ablattmann

Thanks for your interest in our work! We will also release our text-conditional model in the near future.

We used the [transformer implementation of lucidrains](https://github.com/lucidrains/x-transformers), which includes lots of different positional encodings. As we never used relative position encodings, we always set it to `None` to avoid adaptations...

please close if this answers your question :)

We added a [colab GUI-demo](https://colab.research.google.com/drive/1sec1I_80SpG6ielaSE0n_AcI3cEKzk0l?usp=sharing) which is based on `streamlit`. However, since it's currently not possible to dynamically manipulate images with `streamlit`, we created a workaround based on the [`streamlit-drawable-canvas`](https://github.com/andfanilo/streamlit-drawable-canvas)...