only high quality images are without exifs. Like I said in the description, I don't want to take great fun from you, so... :)

@jagona Ok, I will upload some middle quality images with exifs that you can play around and find out what the best numbers and prompts for my model. :)

@stanleypain007313 will upload some middle quality images with exifs that you can play around later. 

uploaded some sample images with exifs. go check it out. And I recommend you to run as many shot as you can and get your own know-how by it. It's that every model has different optimal prompts and setting for photo realism.

@jagona usually when with more than 12 cfg, you must use dynamic thres. otherwise the image gets broken badly. Also, set hires upscale sampler to ESRGAN with 0.3/30. Dynamic thres. is not hard to use if you are not into a profound adjustment on an image generation. Just set the mimic cfg same as cfg. Also, use ddtailer always with smallest numbers otherwise the face changes and becomes more 2D. I wish civitai also showed dynamic thres., upscaler and ddtailer numbers so that people can easily get the sense of what to do with a certain model by looking at sample images...  

@stable_diffusion7388 yeah I get what your point is and thanks a lot for reminding me. I will add some more on the model description.

@stable_diffusion7388 done adding hires, clip skip and extensions info. on the description.

vae-ft-mse-840000-ema-pruned. generally, this vae is used for not my model only but most photorealistic models.

samples with exifs uploaded.

@vitherenegade thanks!

very very nice comment! I use 'esrgan_4x' sampler on 'hires' for a photorealistic image because it does not much 2d on images like R_esrgan. setting numbers are usually 0.3/30 but if I want to create an image with a concept like dark gloomy...etc, I adjust the number to get the best feeling. I mostly use 2x upscale.

And, the ddetailer and threshold! They are very important in generating top-quality images. I put as smallest number as possible for ddtailer (denoising/inpaint number) because it turns face into more 2d when the number is too high. Usually, you don't need to touch anything but denoising/inpaint) number, and those number should change depending on the image composition. (if char is in a very far distance, you increase those numbers and if too close, you turn it off...so on). For the sample images in this page, most high quality images were generated with 'bbox...' as a primary detection model, everything default except for denoising/inpaint about 0 to 20/ 0 to 24. (do not touch other numbers if you don't know what they do for sure. Adjusting other numbers are required for generating a truly top-quality image but you must first know what they are for.)

I want to tell you about dynamic threshold in very detail but it will be too long to write here. Just want to tell you that when prompts are not recognized due to its huge token sizes, you need to increase cfg and sampling steps until it does. I usually prefer high cfg number because I use so many detail prompts. But to find a proper cfg for the high quality image you desire, you adjust not only cfg but also sampling steps and prompts together, otherwise even with dyn. cfg an image will be broken.

As for 'Dynamic Thresholding Advanced Options,' I set 'Top percentile of latents' to clamp to 100 (rarely change), mimic scale to cosine up / half cosine down for most sample images in my model page (with 0.5 to 4), but they all change depending on composition, background, details ... etc.

Hope this helps!

also I uploaded middle quality sample images with exifs. go check em out and play with dynamic thres. and ddtailer to make images more realistic.

for only prompt generation, you put more detailed prompts like (alternated face), (alternated body)...etc and make sure the resolution is wide enough to handle multiple girls. Also, for a better image you can use the extension called regional prompter. I guess once you get how to use the extension, you won't worry about it anymore at all. Hope this helps.

Unfortunately, my gpu is not good enough to make a SDXL model. I guess SDXL creators will come up with a very bad defect (demanding too much gpu for making a model) in the future. Then, I will work on it for sure.