Experiment with combined base+refiner for SDXL

[stronk-dev]

Lets not hard code it like this, maybe enhance the request to be able to request a list of sequential models?

[stronk-dev]

Apparently there's also a big difference between:

• Base model x Refiner:
  Use the base model for the high-noise stage and the refiner for the low noise stage

            kwargs["denoising_end"] = 0.8
            kwargs["output_type"] = "latent"
            image = self.ldm(prompt, **kwargs).images
            del kwargs["output_type"]
            kwargs["denoising_start"] = ["denoising_end"]
            del kwargs["denoising_end"]
            kwargs["image"] = image
            return self.refiner(prompt, **kwargs).images

• Base model -> refiner
  Use the base model to fully denoise the image and then use the refiner to add details

            kwargs["output_type"] = "latent"
            image = self.ldm(prompt, **kwargs).images[0]
            del kwargs["output_type"]
            kwargs["image"] = image[None, :]
            return self.refiner(prompt, **kwargs).images

I am getting good results with both. To support refiner models, we would require a param to set denoising_start/end and a separate num_inference_steps for the second stage. So ideally you would be able to request:

• Base model
• Base model x refiner (with denoising_start/end)
• Base model -> refiner