Faces in V1 are not diverse enough by my opinion. V2 is better in this regard... I would say V2 gives more fun to me. It's just crazy sharp when you do upscaling with I2I.

https://civitai.com/posts/11707807?returnUrl=%2Fmodels%2F1096167%3FmodelVersionId%3D1289716
just look how it is SICK with many LoRas and HiRes applied. It blows my mind with details...
Faces also quite new.

Why don't you post the samples. Don't think @homoludens are not interested in debugging the issues and fix it...

@dioxidin  Yes, HL's always been interested in improving and revising things. I'll have to run them again. Let v2 work overnight and out of 100+ samples, I couldn't find a single one that I felt was worth saving (at least I didn't get frustrated and rage delete the models again, lol). I'm going to try some close-ups, since this model seems very capable of that, but even in the gallery a lot of the mid to full-body portraits seem to exhibit the same results I'm getting. Honestly, I just really hate posting pics that don't have at least some resemblance of photorealism, but I guess I could always delete them later.

@AFD_0 I'm using the 2-pass approach. forex, 832*1216 1st run with Euler A + SGM_Uniform
2nd pass with ADetailer using the Yolo Person as 1st detector and Yolo Face as 2nd detector.
If you set for each detector a resolution 1200*1200 in Inpainting section you most likly will get a closeup on face and hires fix on body.
Also try settin the Heun + Karras samplers for ADetailer 2nd pass. This pipeline mostly give a good result (for my opinion of course)

BTW, there is a thing with V2 and CFG. It can render images at CFG higher than 7 without overburn, but image looks much less realistic. I feel that CFG 3 ... 3.5 gives more real image and let you use higher weights in prompt.

And one additiona thing I discovered: Img 2 Img Noise Multiplier. There is a global parameter in A1111 WebUI and there is corresponding parameter in ADetailer.

1. Value less than 1 - make image smoother (similar to FLUX)

2. Higher than 1 - adds more noise and can give more details on final image.

Maybe you can also play with it j4fun

0.97 ... 1.03 - this range is optimal to me. Outside of this - too smooth or too much noise.

@dioxidin Good suggestions, thanks! Think HL also recommended doing multiple passes with his models to me before. Unfortunately, I'm using Easy Diffusion, which is pretty limited vs A1111 or ComfyUI. It has a decent selection of samplers (but not everything) and no choice of scheduler or ability to do multiple passes afaik. There's an older ControlNet, img2img, 3 upscalers and 2 choices for face correction that actually work pretty well most of the time. It's a lot more versatile than what the CivitAI UI offers, but it certainly has its limitations.

Still plan on switching/learning a full-featured UI eventually (whenever I get a better GPU and build a new PC), but for now I really enjoy the simplicity. Even with limited options, I still find myself constantly wasting massive amounts of time testing different settings rather than enjoying the creative process. Like everyone else, I'm still learning things, but really try to focus on quickly getting an "acceptable" result and then having fun tweaking things and experimenting from there. When I get stuck on that first step, it stops being fun and starts to become frustrating if I keep at it too long before thinking, "maybe this one just isn't for me". It's actually rare when that happens, but HL's early builds often tend to do that for me. And that certainly doesn't mean they're bad or anything, especially given some of the amazing results that are posted (especially yours!)

I really love a lot of HL's models and revisions and just like to give a bit of basic feedback on which ones I enjoy, or if one does or doesn't work for me. Only reason I started posting my results was when HL kindly reached out when I was having trouble with early builds of XenEro (using something that reflected my usual style/workflow and generally turns out "acceptable" for most Pony Realism models with just a slight change in sampler, CFG and steps).

I'll probably never enjoy a simple one-sentence prompt, and after too many words, it's all about trying to get the best image quality without sacrificing too much adherence. At least for what I'm looking for. Some models just hate wordy prompts, negative, certain LoRAs, etc. And sometimes a model is a bit more specialized than what I'm able, willing or wanting to work with, and some models/revisions just excel in certain things and less in others, and that's 100% okay. As long as others are getting good results and enjoying a particular model/revision, then it was absolutely worth creating imo.

@AFD_0 As far you use a simplistic tool which is less popular than mainstream... its hard to help you that much. I suspect your tool had some non-optimal params whaich affect quality. You should try smth more complex, IMHO. With A1111 you can just take any good png and then load it in A1111 PNG info. This way you can learn quickly. This is what I did with images of HL and others.

@homoludens wrote that CFG range made wider and it's seems that yes I can use very high CFG without results damaged. But if I want Tex-to_image Realistic I keep CFG under 6 (usually 3.5-5).
If I need realistic Image-to-Image I prefer lower CFG to 2.0-3.0.

CFG 7 - Semi-Realistic
CFG 9+ - closer to digital art.