Painted Faces - Possibly a training bug?

[XetalSL]

With this face (which was generated by Stable Diffusion):

Example Prompt:

british 25 year old female model in minimalistic formal attire. portrait.
High precision, high detail, realistic, no eye shadow, sharp, focus, documentary, in manor.

Used with the face control (any % strength) and the Digital Artwork trained set and no LoRA's loaded, I seam to get generations with faces being splotchy or painted faces:

[ETCore7]

I can't reproduce your results. Using your original as face reference just inpaints the face on target pictures.

[XetalSL]

I'm not in painting - just generating from scratch.

[ETCore7]

I still can't reproduce it, tried several default models.

[ETCore7]

You may have changed the sampler to DPM 2M SDE Karras, default is DPM 2M Karras (see below).

DPM 2M SDE Karras

Default

[XetalSL]

uhuh! - i did change that - I've tried it out and i get a lot less weirdness. out of 10 generations 1 was funky.

[ETCore7]

(edit)

Anyway, both samplers work fine with their default settings without controlnets, but if you want to use DPM 2M SDE Karras (= default sampler) with the Face controlnet you must lower its Guidance Strength to i.e. 2.0.

[ETCore7]

PS, if restoring defaults via the button results in a crash too, here are the default settings: