Repository navigation
Add a section about how training data resizing might affect the quality of the end models #6397
Description
Have been chatting with @bghira on our Discord forum about the negative effects of not carefully resizing the training data.
@bghira noticed that the way we resize the images in our training examples can introduce unwanted artifacts in the samples generated by the end fine-tuned model. In short, it ruins the fidelity of the generated samples.
We don't want to make the data input pipelines in our training examples super sophisticated, guaranteeing SoTA results. But that said, I think it's best in the community's interest if we made a note about this phenomenon in the READMEs and just added a comment about it in the pipelines.
Curious to know what @patil-suraj @apolinario think about this. Personally, I think it makes sense.
I will let @bghira share some more insights as well.
Activity
github-actions commentedon Jan 29, 2024
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
sayakpaul commentedon Jan 29, 2024
Not stale.
github-actions commentedon Mar 16, 2024
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
github-actions commentedon Apr 13, 2024
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
github-actions commentedon Dec 16, 2024
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
github-actions commentedon Jan 10, 2025
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
8 remaining items
github-actions commentedon Mar 1, 2025
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
crapthings commentedon Mar 2, 2025
What do the 'unwanted artifacts' look like?
github-actions commentedon Mar 26, 2025
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
bghira commentedon Mar 26, 2025
@crapthings they look like JPEG artifacts or resize artifacts, it could be something like a grid of pixelation if you use bicubic / nearest neighbour interpolation on downsample.
bghira commentedon Mar 26, 2025
https://zuru.tech/blog/the-dangers-behind-image-resizing
bghira commentedon Mar 26, 2025
@stevhliu maybe since you're the docs expert we can bypass process and start working on what this would look like? here's the reference from Zuru tech blog;
stevhliu commentedon Mar 27, 2025
Yeah that sounds good to me. Maybe we can make a note of it in the training Overview since that is also where we introduce our training scripts.
Would you like to open a PR for this @bghira or I can too once I've wrapped up some other things 🙂