Preface
One of my bad qualities is that I never give up. Since it bothered me that I could not create a laying down bottle with some content using a LoRA model, I came back to this problem today to see how it could be possibly realized.
I summarize the result of my efforts in this article and publish later on the LoRA model developed accordingly. From my point of view, the results I achieve are getting better and better.
Presumptions
Due to some hardware and software limitations, I am still using Stable Diffusion SD 1.5 for my work an the research of the topics of interest.
Goal
I am pursuing two goals. First goal is the creation of a photorealistic laying down bottle. Second goal is putting content of different kind in the laying down bottle. As the last few days have shown, this is more difficult than expected.
One should be able to exchange the content inside the bottle by other stuff.
Training
I prepared a set consisting of 4 images for the training to perform. Therefor I created different bottles. By overlaying I added a red apple to the images.
Figure 1: Training images
I used buckets and mirroring to get a set of 32 images for the training. A caption text file is not used. The file name are of the type
bottle apple [1..3].jpg
The intention using this file name was , that I should after training be able to remove or exchange the apple from created images.
For my first test I limited the iteration steps to 1500. The training took round about 20 minutes.
First Results
The first created images looks good. The results are at least now heading in the right direction.
Figure 2: Marbled red ball in the bottle
Figure 3: Orange in the bottle
Figure 4: Butterfly in the bottle
Figure 5: Pumpkin in the bottle
Figure 6: Planet in the bottle
Figure 7: Empty bottle
Learning Rate Curve
The curve of the learning rate is shown for the 1500 iteration steps I used to train the new LoRA model.
Figure 8: Learning rate curve
Conclusion
Despite some limitations, I can now produce horizontal bottles more reliably. I have not yet been able to bottle certain things, such as cars, trains or ships.
Final Words
Have a nice day. Have fun. Be inspired!