DDIM Calibration/Replication of diffusers default Scheduler

[petekay]

I would love to replicate one of my dreambooth`ed samples, which produce really good results with this scheduler settings with diffusers (in notebook):

scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)

pipe = StableDiffusionPipeline.from_pretrained(model_path, scheduler=scheduler, safety_checker=None, torch_dtype=torch.float16).to("cuda")

these parameters for the scheduler are the default parameters, so you can write it shorter:

pipe = StableDiffusionPipeline.from_pretrained(model_path, safety_checker=None, torch_dtype=torch.float16).to("cuda")

the default configuration from the DDIM Sampler in automatic111 uses different packages, so it's not that easy to provide the exact same parameters.

is there a simple possibility to a) either extend the sampler GUI with "DDIM (diffusers)" and append the use the parameters or b) define a custom script with the specific DDIM Params?

i found those Samplers, but not the specific DDIMScheduler code:

ldm.models.diffuison.ddim.DDIMSampler
modules.sd_samplers.VanillaStableDiffusionSampler

Summing up, it's still not possible to simply replicate 1:1 from diffusers Pipeline (DDIM with above params) and the automatic implementation (or usage of wrapper-packages) ?

[MugoSquero]

Fining those arguments and changing scheduling_ddim.py and scheduling_ddim_flax.py in “stable-diffusion-webui\venv\Lib\site-packages\diffusers\schedulers_pycache_” directory and also in “C:\Users%username%\AppData\Local\Programs\Python\Python310\Lib\site-packages\diffusers\schedulers\” can be the answer you are looking for.

After training my face in Google Colab, using Dream Booth, I did get pretty good results while my ckpt file is downloaded. And after closing the session and trying locally, I couldn't get results even similar to what I used to. So I searched for and find those files, and changing the arguments did the job for me.

Also, I have marked those two files to read-only in pycache directory and removed diffusers and diffusers==0.3.0 from requirements_versions.txt and requirements.txt files. It gives errors when starting webui-user.bat, but I didn't get any problem.

The results are still kinda weird tho, but I hope we will find a better explanation for this.