Allow seeding initial noise latent

[pomofomo]

This is needed for reproducible generations, if I want to generate the same image again, the prompt and noise is sufficient. Currently, the noise tensor is generated random (no seed) inside StreamDiffusion (pipeline.py).

I added a noise argument (default=None) to StreamDiffusionWrapper (wrapper.py) and StreamDiffusion (pipeline.py) for the txt2img variants and the __call__ method, to allow the called to specify this. It works for batch sizes larger than 1 as well.

The user can either:

• Pass in None (or omit it) to maintain current, unseeded latent vector
• Pass in List[int] (of length batch_size) to use each element as a deterministic seed for one batch element
• Pass in int, which is shorthand for List[int] only when batch_size == 1
• Pass in torch.Tensor, which must be of the proper size, and allows the user to have full control of the noise tensor

I also exposed two more methods from StreamDiffusion, which could be used by advanced users to create and manipulate the noise vectors further (eg. create "similar" variants, but slowly deviating from the noise vector that created the given image). These are noise_size (to show the required input size) and noise_from_seeds, which is the internal generation method for the List[int] case, in case further processing is desired.

Note: I didn't know how to write automated tests on this codebase, but I modified the code in examples/optimal_performace to use this in both single image and batch generations, with different configurations and manually confirmed it works. I will push in a second commit commented out, so you can quickly try this.

[teftef6220]

Thanks for the PR.
We are currently working on it in our team.

[worthy7]

Hi there, bump on this!