current released VideoWorld 2 dLDM pre-training data

[zef1611]

Hi @MaverickRen,

Thank you for releasing this amazing work! I can say this is one of the best papers I’ve read this year.

I would like to confirm whether the currently released dLDM checkpoints were pre-trained on both the Video-CraftBench and OpenX datasets, or only on Video-CraftBench.

Thanks again for the fantastic work, please keep up the good work!

[MaverickRen]

Really appreciate the positive feedback! The current dLDM weights are for Video-CraftBench, though you can directly use the code to train on OpenX as well. We're working on releasing models trained at larger scales—will keep you posted！

[zef1611]

@MaverickRen Thank you for the prompt response. I’m looking forward to the full release of this awesome work!

I have one more quick question: given the pre-trained dLDM weights on Video-CraftBench, can I continue fine-tuning them on RLBench dataset, or would I need to pre-train the model again to extract the best latent dynamic codes for AR transformer training? I would deeply appreciate any guidance you can provide.