š¼ Trained the Blur Background Kontext LoRA with AI Toolkit on an RTX 3090 using ML-Depth-Pro outputs.
Results looked good after 250 steps (~25 mins).
When combined with Kontext Lighting, it outperforms ML-Depth-Pro in just 6ā8 steps.
Thanks to @ostrisai ā¤ @bfl_ml ā¤ ML Depth Pro Team ā¤
Thanks to Steve ā¤ for the GPU resources. (v2)
š§¬code: https://github.com/ostris/ai-toolkit
š§¬code: https://github.com/apple/ml-depth-pro
š¦blur background: https://civarchive.com/models/1809726
š¦unblur background: https://civarchive.com/models/1812015
šNote: Blur, trained with Gaussian blur, works with Gaussian blur. If you have pairs of blurred and sharp images from a camera lens (same image different aperture), I'd love to use them for training.
Enjoy! ā¤
Description
FAQ
Comments (2)
Thank you for your work, but unfortunately this model makes an unpleasant halo around the object. Perhaps you had it in your training data too?
To make a high-quality background blur, it is not enough to simply cut out and blur the background in photoshop. You also need to eliminate the halo on the bottom layer, which is difficult.
Blur, trained with Gaussian blur, works with Gaussian blur. If you have pairs of blurred and sharp images from a camera lens, I'd love to use them for training.