm, euler A or deis with Flexible. karras, exponential or beta schedulers work best too.
@GBRX Thanks for the quick reply!
I managed to fix the blur issue, but so far the Flexible version still gives me more realistic results.
I’ll keep testing. Really appreciate your work - looking forward to the IMAGE2VIDEO workflow guide! 🤝
@GBRX Ah, that helps me too as I was wondering why Flexible looked so smooth and a little underbaked... Wrong settings!
@GenoMachino @acc0nfans921 Yeah, the three different v5.1 versions with different settings is a little confusing
I am fond of Minimalism and you have ended my search of comparing models with minute details, MoP is all my folder contains now, thanks looking forward to more
Very kind of you to say. I'm glad you're liking my models!
Why do women have to get naked just because it says “breasts” in the prompt? Cleavage or covered niples not helping.... Eny idea?
Well, the name of the checkpoint is Mixture of Perverts so that should give you some idea of how the model is biased. That said, I find using "(covered:1.3) breasts" tends to work about half the time.
@GBRX , I really appreciate you providing these excellent models – they've been great to work with! I wanted to reach out about a couple of challenges I'm facing:I've been trying to generate BJ images, but the closest I can get is a person holding a COCK near their mouth. I'm also having difficulty generating images of pussy licking – I've tested multiple prompt variations without success.I've experimented with both the Mop and Gonzalos models, but neither produced the desired results. Do you have any suggestions on prompt techniques or alternative approaches that might work better for these specific scenarios?
Hey, thanks. I've got an examples of both of those image concepts. If you look back through my image gallery in my profile you'll find them with the prompts.
I appreciate a non-DMD model! 👌 It is one of the best models to use with a LoRa. 👏 Thank you for your effort! 🙏
This can't be happening! Seriously!
It's impossible for it to work this fast and smoothly!!
AWESOME!
This must be black magic! It is THE best checkpoint I've ever used. Beautiful, vibrant photo quality images in under 3 seconds. Plays well with loras. Superb at NSFW (if that's your thing), excellent landscapes (post apocalyptic being my favorite), just all around fun to use. Kudos on a freaking amazing job!
Awesome, I'm glad it works so well for you!
This new flex model is my new favourite checkpoint.
SDE and under 10 steps? Now THIS is what i call black magic.
Incredibly good model that blew me away!
Thanks, I'm glad you like it!
hi, I'm only used to anime models, can someone explain how I should be prompting this?
I'm not sure there's any particular right or wrong way to prompt for this model. I tend to use short tags rather than a long narrative but I've seen both styles work equally well. Best thing you can do is take the prompt from an image you like and then adjust is from there.
You do great work, would love to see a zimage turbo version!
Thanks! I'm currently investigating doing a ZIT version too.
Version 3.1 works much more smoothly with the command prompt.
Hilarious (and accurate) model name!