切图

import torch
from sam2.sam2_image_predictor import SAM2ImagePredictor

predictor = SAM2ImagePredictor.from_pretrained("facebook/sam2-hiera-large")

with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
    predictor.set_image(<your_image>)
    masks, _, _ = predictor.predict(<input_prompts>)

https://huggingface.co/qualcomm/Segment-Anything-Model/tree/main https://segment-anything.com/demo#

https://github.com/Render-AI/segment-anything-2/blob/main/notebooks/image_predictor_example.ipynb

sucorn

pronounced as su-kon

A project to automate the collection of AI generated images on different platforms, and to classify which ones are good and which ones are bad, so that time spent during the ML/DL life cycle is reduced The goals of this project include

• deprecated automatic collection of images using pyautogui,
• automatic collection of images through the BingImageCreator reverse-engineered API
• automatic collection of images with Google's ImageFx
• WIP automatic collection of images through stable diffusion's text2img, hopefully with automatic parameter tuning
• Automatic collection of Microsoft Session cookies using WebDriver
• Displaying these images on Discord for remote labelling
• Labelling these images from both remote (Discord) and local (tkinter) for training
• Using Discord for easy, persistent remote access of images
• use pandas and matplotlib for displaying statistics locally and on Discord
• WIP training of Keras classifier
• WIP training of Stable Diffusion LoRA (LoRA not included in this repo)

Setup

Warning

There is no .exe file because I am a stupid smelly nerd

It is recommended that a venv is used as there are a lot of dependencies (50>)

python setup.py

Bot

Requires your bot token in .env, follow the setup guide

python sucorn.py

Data collection phase

Warning

Use this ethically, as this makes requests to BingImageCreator. Avoid using too many instances as it will result in executing a DoS attack on this host.

utilities/reverse_api/get_cookie.py

Load account names in the normal list in usernames.json and run this file. This will automatically get the session cookie for each account and store it into cookies.json. Credentials are taken from .env. The loaded list contains account names with the cookie acquired and stored

utilities/reverse_api/run.py || run_linux.py

Note

Replaces all double quotation marks in prompt with single quotation marks in Linux, keyboard ctrl shift r does not work either

Creates n-number of instances that will use the reverse engineered API to generate images from imagedir/prompt.txt. n-number of instances depends on how many pairs are found in cookies.json If imagedir does not exist then utilities/reverse_api/prompt.txt will be used to create imagedir/prompt.txt

python run.py imagedir --delay 10 --max 20

Each instance here will have a cooldown of 10 seconds after downloading all images from the query before

Keys:

• Tab: Brings the child processes to front
• End: Terminates all child processes
• Ctrl + Shift + R: Reloads prompt from utilities/reverse_api/new_prompt.txt. Currently unavailable

If it doesn't work

Try changing the parameters located in BingImageCreator.py TODO: instructions here

utilities/reverse_api/.env

WIP

GENERAL=your_password_here
speckly=your_password_here

utilities/automation.py

Warning

This feature is deprecated, use utilities/reverse_api

Makes use of pyautogui to automatically navigate the GUI, using mouse clicks and keyboard presses. The output is links pasted in a Discord channel. Run the Discord slash command /nuclear_cat copy=<copy channel>

Data preparation phase

features/sucorn_statistics.py

This is both a feature and utility, feature used for the /statistics slash command in Discord, utility can be used in the command line.

python sucorn_statistics.py

Providing no options will iterate through all subfolders in ./images

python sucorn_statistics.py --csv sample_statistics.csv --mode write

Writes these results out to a given csv file, ommitting the --mode flag will read the csv and display it using matplotlib

utilities/labelling.py

For supervised learning, we need to label the images

Warning

Do not run this with reverse_api/run.py, as there will be duplicate file names and overwriting will occur, will fix this later

labelling.py foldername

Ensure that foldername is a valid foldername in ./images The following keys will label the image by moving it into its subfolder

• 0 for negative
• 1 for positive
• 2 for neutral
• ESC to quit

utilities/reverse_imagen3/imagen3.py and get_auth.py

Note

With effect from 3 September 2024, you no longer need to collect the Authentication header value, please follow the new steps below

Reversed API for Google's ImageFX also known as Imagen 3

Login to ImageFx, and then run

python utilities/reverse_imagen3/get_auth.py <variable_name>

This will collect your cookies and store it in .env in the same directory as these files. The variable name is a way of differentiating between different cookie strings, therefore allowing the use of multiple accounts. See the example with run_xfce4.py. Suggest to name your variable name similar to the username you use for your Google account.

This will retrieve the cookie string from .env and use it to get an access token which can be used for generating images for the day. You can expect to generate about 50 images daily before getting a HTTP 429, stopping you from generating more for the day.

python utilities/reverse_imagen3/imagen3.py <variable_name>

sudo python run_xfce4.py <foldername> (more versions coming soon) will create multiple instances depending on how many variables are found in .env A default of 8 images maximum will be returned per request

If a HTTP 401 is returned with the following message: Request had invalid authentication credentials., it most likely means that the access token has expired. Follow the steps above starting from get_auth.py. For now, you will have to manually remove the string from .env. TODO: automate this

Occasionally a HTTP 400 with message: Request contains an invalid argument is returned. It could mean that inappropriate content has been generated and therefore is not sent back to the client.