@iamddtla Is there a specific resolution? 🤔 I can see the preview images you posted also have some issues with the teeth and such. V7 def looks much better to me. But i give it another round of testing tonight

@iamddtla I see. I don't entirely understand how training on a base checkpoint works like sdxl. I do understand lora training though. I was under the impression that if you trained it enough with images of it, considering you're training it with various nsfw images, that it could learn new concepts as a lora works in the same way. I guess I was thinking it could be better than a lora considering your training the entire model with it. Is that incorrect thinking? Can new concepts not be learned from training a model on base models with enough various images?

@iamddtla, can you put a URL/link here for me please? I'm having trouble finding how to link right back in. :/ Been with you like almost 2 years now. Definitely don't want to fall out on this project.

Invert Mask technique. Mask the face and everything else you want to keep, then invert the mask, so the masked area becomes unmasked and vice versa. I don't remember if the built-in masking tool in ComfyUI has this option or if it was a custom node.

it is called img to img inpainting

@iamddtla You make no mention of what scheduler it uses? Karras?

comfy ( from Github), wan2.2 (High and Low), Workflow, all on civitai or you use the template option in comfy

Yeah I get the same with V8, V7 is amazing but V8 has messed up eyes and fingers. Even shows that problem in the example images shown on the model itself. Maybe just unlucky. Hoping V9 will be as great as V7 was!

This is a problem with absolutely all models. Not just SDXL. When moving away from the object. Use ROPE DeepFake to overlay the face onto the final image. Or work with a higher image resolution. I use 1408×1152 or 1408×1024.

Adamw8bit work very good too. Adafactor is better?

@broncedelic My main goal is the most accurate transfer of the quality of the original dataset photos. I don't need the model to make improvements on its own, because on Prodigy the model burns the skin texture. I use separate datasets of 100-200 photos for each. Face separately, body separately. With a different number of repeats for the datasets. The photos in the datasets have already undergone 4-level preprocessing and do not need any improvements. I copy the quality 1:1. Adafactor is the most honest optimizer in this sense. It does not smooth out or embellish the data. If the photos are sharp, with pores, highlights, hair volume — the generation will be the same. Prodigy or Adam often give a slightly more glossy and polished look even with the right settings. The only problem with Adafactor is large datasets. If the face dataset is, say, 30-50 photos, it won't remember any features even with a large number of epochs, but will overfit.

saran model paling bagus bang, pakai bigAsp kok blurry ya

https://pastebin.com/5DHdv3BL
try this one

copy all the text and the paste (ctrl-v) on a unsaved workflow

Sometimes that happens when you set clip skip to -2. Try setting it to -1 and see if that fixes the issue.

hahahaha I think you just need to look at other users' creations here to see what a checkpoint like “pornmaster” does.

If you're new to inference, I recommend watching YouTube tutorials, but to make it very quick.

If you want to create an NSFW image and then create a video with that image, you need to do the following:

First, create an image by choosing the model you like (SDXL, Pony, Illustrious, Flux, etc.). Then choose a checkpoint. These are models that have been retrained to often produce NSFW content. Here, for example, pornmaster is based on SDXL, but realistic pony is based on... pony!

Then you will have your beautiful image, which you will need to animate. This is where the video models come into play. Since you are starting from an image and not from scratch, you will need to find an I2V (image-to-video) model. Currently, wan 2.2 is the most popular locally (especially for NSFW), and again, you'll look for the re-trained models available on Civitai, for example, WAN2.2 ENhanced NSFW.

In short, allow a few days if you're starting from scratch, because it takes some learning. And above all, watch YouTube!

And if this was a joke that I took at face value, then RIP xD

LOL, it's not as easy as you imagine. First you have to see what your PC (video card) can support, then pick a base model like SDXL or even stronger depending on what your PC can support, then download checkpoints for that, figure out the parameters and see which ones give you the images you want and which don't. Or create the images online, which is a whole different thing than using your own video card. Pony is another base model for anime, not realistic images. There are various kinds of models even within a base like SDXL, these are called DMD2, lightning, hyper, LCM, etc. They require different parameters and some are much faster than others. Have fun.

I'm not getting this one, for example because it doesn't have a DMD2 or other fast version so it would take 2-3 minutes to generate a single image on my PC.

@forum2233726 Absolutely true. That's why I've pretty much given up on using checkpoints and LORA's from individual contributors on CivitAI. A ComfyUI template for, say, Z-Image-Turbo or Wan 2.2 is usually super simple. Just download and it runs fine. But these freakin' LORA's with either no documentation or very poor documentation, no button that says "Download Basic Workflow", and when you try to get a workflow by downloading a provided image or video often it doesn't work, or doesn't include workflow metadata. And then if you do get a workflow it's 326 custom nodes, and it takes hours figuring out how to install them, or there's an error in the workflow that you can't figure out. And if you're somehow lucky to get it running it either doesn't run on your hardware, or gives terrible results. And then you spend hours and days trying to figure out if it's a setting problem, or a LORA limitation, or you're not using the right magical prompt or trigger word. And then you ask online and you get responses from people who have spent the last 2 years figuring it out and just want to sound smart, so they tell you to figure it out yourself. I've spent so many wasted days with this nonsense and getting nowhere. I'll just stick with the basic Z Image and Wan 2.2.