A.F.R.O. LOGOBOT
Automotive Fast Responsible Open source
“Hello! My name is AFRO. I’m here to help you to quickly generate a logo for your company using a few words.”
PROJECT OBJECTIVES
SAVE MONEY
OPTIMIZE TIME
BE VARIOUS
INSIDE THE MODEL
User inputs text
Image denoising and super resolution
StyleGAN2 takes random vector and returns image
Bot returns image blended with text
01
03
02
04
GAN IN ESSENCE
GAN consists of a generator and discriminator. The generator tries to create an image that will be accepted by the discriminator
At the same time, discriminator tries to decline generated images, which do not look real
GAN TRAINING
It hurts...
First attempts
Some progress!
Not bad
iWGAN-GP - https://arxiv.org/abs/1704.00028
StyleGAN2
If it even can generate waifu for you..
StyleGAN2 - https://arxiv.org/abs/1912.04958
OUTPUTS
SUPER RESOLUTION
32x32
128x128
https://arxiv.org/abs/1809.00219
https://pypi.org/project/ISR/
TEXT TRANSFORMS
https://github.com/albumentations-team/albumentations
https://opencv.org/
What didn't work
MONETIZATION
Advertising sales in the bot messages
PHASE 1
PHASE 3
Propose subscription for users with advertisement removal
PHASE 2
Participation in the government orders (tenders)
FUTURE WORK
Decrease the time for image generation
OPTIMIZATION
Share the bot at the tg-channels
COVERAGE
Increase the image reality and resolution
QUALITY
Prepare for the 1st investment presentation
INVESTMENT
Prepare clusters for input text embeddings
CLUSTERING
TRY IT YOURSELF�(t.me/LogoAfroBot)
TEAM
01
03
02
04
05
ALEKSANDR FROLOV
Team management, text to embeddings, image preprocessing
RODION BUKHANEVICH
Research using CycleGAN, clusterization, data preprocessing
KIRILL BUROVIN
Research using iWGAN-GP, StyleGAN2, SRGAN, literature review
BOGDAN TYSHCHENKO
Telegram bot, user interface, model release to production
ALEKSANDR ISAKOV
Image postprocessing, MVP, presentation, literature review
Contacts:
@kupchanski @AllexFrolov @rd_magic @jityshchenko @asisakov
Last slide behind