Telegram Bot using Fast Neural Style Transfer

This repository contains Telegram Bot implementation, that uses for Style Transfer. Style Transfer implementation entirely based on Pytorch Fast Neural Style implementation, shown below.

PyTorch Fast Neural Style Example

Bot

You can find this Telegram Bot after searching for @StatixX_StyleBot. This repo contains other styles, besides the standart ones, contains in Pytorch Fast Neural Style repository, that has been moved into artwork directory.

Requirements

torch
torchvision
pyTelegramBotAPI
pillow

Available styles

psychodelic
abstraction
artwork
photo style

Available style photos store in the style folder, so you can find them there or use the bot, which will show you all styles it has.

Usage

Click start
Upload an image when the bot asks you
Choose the style you want
Wait for the stylized image
Click "Finish" if you're done.

Learning process

TransformerNet was taken from Pytorch Fast Neural Style github repo and was trained on COCO dataset with 2 epoch on (128 x 128) image size. The train_styles.py trains models for all styles, contained in the style folder iteratively.

Examples

Example 1:

Example 2:

Quality

For the purpose of make possible deploy this app on heroku, the image size of pictures will be decreased to 512 px in width. If it'll be a purpose to get better results, you can clone git and perform stylize on your own.

Config

Config file was not added as it contains API key for TelegramBot