Correct timestep for img2img initial noise addition

[wfng92]

There seems to be difference between huggingface's diffusers and compvis's implementation for img2img as the wrong timestep is used for the additive noise for image-2-image.

Summary from the relevant PR raised by the HuggingFace team:

t+1 is used as the timestep to add noise to the original image, but [0, ..., t] is used afterwards for the denoising process. We should however also use t when adding the noise to the original image. This can be quite easily verified by doing the following. Run a img2img with a small number of update steps and a very low strength because then differences between t and t+1 become quite clear.

From

torch.tensor([t_enc]).to(self.model.device)

to

torch.tensor([t_enc - 1]).to(self.model.device)

For example, using the following init_img and configuration:

• prompt: A fantasy landscape, trending on artstation
• strength: 0.1
• seed: 42
• sampler_name: DDIM
• steps: 10

The existing code will generate an image with the following noise:

The new implementation will produce an image with lesser noise:

This PR affects only the img2img script. I realized that text2img2img and inpainting share the same code base but I am not sure if the same implementation should be applied as well.

[lstein]

Thanks so much for finding this problem. I will test it out the proposed solution later today.