Request for Combined Dataset Used in ReconX Training

[Yang-xiao-uts]

Dear authors,

Thank you for the impressive work on ReconX — it’s a fascinating and promising approach for sparse-view 3D scene reconstruction.

I have a quick question regarding the training setup:
In the paper, you mention that the video diffusion model was trained on 3D scene datasets (RealEstate-10K, ACID, and DL3DV-10K). It seems that these datasets were used in combination for a unified training process (rather than separate training on each dataset).

Would it be possible for you to provide the combined version or the preprocessing script used to merge and sample from these datasets for training?

Having access to the same setup would be incredibly helpful for reproducing the results and for further research on cross-dataset generalization.

Thank you in advance, and looking forward to your open-source release!

[Yang-xiao-uts]

In addition, I would like to ask a few technical details about the training configuration:

What batch size did you use during the fine-tuning of the video diffusion model on the 8× NVIDIA A800 (80G) GPUs?

Do you think it is feasible to fine-tune the video diffusion model on three L40 GPUs (48GB each)?