YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Movens-Zero training workspace
This repository is the server-local workspace for the absolute-height startup continuation stage.
The only supported training entry is:
source /cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/code/Movens-Zero_train/train_8gpu_absolute_height_startup.sh
The launcher restores the configured full training state and writes new snapshots, logs, and checkpoints below runs/. Runtime checkpoint and visualization paths are configured in config/runtime.json; the launchers reuse the existing Python environment and public Qwen model. See PROJECT_GUIDE_ZH.md for the migration record, training/data flow, and external dependencies.
The actual training data is stored below training_set:
training_set/
text/ # text-conditioned G1 motion and indexes
music/ # music-conditioned G1 motion, audio, and indexes
speech/ # speech-conditioned G1 motion, audio, and indexes
video/ # video-conditioned G1 motion, media, and indexes
config/ # four-modality runtime configuration
integrity.json # repository-relative manifest integrity checks
Validate the local dataset without allocating GPUs:
python -B -u prepare_training_set.py
Start the unified training dashboard, dataset comparison page, and automatic checkpoint inference/rendering:
source start_visualization.sh
Before the first training run exists, the same command still serves the dataset page. After training creates runs/latest_training_run.json, restart it to enable curves and automatic checkpoint cases.