Google Colab Non WebUI Stable Diffusion XL Image Generator

This is a simple Google Colab notebook made by an 18-years-old junior programmer for generating images using Stable Diffusion XL.

Features Overview

• Image Generation: Uses Transformers and Diffusers as the major components for image generation, along with ControlNet, Inpainting, and IP-Adapter.
• Download and Load: Ability to download and load images, checkpoint, LoRA weights, and VAE using direct URLs.
• Textual Inversion: Ability to load embeddings for more output control.
• Image-to-Image: Turn your images into something more creative and unique.
• Prompt Generator: Generate creative prompt powered by GPT-2. (Model by Gustavosta)
• Image Upscaling: Using Real-ESRGAN made by xinntao and Real-ESRGAN's contributors.
• Google Drive Integration: Can connect to your Google Drive or disable it.
• User-Friendly UI: Simplified complexity with an easy-to-understand user interface.
• IPyWidgets: Simplified UI so that you don't need to scroll up and down.
• Built-in History System: Can show all of your previously generated images in Google Drive without opening it, just with a few simple clicks.
• Preset System: Enables saving, loading, renaming, and deleting presets containing your saved parameters.
• CivitAI Token Support: Pass your CivitAI token for additional functionality, but remember never share it with anyone.
• Hugging Face Token Support: Pass your Hugging Face token for accessing private files, but remember never share it with anyone.
• Simplicity: Consists of two cells only.
• IPyCanvas: Create your own mask image for Inpainting immediately. (unstable and may introduce issues unexpectedly)

Usage

This notebook is designed for artistic purposes and to spark inspiration. Please use it responsibly. Creating deepfakes and uploading them online are strictly prohibited.

Installation

Simply open the notebook on Google Colab to start using it. No installation is needed.

Link:

Note

Check out the Legacy branch to use the old versions of this notebook.

Disclaimer

• This tool is intended for creating art and exploring creative image generation.
• Use responsibly. Please refrain from using it for any malicious or harmful activities.
• NSFW generation is supported, but do it at your own risk.

Feature Table

Click here

	Features
1.	Base pipelines and autoencoder (ControlNet, Inpainting, VAE, Text2Img)
2.	Base adapters (LoRA, IP-Adapter)
3.	IPyWidgets
4.	Saving and loading parameters
5.	Interactive UI
6.	Linking widgets
7.	History system
8.	Upload images directly
9.	Image-to-image
10.	Textual inversion or embeddings
11.	Send images from history to Image-to-image, ControlNet, Inpainting, and/or IP-Adapter
12.	Reset button (defaulting the parameters)
13.	Compatibility with saved parameters from previous versions
14.	Preset system (saving and loading custom parameters)
15.	GPT-2 Prompt Generator
16.	Hugging Face token integration
17.	Real-ESRGAN Image Upscaling
18.	New parameters saving system (from a flat list to a dictionary) and a conversion method to convert old saves to the new ones
19.	URL saving system to save any previously-inputted URLs into a JSON file and access them with filenames only
20.	One model at a time for all pipelines instead of multiple models for each pipeline to avoid OutOfMemory error
21.	Preview buttons to display ControlNet's converted image
22.	components integration for memory-sharing between pipelines to avoid OutOFMemory error due to independent tensors
23.	Implementation of ControlNetUnion for better memory management and scalable approach
24.	Corrupted-download deletion logic
25.	Preview buttons to display uploaded IP-Adapter's images and URLs, as well as buttons to remove them
26.	ipycanvas canvas to create mask images for Inpainting (unstable)
27.	Pagination feature in history system
28.	Multiple images per generation (requires higher VRAM, so it's not T4-GPU friendly)
29.	Default model selections (12 models in total)
30.	Hires.Fix with Real-ESRGAN and LANCZOS upscalers (requires higher VRAM for high upscale factor)
31.	Delete button for history system
32.	Combobox widgets for VAE, checkpoint, LoRA, and Textual Inversion to introduce selections based on previously-inputted URLs

Preview

These previews were taken before adding the feature for multiple images per generation (can be seen in Feature Table number 28).

Text-to-image

Image-to-Image

ControlNet

Inpainting

Settings:

Canvas:

LoRA

Textual Inversion

IP-Adapter

Real-ESRGAN

History

Preset

Save:

Load:

Rename:

Delete:

Token Saving (Optional)

This notebook doesn't save your Hugging Face's token and CivitAI's token due to security concern. But, if you wish to save your tokens to your Google Drive or to your runtime, you can follow these steps.

Caution

I strongly don't recommend you to save your tokens anywhere else, especially if you share your Google Drive storage with anyone. Any potential exploitation or misuse of this method won't fall under my responsibility.

Here are the steps to save your tokens to Google Drive storage or to your runtime:

1. Open a notebook in Google Colab and create a new cell first. This step is required to run the code.

2. Copy this code and paste it into your new cell.

   from google.colab import drive
   import json
   import os
   
   hf_token = "" #@param {type:"string"}
   civit_token = "" #@param {type:"string"}
   gdrive = True #@param {type:"boolean"}
   
   token = {
       "hf_token": hf_token,
       "civit_token": civit_token
   }
   
   if gdrive:
       drive.mount('/content/gdrive')
   
   base_path = "/content/gdrive/MyDrive/Saved Parameters/Token" if gdrive else "/content/Saved Parameters/Token"
   os.makedirs(base_path, exist_ok=True)
   
   with open(os.path.join(base_path, "token.json"), "w") as f:
       json.dump(token, f, indent=4)

3. Fill the textboxes with your tokens. You can save it to your Google Drive storage by enabling gdrive.

4. Run the cell. You'll see a file named token.json in the "Saved Parameters/Token" folder, indicating that the tokens are saved.

5. To use the tokens, simply run the second cell as usual. If the GPU has loaded a model, restart the runtime first before rerunning the second cell to avoid OutOfMemoryError.

These steps allow you to reuse the tokens without inputting them manually in the widgets. But, beware of any potential leak!

To-do List

Details

• Adding textual inversion ✅
  
• Implementing Inpainting using IPyCanvas ✅
  
• Creating the UI preview ✅
  
• Modularizing the code ✅
  
• Overhauling the save system ✅
  
• Adding Hugging Face's token integration ✅
  
• Implementing Img2Img pipeline ✅
  
• Pagination for history ✅
  
• Use other presets as the main preset ✅
  
• Enable custom upscale factor for Real-ESRGAN ✅
  
• Multiple batches per generation ✅ (requires higher VRAM)
  
• Default model selections ✅
  
• Hires.Fix (Real-ESRGAN and LANCZOS upscalers) ✅ (requires higher VRAM)
  
• Remove images from history ✅
  

Feel free to open an issue if you found one. Each time you submit an issue will improve this notebook's capability and further enhance your experience.

License

This project is open-source and free to use under the MIT License.

---

This README was partially made by ChatGPT.