Questions: anchors, model's forward and code ETA

[Jay-D13]

Thank you for your previous response regarding the VLM. It was very helpful in clarifying certain aspects. I have a few additional questions, and I hope you don’t mind providing further insights:

1. Could you elaborate on the reason for selecting only the last anchor during inference (instance_feature[:, 900:, :])? Is it purely to serve as the global condition for the diffusion process, or is there another purpose?

2. In the provided implementation, some attention mechanisms and feature refinement steps seem to be commented out. Is this a simplification for testing, or does it reflect the final design in the paper? If it’s a simplification, could you share the rationale behind excluding attention in this version?

3. If it’s not too much trouble, could you share which conference the paper was submitted to? This would help anticipate the potential release date for the full code, as mentioned in your other communications

Thank you again for your time and for sharing your amazing work. Greatly appreciate your support!