Training sample parameters

Hi
i am currently training loras with sdxl 1.0. The sample images during training are flawless, hovewer if i rty to use the resulting lora i become disorted results, for example the faces are ugly. It seems that the first preview is ok but the final image becoming ugly details.

I tryd:
A1111 and Confy ui.
change lora weight,
reload, restart/ i am using runpod so everiitime i start it is a clean install
tried different docker images
tweeked all samplers denoise settings, cfg scale, ewerithyng.

now it loks like sampler in kohya ss is using somewhat different "renderer" as i.
for sampler i seen euler a is the default, and this is the only setting i can change.

So what method is sampler using, or any advices in this case?

Thx

Start by reproducing the SDXL LoRA training and preview workflow described in the issue, comparing the Euler a sampler and settings used by kohya_ss with A1111 and ComfyUI. The issue does not name files or tests; done would require isolating the cause of the difference or documenting a reproducible compatible configuration.

Written by the indexing model from the issue text.