This is the updated version of Ambifusion. Our proposed model can generate ambigram images from two different text conditions. You can see the details of ambifusion(ver.1) in this paper: ArXiv.
In the flowing experiments, the pretrained large-scale Text2Img-diffusion models are used in our proposed ambigram generation modules.
Therefore, The model can generate a variety of ambigram images, not only alphabet letter pairs but also diverse image pairs.
In the following demo codes, we use deep-floyed/IF models as generation modules in the reverse process.
♦ "A↕Y" ambigrams with different styles
Set up diffusers environment.
- Start gradio web app as following.
python demo.py
- Access
127.0.0.1:11111
with your web browser.
- Set
TestConfigs
inambigram_sample.py
. - Run the sampling code as following.
python ambigram_sample.py
- Set
TestConfigs
inambigram_sample_hr.py
. - Run the sampling code as following.
python ambigram_sample_hr.py
You can change base generation model to other models such as StableDiffusion
.
But you might change some of the codes in the ambigram_pipeline.