How data generated?
Hi ,your work is excellent! I am pretty curious abot how you get dataset? It would be great if you like to share with us ,many thanks!
EDIT: My bad, thought I was in the comments for a different LoRA that adds objects.
ORIGINAL:
I'm just some passerby, so I can't tell you for certain how they did it, but I can give you two plans of action for creating a dataset for this LoRA.
- Set up a digital camera on a tripod so it doesn't move. Take a picture. Physically add an object in frame. Take another picture. Use an image editor to cut the object from the second image and paste it into the first, so that it is there but not casting shadows. This is your "Before" image. Your "after" is the real life image with the object in it.
or
- Use Kontext to do the same process with an object removal LoRA. Cut the object out, create an image without the object in it, paste the cut-out object into the altered image. This is your "before", the after is the original image.
I'd say the first one is the less time consuming and would produce better results, if you have a camera and a tripod. Oh and just thought of a bonus third option: Do it with a 3D rendering program, or a video game.
I created day and night version of same image and masked the day version of the object on top of night version and used day version as output. I used grounding dino for masking. I also do the reverse of this. And collected aprroximately 40 image pairs with this strategy.