Question for source and target

[planemanner]

Hi, thank you for your great work sharing !

I had question while reviewing your work and source code you shared.

Apologies if this sounds silly — I’m just starting out in image editing research.

Question 1

• Why source and target points must be same even if source and target regions can have different area sizes ?
  • Examining this line, I found all the source and target points sizes are same.

Question 2

• Why this mask should have 1 for all locations ?

Question 3

• If I arbitrarily create source and target masks, the number of points in each mask may differ.
  How can I make them have the same number of points while still reflecting your intended design ?

Question 4

• Actually, I'm trying to make an application which progressively generates a new new part of an original image.
  Have you tried this approach ? And, Did it work well ?

Thanks !

[LuJingyi-John]

Thank you for your interest in our work! I appreciate you taking the time to review our code so thoughtfully. Here are answers to your questions:

Q1: Why source and target points must be the same

Answer: The one-to-one mapping between source and target points creates a direct correspondence that simplifies transformations. This approach helps our algorithm create more natural-looking results while maintaining the relationship between where points begin and where they end up.

Q2: Why the mask has all 1s

Answer: The mask defines which areas are editable in the image. We set it to all 1s to make the entire selected region fully editable, which provides a more straightforward user experience.

Q3: Handling different numbers of points in masks

Answer: Our implementation in this section of the code addresses this issue. The algorithm automatically finds corresponding source points for each target point, so you don't need to worry about ensuring they have equal counts.

Q4: Progressive generation

Answer: I'm not entirely sure which approach you're referring to. There are two possibilities:

1. Autoregressively generating image patches - this isn't something we've extensively explored in this project
2. Visualizing the diffusion denoising process - which can be accomplished by using a VAE to decode latents at various stages

I'd be happy to discuss your specific application further if you'd like to share more details!