video-acceleration
video-world-acceleration-qualitative
Acceleration quality ladder (qualitative)
Videos showing what WorldCache (whole-step diffusion caching) does to a video
world model as it is pushed from a quality-preserving operating point all the way
to collapse. Two cells of a robot video-generation benchmark, same episodes at
every rung, so identically-named files are directly comparable.
Backbone: Cosmos-Predict2 2B DiT, post-trained on a GR1/AIRBOT humanoid
manipulation corpus, 480p, 35 denoising steps, autoregressive… See the full description on the dataset page: https://huggingface.co/datasets/doanh25032004/video-world-acceleration-qualitative.agent_video_acceleration