Research Context​​

​​Objective​​:

To verify whether a diffusion model can generate images that match target features when conditioned on two paired input features(initially simplified by using features extracted from the same source image).

• •​​Ideal outcome​​: The model should generate images whose features align with the original image's features.
• •​​Current setup​​: In real-world scenarios, the correspondence between features and original images is unknown. This experiment uses paired featuresas a simplified proxy.

​​Methodology​​

• •​​Input​​: Two paired features extracted from the same image (e.g., Feature A + Feature B).
• •​​Control​​: Constrain the generated image’s features to match the target feature (e.g., Feature A) via similarity metrics (L2 loss or feature-space distance).
• •​​Model​​: Modified DDPM architecture with dual-feature conditioning.

​​Observations​​

​​Attempted Solutions​​

• •Adjusted loss weights (feature-matching loss vs. diffusion noise).
• •Tested different noise schedules (linear vs. cosine).

​​Key Questions​​

​​Model Selection​​

​​Feature Representation​​

​​Requests for Advice​​