Hi authors, thanks a lot for open-sourcing your great work! I’ve been studying your code and have a few questions about the training process, hoping to get your insights.
In the training loop, I noticed that to obtain the Agent Token ($\hat{x_0}$) for interacting with the traffic scene encoder, the model starts from random Gaussian noise and performs full denoising sampling with 50 diffusion steps. This seems to introduce noticeable training time costs—since each iteration requires 50 denoising steps, the time complexity is significantly higher than a standard approach that computes $\hat{x_0}$ in a single step. I didn’t find obvious further speed-up strategies (like reduced steps or optimized schedulers) in the code. Do you have plans to optimize this part in future work? Additionally, I’m working on a similar task where I need to use $\hat{x_0}$ for some constraints. I’m wondering if an alternative approach is feasible: instead of full 50-step sampling, compute $\hat{x_0}$ directly using the predicted noise $\varepsilon_\theta$ and the forward noising formula (i.e., $\hat{x_0} = \frac{x_t - \sqrt{1-\bar{\alpha}t} \cdot \varepsilon\theta}{\sqrt{\bar{\alpha}_t}}$). Would this method preserve the necessary properties of $\hat{x_0}$ for downstream interaction with the scene encoder?
Thanks again for your time and contributions!
Hi authors, thanks a lot for open-sourcing your great work! I’ve been studying your code and have a few questions about the training process, hoping to get your insights.$\hat{x_0}$ ) for interacting with the traffic scene encoder, the model starts from random Gaussian noise and performs full denoising sampling with 50 diffusion steps. This seems to introduce noticeable training time costs—since each iteration requires 50 denoising steps, the time complexity is significantly higher than a standard approach that computes $\hat{x_0}$ in a single step. I didn’t find obvious further speed-up strategies (like reduced steps or optimized schedulers) in the code. Do you have plans to optimize this part in future work? Additionally, I’m working on a similar task where I need to use $\hat{x_0}$ for some constraints. I’m wondering if an alternative approach is feasible: instead of full 50-step sampling, compute $\hat{x_0}$ directly using the predicted noise $\varepsilon_\theta$ and the forward noising formula (i.e., $\hat{x_0} = \frac{x_t - \sqrt{1-\bar{\alpha}t} \cdot \varepsilon\theta}{\sqrt{\bar{\alpha}_t}}$). Would this method preserve the necessary properties of $\hat{x_0}$ for downstream interaction with the scene encoder?
In the training loop, I noticed that to obtain the Agent Token (
Thanks again for your time and contributions!