Sekai-Real-HQ used in Yume 1.0 vs 1.5

[flt-ime]

Thank the authors for the great work!

I have a question regarding the Sekai-Real-HQ dataset used in Yume 1.0 and 1.5 - aside from that version 1.5 used InternVL3-78B to generate event-aware descriptions, are there any other differences? If possible, could you also share the specific prompts used?

Thanks in advance!

[stdstu12]

We synthesized a T2V dataset on Wan2.1 14B using prompts from OpenVid. We also asked volunteers to write event descriptions to synthesize an I2V dataset on Wan2.2 14B. I expect these datasets to be released on Hugging Face in batches.

[flt-ime]

Hi @stdstu12, thanks for the clarification! I understand the release process may take time - releasing training data is already greatly appreciated! In the meantime, would it be possible to share the specific prompts or even the scripts used to construct the event-aware descriptions? I'd like to better understand how the event-aware descriptions are formulated.

+) Just to double-check that with the Sekai-Real-HQ dataset currently available on Hugging Face and the new event-aware descriptions, it is sufficient to runtrain_5b.py given in #16, right?

Thanks again for your time and for maintaining this project!

[flt-ime]

Hi @stdstu12, I just noticed that Sekai-Real-HQ is described as including both walking and drone videos (the YUME 1.0 paper states that "there are 8409 hours of walking videos and 214 hours of drone videos after downloading"). However, at the moment, only the walking subset (sekai-real-walking-hq - https://huggingface.co/datasets/Lixsp11/Sekai-Project/blob/main/sekai-real-walking-hq.zip) is available.
Could you clarify whether the full version (sekai-real-drone - https://huggingface.co/datasets/Lixsp11/Sekai-Project/blob/main/sekai-real-drone.zip) was directly adopted as sekai-real-drone-hq, or if the HQ drone subset has not been released yet?

[Aurora-edu]

@stdstu12 Great work! Thanks for your plan to release dataset! 2 months passed, Will the dataset be released recently?