Hi, thank you for releasing the project.
While reviewing the paper and the current codebase, I noticed that the implementation seems to be missing the Path Diffusion module described in the paper.
According to the paper, RFGen consists of three diffusion-related components:
-
Object Diffusion
Generates target object geometry and motion from RF sensing prompts.
-
Environment Diffusion
Generates environment-related PIR representations, such as structured multipath or ambient noise.
-
Path Diffusion
Refines the mixed PIR produced by combining the path tracer output and the environment diffusion output.
From my understanding, Path Diffusion is important because it acts as an image-to-image refinement module. It takes the imperfect mixed PIR as input and improves the realism of the RF path representation before the signal generator produces the final radar signal. The paper also mentions that ControlNet is used to condition the diffusion process on the mixed PIR.
However, I could not find the corresponding implementation for this module in the released code. In particular, I could not locate:
- the Path Diffusion model definition,
- the ControlNet-based refinement pipeline,
Could you clarify whether the Path Diffusion implementation is included in the repository? If it is not currently released, are there any plans to release it in the future?
This module is important for reproducing the full RFGen pipeline described in the paper, especially for improving the realism of synthetic multipath and reducing the sim-to-real gap.
Thank you.
Hi, thank you for releasing the project.
While reviewing the paper and the current codebase, I noticed that the implementation seems to be missing the Path Diffusion module described in the paper.
According to the paper, RFGen consists of three diffusion-related components:
Object Diffusion
Generates target object geometry and motion from RF sensing prompts.
Environment Diffusion
Generates environment-related PIR representations, such as structured multipath or ambient noise.
Path Diffusion
Refines the mixed PIR produced by combining the path tracer output and the environment diffusion output.
From my understanding, Path Diffusion is important because it acts as an image-to-image refinement module. It takes the imperfect mixed PIR as input and improves the realism of the RF path representation before the signal generator produces the final radar signal. The paper also mentions that ControlNet is used to condition the diffusion process on the mixed PIR.
However, I could not find the corresponding implementation for this module in the released code. In particular, I could not locate:
Could you clarify whether the Path Diffusion implementation is included in the repository? If it is not currently released, are there any plans to release it in the future?
This module is important for reproducing the full RFGen pipeline described in the paper, especially for improving the realism of synthetic multipath and reducing the sim-to-real gap.
Thank you.