Skin smoothness on realistic characters

wakwak

Hi! I'm currently trying to find good seeting for realistic character models. However, I don't like the skin texture; it looks a bit as if the AI would try to imitate pores/structure, but it looks very rough and unnatural.

See the following example:

Settings are:

Cyberrealistic v120
Style: photographic
Steps: 25
CFG_Scale: 2.2
Denoise: 1
Sampler: DPM++ 2M
Scheduler: sgm_uniform
Face_Detailer: on
Realism-Slider: on

No lora, and of couse high res pictures.

I also added "smooth skin" to the prompt, but that didn't help, so I assume it is about the technical details of the image generation.

Do you have any tips on how to achieve better results here? I am quite happy with the rest, though.

KairosJS

wakwak Hey, I've just tested a bit and you can try using a realism lora in addition to the model. I've tested a few this morning and this one was doing decent:
Illustrious Realism Slider

At strength = 2 for this lora:

Cyberrealistic v120
Style: none
Steps: 20
CFG_Scale: 3.5
Denoise: 1
Sampler: DPM++ 2M
Scheduler: Karras

Also I would recommend not using any style/medium with Illustrious.

You may try to get better results but I think this 'plastic' skin is inherent to older SDXL models such as illustrious. To get a real improvement you'll probably need to run more recent but also larger models such as: Z-Image turbo or Krea2.

KairosJS

Also this model:
LucentXL Pony by klaabu seems to be doing better, but I haven't tested it in other departments.

wakwak

Thanks, your settings are doing better. The LucentXL produces ₁₅% of pictures with buggy teeth/mouth, though.

As for Z-Image and Krea, I unfortunately get an error message: "An error has occurred during image generation, please ensure that the parameters are correctly filled in and that the paths for the checkpoint and Lora are correct."

I do have the the Z-Image and Krea Turbo models, and qwen in /clip as well as the ae.safentensors in /vae. Any idea on how to get this running?

Edit: z-image works now, the trick was to select clip method qwen instead of "default". However, the images are not being created properly, i.e. it shows the wrong poses/body parts.

Krea2 doesn't work, error message is: [ERROR] !!! Exception during processing !!! Krea2 expects conditioning with 12x2560=30720 features (a 12-layer Qwen3-VL stack) but got 2560. Load the text encoder with CLIPLoader type 'krea2'.
I selected Krea 2 in the settings, though.

KairosJS

wakwak For Krea2 you must follow the settings below:

That should work but it's really large to run locally, the clip model + krea2 take lots of VRAM.

wakwak

With these settings it works. I haven't had the Qwen3VL version, just Qwen3.

I do have a 7900xt wigh 20gig vram, but it was indeed not possible to create pictures with krea2, every step took like 4 minutes.

I will probably play around with Z-Image turbo and the cyberrealistic right now. Cause for the skin was the SDM uniform sampler; karras works better here.

wakwak

Following your remark on Pony Models, I gave them another try, after I was initially not a big fan of them.
What should I say... it was quite successful and I was basically swallowed by the rabbit hole of image generation.

I'd like to share my settings for realistic looks here:
CyberRealistic Pony (v18): https://civitai.com/models/443821/cyberrealistic-pony
DMD2 Speed LoRA: https://civitai.com/models/1608870/dmd2-speed-lora-sdxl-pony-illustrious

The DMD2 Lora allows you to quickly generate pics with quite good quality. On my machine, a picture takes about 10-15 seconds, as you only need very little steps and cfg. Settings are:
Steps: 8 or 12
config: 1.0 or 1.5
Sampler: LCM
Scheduler: normal

For finetuning, I am using the following LoRAs:

Explicit photography around 0.5: https://civitai.red/models/719983/pony-nsfw-explicit-realistic-photography
LowLight around 0.5: https://civitai.red/models/482151/lowreal-low-light-for-ponyxl?modelVersionId=536226

Explicit photography adds some nsfw details to the standard model and gives the skin a better texture. However, the pictures then tend to have overexposure; this I fixed with LowLight.

It works like a charm with many character LoRAs. Out of curiosity, I tried the Diablo 4 Barbarian Lora here: https://civitai.red/models/186984/female-barbarian-from-diablo-or-pony-sd15?modelVersionId=472453

And this was the result:
(Disclaimer: I fine-tuned it a bit with the Pony Body Slider).

It seems that some (mostly older) LoRAs don't work well with the newer versions and/or DMD2, but with those LoRAs which actually work it is really fascinating what you can do. So I thought it should share my experience.

What I would like to have as a feature is to save more than 1 LoRA per character. But I will make a separate feature request for that, I just wanted to share my settings for decent realistic pictures here.

KairosJS

wakwak Yes, the DMD2 LoRA is very performant, and fine-tuning characters while looking for the best composition is very addictive!

Using multiple LoRAs can give excellent results, especially for style, but be careful as it may also dilute prompt adherence.

The single LoRA you can set on the character sheet is reserved for a character LoRA, and its trigger words are inserted at a specific place in the prompt.
I may add a way to save a list of style LoRAs for a character, so you don't have to tune it yourself every time you start a session.