Geral-Yuan/actionWM_vfx_h3
ActionWM VFX — 600 H3 videos 600 synthetic VFX videos generated with MiniMax H3 Max text-to-video and balanced prompt expansion. Each unchanged original videos/TASK_ID.mp4 matches prompts/TASK_ID.txt. The TXT and viewer text column contain the exact raw prompt submitted before the model's server-side prompt expansion. Category First person Third person Total Global environment 120 120 240 Local object 120 120 240 Character effects 0 120 120 Total 240 360 600… See the full description on the dataset page: https://huggingface.co/datasets/Geral-Yuan/actionWM_vfx_h3.
ActionWM VFX — 600 H3 videos
600 synthetic VFX videos generated with MiniMax H3 Max text-to-video and balanced prompt expansion. Each unchanged original videos/TASK_ID.mp4 matches prompts/TASK_ID.txt. The TXT and viewer text column contain the exact raw prompt submitted before the model's server-side prompt expansion.
There are 36 effects, each with ten scenes per applicable viewpoint. Third-person prompts request full-body rear gameplay framing with a continuously following camera. The stone-arch reconstruction prompts additionally require persistent stone trajectories, a fixed assembly site, and no camera resets at timeline boundaries.
Generation requests: 5 seconds, 768P, 16:9. Actual dimensions, frame rate and duration are provided per row; videos retain the original H3 frame rate and audio if present. Existing outputs are reused only when their exact prompt, seed and request parameters match this final plan. These are independently generated event clips, not paired source/edited videos.
All files are checked for prompt pairing, hashes, video decoding and basic video specifications. An adjacent-frame scene-change heuristic marks continuity candidates for review; VFX changes can also trigger it, and a negative result does not guarantee continuity. The dataset has not received exhaustive visual acceptance. The train split name is a storage convention.