parediffusers 0.0.0

The library `pared` down the features of `diffusers` implemented the minimum function to generate images without using huggingface/diffusers to understand the inner workings of the library.

PareDiffusers

The library pared down the features of diffusers implemented the minimum function to generate images without using huggingface/diffusers to understand the inner workings of the library.

Why PareDiffusers?

PareDiffusers was born out of a curiosity and a desire to demystify the processes of generating images by diffusion models and the workings of the diffusers library.

I will write blog-style notebooks understanding how works using a top-down approach. First, generate images using diffusers to understand the overall flow, then gradually replace code with Pytorch code. In the end, we will write the code for the PareDiffusers code that does not include diffusers code.

I hope that it helps others who share a similar interest in the inner workings of image generation.

Table of Contents

Ch0.0.0 PareDiffusersPipeline

• Generate images using diffusers
• Without StableDiffusionPipeline
• Without DDIMScheduler
• Without UNet2DConditionModel
• Without AutoencoderKL

Ch0.0.1 Test PareDiffusersPipeline

• Test PareDiffusersPipeline by pip install .

Ch0.0.2 Play prompt_embeds

• Play prompt_embeds, make gradation images by using two prompts.

Usage

import torch
from parediffusers import PareDiffusionPipeline

device = torch.device("cuda")
dtype = torch.float16
model_name = "stabilityai/stable-diffusion-2"

pipe = PareDiffusionPipeline.from_pretrained(model_name, device=device, dtype=dtype)
prompt = "painting depicting the sea, sunrise, ship, artstation, 4k, concept art"
image = pipe(prompt)
display(image)

Contribution

I am starting this project to help me understand the code in order to participate in diffusers' OSS. So, I think there may be some mistakes in my explanation, so if you find any, please feel free to correct them via an issue or pull request.