(see license restrictions below)
What is this thing?
MoP Pro is something I've been working on for a while. The base for this model is the latest version of my MoP (Mixture of Perverts) model. I've done a significant amount of training with various datasets I've compiled. These datasets contain 500 images and have been generated with my GonzaLomo ZPop Insta and MoP Mix Big ASP 2.5 models. I really enjoy the aesthetic qualities of those models and wanted to try to replicate them in a fast. normal SDXL model.
MoP Pro is not going to provide the highest detail or the crispest image, but that is by design. The focus on the training is in overall image composition, lighting, facial diversity/expressiveness and a certain everyday photographic quality. It's one of those things that's not easy to explain but I hope the example images I've provided do the explaining for me.
The recommended settings are:
Sampler: LCM
Scheduler: karras, beta, beta57, bong_tangent, ddim_uniform and exponential
Steps: 6-24
CFG: 1.0-1.7
It's very flexible with steps. In general, the lower the steps the higher the creativity but the lower the pixel quality. The reverse is true when you raise the steps. I've gotten good results with anywhere from 6 to 24 steps. bong_tangent really needs at least 16 steps and a cfg of 1.4 at least though.
It also isn't restricted to 1024x1024 sized latents. I reguarly use 1152x1536 for 4:3 ratio and 960x1728 for 9:16.
MoP Pro License
1. You may use MoP Pro non-commercially without crediting the creator.
2. You may not sell the images that it generates without obtaining a separate license or permission from the creator.
3. You may only run the official MoP Pro on Civitai. Running it on other public or shared servers commercially requires a separate license or permission from the creator.
4. You may not sell this model or merges of this model.
5. You may not merge, share merges of, or share LoRA extractions incorporating MoP Pro.
6. Redistribution of MoP Pro is prohibited. It may only be obtained from official sources designated by the creator.
7. Any restrictions in these additional terms may be waived only with explicit written permission from the creator.
Description
This first version of MoP Pro has been trained with 500 carefully curated images.
While this is technically a DMD2 model and requires LCM as the sampler, the scheduler options are pretty flexible. I find karras, beta, beta57, ddim_uniform, bong_tangent and exponential to work best.
The quality of the image is controlled by the steps, which could be anywhere from 6 to 24. In general, I would say that more complicated prompts or images in which the subject is further from the camera require more steps whereas simpler prompts or closeups on the subject work better with fewer steps.
The CFG should normally remain at 1.0 but can go as high as 1.7 if using bong_tangent as the scheduler.
The model can handle higher resolutions than 1024x1024. I tend to generate at either 1024x1344 or 1152x1536.
You can use the workflow embedded in this image. The workflow uses the Ultralytics FaceDetailer node to do targeted refinement of the face and eyes. This refinement can be turned off by disabling the node(s). You can find the bbox files for these nodes here.
Finally, one of my private loras was merged into this model. Using a trigger word in your prompt to activate that lora is not strictly necessary but it may help in some cases. The trigger word is '45p'.
FAQ
Comments (22)
This looks natural as it should be. very real looking skin.
Thanks very much for the feedback! This is going to be a work in progress but I think this base has a lot of potential for realism.
Maybe I'm doing something wrong, but I'm getting completely black images, using my own workflows or the suggested one embedded in the example images. All other checkpoints (both LCM and not) working fine.
That's odd, I've never seen that. Are there any errors in the console?
Could you please send me one of the image generations to [email protected] so that I can see what the problem might be?
It doesn't work if I don't include the "DiscreteModelSaving" node. That happens to me.
Discrete Model Sampling
@Tarendax That's weird, I've never had to use that node before. I'll try it out and do some investigation. Thanks
@GBRX No errors in the console. I've stripped the workflow back to the bare minimum and still black image. The workflow is embedded fine. I'll send it over to you. Tested with other checkpoints (both LCM and non-LCM) and they are all producing images fine.
@GBRX I've emailed you over a generation. I've also included the same workflow generation using a different checkpoint for comparison. subject: MoP Pro image generation
@GBRX If it's helpful the generation log is as follows:
got prompt
model weight dtype torch.float16, manual cast: None
model_type EPS
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load SDXLClipModel
loaded completely; 1560.80 MB loaded, full load: True
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
Requested to load SDXL
loaded completely; 5272.20 MB usable, 4897.05 MB loaded, full load: True
100%|██████████████████████████████████████████████████████████████████████████████████| 12/12 [00:08<00:00, 1.36it/s]
Requested to load AutoencoderKL
Unloaded partially: 3432.17 MB freed, 1464.92 MB remains loaded, 75.06 MB buffer reserved, lowvram patches: 0
loaded completely; 270.47 MB usable, 159.56 MB loaded, full load: True
Prompt executed in 22.36 seconds
Sorry for the multiple comments. Just noticed the SHA256 hashes don't match between the file I downloaded and the Hash reported on CivitAI. I'll download it again and re-check the hashes.
@hoopenball Ah yes, I forgot that could be the issue. Civit has been having download issues lately and I've had something similar happen with WAN models. You probably do have a bad download.
@GBRX Yepp, re-downloading the checkpoint worked. All good now, sorry for the interruption!
@hoopenball Oh great!
It's time to switch to Klein)
The Klein license is non-commercial only, too restrictive.
Ernie Image Turbo then
Do you use your gonzaLomo DMD v31 workflow for this one?? or is this another workflow???
I'm using a simpler workflow for this model right now. At some point I'll publish an update to my SDXL workflows.
Is there a way to add bong tangent to Forge or Forge Neo?
Sorry, I don't use those tool so I wouldn't know.