How to reproduce the demo video results.

[liuquande]

First of all, thanks for you great efforts on maintaining this project.

The outpaining results on my own images are relatively bad, so I am trying reproduce the reuslts of the demo video to make sure my implementation is correct.

However, I find that when using the same init image and prompt (with SD 1.4 ckpt), I cannot get results as good as the authors, see the figure below:

So I would like to ask, for the demo results in the video, did the author get pretty good results after each "outpainting click", or try outpainting for each place many times, then cherry pick the best one to make the video?

Many thanks.

[lkwq007]

The video is edited to remove safety-checker affected results (e.g. I intended to outpaint a wooden closet on the left, but the safety-checker was triggered. so you may notice a selection box shift at the second outpainting) and reduce the length to 30s (e.g. chandelier was tried multiple times). However, the girl and the right part of the image are generated continuously.

Please note the all generations are random, so it may be difficult to reproduce a certain generation sequence.

---

Update: For completeness, the demo video is not the current one in the readme. The video mentioned in this issue can be found at https://github.com/lkwq007/stablediffusion-infinity/blob/photometric-correction/readme.md