Question about Identity-RoPE / Ge-RoPE implementation and their interaction

[PengJingchao]

Hi, thanks for releasing the code.

I have a question about the implementation of the two RoPE variants described in the paper: Identity-RoPE and Ge-RoPE.

According to the paper, Identity-RoPE seems to use mask/bounding-rectangle based local normalized coordinates for object regions, while Ge-RoPE uses displacement / flow and confidence to build geometry-aware warped positional encodings.

However, when checking the current public code, I could not locate where these two modules are actually implemented.

In diffsynth/models/wan_video_dit.py, the main RoPE frequencies appear to be constructed from the standard (f, h, w) token grid:

freqs = torch.cat([
    self.freqs[0][:f]...
    self.freqs[1][:h]...
    self.freqs[2][:w]...
], dim=-1).reshape(f * h * w, 1, -1)

Then the block is called as:

x = block(x, context, t_mod, freqs, vggt_tensor=kwargs.get("vggt_tensor", None))

So it looks like flow_tensor is not passed into the DiT blocks here.

Inside DiTBlock.forward, I noticed the signature includes:

def forward(self, x, context, t_mod, freqs, vggt_tensor=None, freqs_vggt=None, flow_tensor=None):

but I could not find where flow_tensor is used to generate warped RoPE frequencies, or where the mask/bbox-based Identity-RoPE positional remapping is computed.

I also saw that VGGT tokens are appended to the latent tokens before self-attention, which is clear. My question is specifically about the RoPE variants:

1. Where is Identity-RoPE implemented in the released code?

• Is there a mask → bbox → local normalized coordinate → RoPE frequency remapping step?
• If yes, could you point to the relevant file/function?

2. Where is Ge-RoPE implemented?

• Is flow_tensor used to warp the spatial grid / positional frequencies?
• If yes, where is the displacement resized, normalized by patch size, smoothed, and added to the original grid?

3. How are Identity-RoPE and Ge-RoPE combined?

• Does Identity-RoPE define the base positional coordinates and Ge-RoPE further warps them?
• Or are they used in separate attention heads / separate frequency channels?
• If object-local normalized coordinates are used for Identity-RoPE, how are they made compatible with the standard token-grid coordinates used by Ge-RoPE?

4. Is the current public inference code a simplified version that only uses VGGT token concatenation, while the full Identity-RoPE / Ge-RoPE implementation is in the unreleased training code?

Thanks in advance. I may have missed the relevant implementation, so a pointer to the exact file/function would be very helpful.