[Question] Potential color inconsistency: ImageNet Normalization inputs vs. Raw 3DGS output

Hi, thanks for the impressive work on Depth Anything V3.

I have a question regarding the color reconstruction consistency. The model input requires **ImageNet normalization** (mean/std), but the 3DGS head is expected to render **raw RGB colors** to compute photometric loss against the original un-normalized images.

Does this design force the network to implicitly learn "color denormalization" from the feature space? Have you considered explicitly feeding the **raw, un-normalized image** to the Gaussian head? This might provide better color guidance than decoding raw 3DGS colors from normalized features.

Thanks!