Fail to reproduce the results on re10k

[fangchuan]

I've attempted to reproduce the results on the RealEstate-10K dataset following the [DFoT](https://github.com/kwsong0113/diffusion-forcing-transformer) protocol (**measure the difference between the generation and gt dataset**), but the generated outputs show significant discrepancies from the reported numbers in the paper.

Setup and preprocessing

• Randomly selected 150 scenes as test set following DFoT protocol

• Annotated test set with [ViPE](https://github.com/nv-tlabs/vipe) to obtain metric depth videos and camera poses

• Generated detailed prompts using Gemini-3-pro from the given video

• Adapted data_engine/create_input.py to match expected input conditions

Evaluation metrics

• FVD, PSNR, LPIPS, SSIM for comparing generations against ground truth

Results comparison

Here are the input and output:

video_voyager.mp4

The result shows that the generated video contain sever color shift and restoration shift. Could you release the evaluation code on re10k, or help me debug the reproduction problem?