Dreambooth SDXL example inconsistent results

[ComandoLando]

Describe the bug

When running the dream booth sdxl code exactly as per the example readme, the results generated are not consistent.

I have had a look at others users push to hub uploads and it appears others have the same issue, i have seen people uploading the same example out of the box and all of the dogs look different. Looks like an issue for the past few months.

Reproduction

Just run exactly as per readme and verify results are not consistent.

Logs

No log issues, just final result inconstancy with random dogs getting generated.

System Info

Windows 11, python 10.8

Who can help?

No response

[sayakpaul]

Can you ensure you're on the latest version of the script? We performed some changes to fix these issues. Also check out: #6119 (comment).

[ComandoLando]

Thank you for getting back to me. I just have cloned and rerun a new branch at a new location. Using the same input as the readme.

This was the results still somewhat nconsistent: https://huggingface.co/Landon69/lora-trained-xl

You can see here https://huggingface.co/models?search=lora-trained-xl lots of people running the example, most of them are not the same clean constant results as the readme file.

Hope this helps

[sayakpaul]

Not sure still as it works for me: https://huggingface.co/sayakpaul/corgy_dog_LoRA.

Colab: https://colab.research.google.com/gist/sayakpaul/9615b89369f3ef23cc29d0dac58253dd/scratchpad.ipynb

I would suggest a clean clone of the diffusers repository and a fresh from-source installation and retrying again.

[ComandoLando]

I did a complete re-clone last night of everything and spent two hours setting up a fresh environment.

Looks like lots of people have the same issue. I'm using windows 11 - 4090. If it works for you might, it might be an environmental issue.

Given that your collab works.

https://huggingface.co/zhongzero/lora-trained-xl

https://huggingface.co/AvesM/lora-trained-xl

https://huggingface.co/GordonMcGregor/lora-trained-xl

https://huggingface.co/formido/lora-trained-xl

https://huggingface.co/Salatof/lora-trained-xl

[ComandoLando]

I have also had a look at the colab, this is not the same as the readme file. I am following the readme exactly assuming it will generate the expected results.

[sayakpaul]

Does this work for you? #6432

[ComandoLando]

Yes, that started generating some better results. Thank you.