Thank you :)

@coyotte could you please share your workflow for inpainting?

@coyotte That's a great way of adding detail without adding additional training on top that might disturb the results! Good job!

@coyotte would it be possible to turn that skin texture fix into a Lora that also works with other checkpoints? or could you describe somewhere how to apply that fix other checkpoints?

@coyotte What loras u recommend using for more realism? or just overall use, what model u recommend to use now I saw the new version OLT?

Try OneTrainer. Much easier to use & the default settings are already quite good. Just switch off image augmentation. I use WD14 & Joytag for tagging..Check out my loras at https://civitai.com/models/1115264/subtle-styles

@SubtleShader I see, I'll look it up, thank you. Is it also good for capturing and then generating the same face consistently? If I understood right, your LORA's are trained for entire styles instead of just one persona. Thanks for the recommendation brother.

@JKCsaba My loras manage to reproduce faces that were on just 1-3 training images out of 220. If you use 10-20 images of a face, it should do fine. But train on Lustify or my Big Love model (not on base SDXL) for higher quality. OneTrainer lets you tag with WD14. For additionally adding Joytag tags I use https://github.com/Particle1904/DatasetHelpers

Neighter of those, for me DMD2 Endgame is the best, faster, more consistent and also better quality

@NinehundredBr Try the new OLT with the DMD2 lora. You might like it better.

@SubtleShader Where can I find this lora?

@geocampanha745 https://huggingface.co/tianweiy/DMD2/blob/main/dmd2_sdxl_4step_lora_fp16.safetensors

@coyotte What would you recommend as prompt for amateur style 'look'. for whatever reason amateur photo or analog photo doesnt seem to be giving me quite the result im looking for in the latest model.
I've been using Euler + Karras @32 steps with 4x_NMKD-Superscale-SP_178000_G chosen as my upscaler, if info is useful to you.

@coyotte Yep. I never use negatives. Only if cannot get a woman to look older any other way.

@SubtleShader Thx gonna try it, because i tried few settings during hours with different Lustify models and i always end up with a subpar quality over their DMD2 counterparts while losing speed too, when with DMD2 Endgame it takes less than 10s per image with almost 4k quality and almost no weird stuff.

@NinehundredBr Yep. I always use DMD2. More photorealistic than anything else and does anatomy very well. Only problem is overexposure sometimes and low cfg (1-1.5). But can live with it. Maybe I will do a lora to counter it. My 4070S GPU renders a 1024x1536 image in 3 seconds. Usually I don't even need to upscale. Very fast and intuitive way to generate images.

@SubtleShader i got ~9s for a 1104 x 1640 image with 4x Ultra Sharp upscaler on a 7900XT, sometimes the quality that i get is so real that you can't tell the difference if its AI or not.

@SubtleShader THX for the Lora it works, didn't lose quality nor speed, what a sorcery is this ? Can i use it on different type models or its only for SDXL/Pony ones ?

@NinehundredBr You can use DMD2 with any SDXL model. I'm no fan of hires fix and upscalers. I get the best quality with img2img, but seldom use it as it is not such a big difference to txt2img without any upscaling.

You lose ton of details with dmd2, but here everybody is into 1girl.

@paulbarrientos606 Still has enough details for my taste. A lot of real photos don't have a ton of details either.

@SubtleShader Yes, its a matter of taste. Common girls are not enough for me anymore:)

@coyotte Alright and does it matter a whole lot what upscaler one uses?, if anything specific you could recommend?, again keeping the amateur style look in mind.

@paulbarrientos606 Don't know what details you mention because everytime i tried, i always ended up with subjects looking bad, weird stuff and terrible quality, like grain all over and blur everywhere. And to get "almost" the same quality as i got with DMD2 it takes like ~40s to more than a minute per image while DMD2 its like 9s.

@paulbarrientos606 What is the better alternative to DMD2 in your opinion?

@coyotte and that would be the '4x_NMKD-Superscale-SP_178000_G' or the '8xNMKDSuperscale_150000G'
Also, what amount of Hires steps, Hires CFG Scale do you go with and the Denoising strength?
Thanks again.

@coyotte thanks. I found 'photorealism, cinematic photo of' also gives decent 'realistic' results you may give it a go.

And yet probably one of the best models out there for porn realism

@NinehundredBr i thought SDXL would die after Flux came out. Looks like SDXl has the capability of redemption. i cannot wait to try this out.

@Noob_ee Flux (currently) is a technological showcase (according to me), SDXL is what we expect from AI: flexibility and freedom. We also tend to forget that SDXL is made up of 6.6 billion parameters

@WearCatEars agreed. i enjoyed SD models for the longest and others were so hyped with SDXL, but never gave me that extra pazazz and shy away from AI art gens almost entirely.

It is not really base SDXL. Lustify is based on bigASP, which was completely retrained from scratch.

@coyotte But you started with it. With my Big Love merge I pilled a lot of other stuff on top of Lustify, but its character is still there, despite Big Love doing some other stuff too.

It should work better with celebs, if there is no Bigasp in it anymore.

@coyotte Hi, in the same boat, and based on my compares using the lora produces visibly worse result for the same settings than with the merged models. I'm not sure why is that but some models perform better with the DMD2 lora, some not.

@EfficientMitosis +1, Models with lora already in are better imo

Man, you built that model at Flash speed! This is why I enjoy working with and studying your models. I also really appreciate how you hardly ever ignore people's feedback on your models.I just love the way you work. You've done an amazing job in this field, and I’m really looking forward to seeing what you create next. Maybe something more on the artistic side? Haha, who knows, it's all up to you. Take care and stay awesome!