search

kohya-ss / sd-scripts

Apache License 2.0
5.34k stars 881 forks source link

Training sample parameters #1169

Open shockman100 opened 8 months ago

shockman100 commented 8 months ago

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

kohya-ss commented 8 months ago

The samplers used in sampling images are Diffusers' sampler. I think Diffusers samplers are same as A1111's or ComfyUI's mathematically.

Please check the resolution, the negative prompt or steps etc. The default resolution for sample images is 512x512 even with SDXL, and the negative prompt is empty.

shockman100 commented 8 months ago

Something is definietly different. resolution and the prompts are set correcly with --w and --h. it is set in A1111 to I keep getting mouch better images during trainingsampling... jus tested confy, and it is way better.

kohya-ss commented 8 months ago

I had an opposite experience: the sample images are bad, but the final images are good, but don't have the similar issue. Could you please share the prompt for training sampling, and the screenshot or workflow for A1111/ComfyUI?

suede299 commented 8 months ago

I had the same problem as you. The images generated by the self-loading lora are not only ugly, but wrong images full of black stripes. Whereas the SAMPLES from the training process are fine. And, with the exception of auto1111 and comfyui, even if I generate it using kohya-ss's sdxl_gen_img.py, I get the same result. But this is generated when using network_module=lycoris.kohya, so I suspect that this is a problem only when lycoris has "dora_wd=True". I've only varied this parameter recently, and before that, even though the sample images were different from the ones produced by loading lora myself, at least there weren't too outrageous wrong images.