MoP Mixes are a series of checkpoints that merge the MoP (Mixture of Perverts) model with one other model that I find interesting. These are completely separate from the normal MoP releases because those are a much more complicated merge that requires a lot of testing. MoP Mixes are much simpler to produce with the goal of creating a MoP variation that hopefully some people may find useful or just fun.
To be clear, these MoP mixes should in no way be considered better than or a replacement for the checkpoint being merged with MoP. I strongly encourage you to download and test out the checkpoint for yourself and use it in the way that its author intended.
Each checkpoint will be set up in the manner that I think works best. So some of them will have the DMD2 LoRA fully integrated, some partially and some not at all. The recommended settings for each mix will most likely be different so check the About This Version section for all the details, as well as the recommended workflow.
Note: If you want to convert any of these DMD checkpoints to a non-DMD checkpoint then all you need to do is add the DMD LoRA and set the strength to a negative value. I would start with -0.85 and adjust from there. That will allow you to use a higher CFG and different samplers than LCM.
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
Note: This BigASP High/Low model may only work in ComfyUI. It requires a special workflow that I'm not sure other image generation tools are capable of.
This MoP Mix is a bit different because it uses two different model formulations of MoP v6.1 and BigASP v2.5, similar to how WAN 2.2 works. The High model is used to create the initial structure of the image whereas the Low model is used to provide the finishing details.
In my experience, this combination results in an image generation that is more stable, predictable and detailed. However, it probably also lessens some of the amazing creativity that BigASP v2.5 offers. Therefore, I strongly recommend that you also try Big ASP v2.5 in all it's glory, without my bastardization. I even have a workflow for it.
I also have another model that uses BigASP v2.5, imaginatively called LittleASP, if you're interested in that.
Unlike the other MoP Mixes, this is not a 100% DMD model. The low model uses DMD but the high model doesn't. The recommended settings are:
Total Steps: 25-35 (18+ on high, 7+ on low)
High CFG: 3.5-5.5
Low CFG: 1.0
High Sampler: Euler
Low Sampler: LCM
Scheduler: Beta (very important that both high and low use the same scheduler)
ModelSamplingSD3 shift: 1.0-2.0 (you may start seeing artifacts above 2.0)
FAQ
Comments (17)
I'm trying to use it in Forge Neo and it only gives me meaningless images, mainly lights.
In ComfyUI it requires a ModelSamplingSD3 node with a shift value between 1.0 and 2.0. I have no idea if Forge Neo or any other tools are capable of this.
It could be used in different SW than ComfyUI, but its effectively one of the kind model, flow matched SDXL. I think a bit of AI and dedication could pull it.
Es maravilloso e impresionante, después de adaptarlo a mi estilo de workflow, me he dado cuenta de que para usar LoRAs tienes que conectar el modelo LOW también a otro nodo de Lora loader y cargar ahí también el mismo LoRA. Y de esta manera, el cielo es el límite. Saludos y muchas gracias maestro.
Thanks for the great feedback!
Can someone tell me Ksampler settings to get result to look like my lora?
No one would have a clue without more info. Was your Lora trained with the model you're trying to use? What does the model you're trying to use say.
@woodenpickle Hi, Lora was trained with sdxl. I work in comfyui
Hey @GBRX could you please explain what is the effect of ModelSamplingSD3 with shift 2.00 in your workflows?
I'm no expert on this but the shift value has an impact on the denoising schedule. A higher value means that more time is spent in the high noise section of the sampling operation. Consequently, more emphasis is placed on the composition of the image.
In my workflow I've found that a 2.0 shift will provide a nicer composition. Going higher than that can start to cause artifacts to develop in the second sampler.
@GBRX unexpected thing is that the node for SD3 works with SDXL
@homoludens Big ASP 2.5 is not a normal SDXL checkpoint that uses classic diffusion. It uses Flow Matching. Don't ask me to explain it because I'm no expert.
@GBRX I herd this term. This explains all
@GBRX can you explain the section on Fooocus in the model description, I didn't see it in the workflow.
The workflow is for ComfyUI. Fooocus doesn't use workflows which is the reason for the step-by-step instructions. Fooocus won't work with the Big ASP 2.5 model though.
Ah you were talking about using the standalone Fooocus app right? lllyasviel/Fooocus: Focus on prompting and generating. Not a Fooocus node inside ComfyUI..
@nonamezoname7621 That's right, the Fooocus app doesn't work automatically with DMD checkpoints so that's what the instructions are for.