Fork of Text-to-image model editing (TIME) 🤗

This is a forked version of the diffusers 🤗 implementation of Text-to-image model editing (TIME, Editing Implicit Assumptions in Text-to-Image Diffusion Models).

The code in this repository is managed at py-img-gen/diffusers-text-to-model-editing.

Here are the minor changes:

• Changes due to the renaming of the CrossAttention class to the Attention class
• Removed the restart_params with large side effects from the edit_model function

How to Run

import torch
from diffusers import DiffusionPipeline

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

pipe = DiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5"
    custom_pipeline="py-img-gen/stable-diffusion-text-to-model-editing",
)
pipe = pipe.to(device)

prompt = "A field of roses"
source_prompt = "A pack of roses"
destination_prompt = "A pack of blue roses"

pipe.edit_model(
    source_prompt=source_prompt,
    destination_prompt=destination_prompt,
)

output = pipe(prompt=prompt)
image_edited = output.images[0]
image_edited

You can find the notebook example in notebooks/run_pipeline.ipynb.

Comparison of the generated results

• Prompt: A field of roses
• Source Prompt: A pack of roses
• Destination Prompt: A pack of blue roses

Original	TIME

Acknowledgements

The code in this repository is based on the Text-to-image model editing (TIME) implementation by 🤗 diffusers.