In this blog, a simple diffusion model with a 2-layer Unet and a simple scheduler will be implemented to introduce the use of Diffusers

What is Diffusers

Transformers-like library for the diffusion model

Customize Components in Diffusers

We need to write wrapper classes to embed our modules into the Diffusers. Inherit from f“{Union[Scheduler | Model]}Mixin” and DiffusionPipeline and achieve basic APIs for Diffusers to run. Some comments are generated by AI.

1. Scheduler

   • APIs: set_timesteps, step
   • code

2. Models

   • APIs: forward
   • code

3. Pipeline

   Strictly if the Scheduler and the Model are well prepared, just use Diffusers.DiffusionPipeline and pass the Scheduler and the Model as parameters(, as pipeline = cls(unet=unet_wrapped, scheduler=scheduler_wrapped)). But I still put an sample code here, in case I forget the usage one day

   • code

4. Usage example

   • code