Using a scheduler that works for pretrained dreambooth weights and default weights

[zeke]

We are in the process of consolidate the prediction from replicate/dreambooth-template into this repo, so the DreamBooth API will use the same predictor code that https://replicate.com/stability-ai/stable-diffusion uses, so all the nice features we add to the "canonical" model will become available to newly trained DreamBooth models.

From #59

The only other major difference between this and dreambooth-template is that it has a hardcoded scheduler:

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

The default scheduler seems to work - although I don't know if those "magic numbers" in the DDIMScheduler in dreambooth-template are to maximize the quality from the dreambooth generations?

[zeke]

@chenxwh I know you had a hand in writing the predictors for both of these models, so I'm guessing you might have some context on why the template repo is hardcoded to use DDIM. Is that necessary?

[zeke]

@anotherjesse did you end up working on this?