Discrepancy between paper and notebooks regarding spatial modification phases

[mirkozaff]

Hi,
great work on the paper! I’m trying to reproduce the spatial intervention results using spatial_intervention_{timestep}_stage.ipynb, but I'm a bit confused and wanted to double-check my understanding with you. While the paper states that spatial control happens in the early stages, my results with the notebook suggest the interventions are taking place in the late phase. I suspect I might be interpreting the timestep definitions incorrectly in the code.

Would you mind clarifying how the phases in the code relate to the paper to help me reproduce the results correctly?

Thanks in advance!

[berktinaz]

Hello,

Thank you for your interest in this work! I apologize for the confusion (and the late response). As you pointed out, there is a discrepancy between the phase definition in the paper and in the code. This is due to whether the early/late definition is taken with respect to the forward diffusion process or the reverse diffusion process.

Here is the explicit mapping between the timesteps_to_cache parameter in the code and the phases described in the paper:

• Early phase: timesteps_to_cache = 1000
• Middle phase: timesteps_to_cache = 500
• Final phase: timesteps_to_cache = 0

I will fix the file naming as soon as possible. I hope this helps in the meantime.

Thanks!