Dancing Stick Figures
A lightweight video generation dataset for educational purposes, containing stick figure dancing animations. It is available in 64×64 and 128×128 configurations and includes multi-task annotations such as skeletal joint points, depth maps, and normal maps, accompanied by a free Colab training pipeline.
Timeline (last 90 days)
Developer sprited-ai shared the educational dataset project named Dancing Stick Figures on Reddit.
The Dancing Stick Figures dataset contains 1,430 six-second short video clips, totaling 514,800 annotated frames.
The dataset offers a 64×64 mini configuration with a total size of only 0.85 GB, designed specifically for free Colab.
The dataset also offers a 128×128 full configuration that retains richer information for advanced experiments.
The full data includes 2D and 3D positions and visibility labels for 27 joint points, depth maps, normal maps, part segmentation maps, camera parameters, and raw motion data.
Each limb in the dataset maintains a fixed color coding, enabling a NumPy scorer to automatically detect missing or disconnected limbs in model-generated outputs.
The training pipeline involves first training an image baseline model, using the image model to warm-start an eight-frame video baseline model, and ultimately generating 5.6-second inference outputs.
The warm-started baseline model reached the loss level of a model trained from scratch at 10,000 steps in only about 4,000 steps.
The author defines the project not as a finished video model, but as a small dataset and learning tool for hands-on building, dissecting, and understanding.
The dataset is licensed under CC0, and the code is licensed under MIT.
2 more timeline events
All Facts (12)
The code repository is located at sprited-ai/dancing-stick-figures on GitHub, and the dataset is located at sprited/dancing-stick-figures on Hugging Face.
50%UnverifiedThe Dancing Stick Figures dataset contains 1,430 six-second short video clips, totaling 514,800 annotated frames.
50%UnverifiedThe dataset offers a 64×64 mini configuration with a total size of only 0.85 GB, designed specifically for free Colab.
50%UnverifiedThe dataset also offers a 128×128 full configuration that retains richer information for advanced experiments.
50%UnverifiedThe full data includes 2D and 3D positions and visibility labels for 27 joint points, depth maps, normal maps, part segmentation maps, camera parameters, and raw motion data.
50%UnverifiedEach limb in the dataset maintains a fixed color coding, enabling a NumPy scorer to automatically detect missing or disconnected limbs in model-generated outputs.
50%UnverifiedThe training pipeline involves first training an image baseline model, using the image model to warm-start an eight-frame video baseline model, and ultimately generating 5.6-second inference outputs.
50%UnverifiedThe warm-started baseline model reached the loss level of a model trained from scratch at 10,000 steps in only about 4,000 steps.
50%UnverifiedThe author defines the project not as a finished video model, but as a small dataset and learning tool for hands-on building, dissecting, and understanding.
50%UnverifiedThe dataset is licensed under CC0, and the code is licensed under MIT.
50%UnverifiedThe project comes with a free Colab notebook, a small reference baseline model, pre-trained checkpoints, and a NumPy scorer.
50%UnverifiedDeveloper sprited-ai shared the educational dataset project named Dancing Stick Figures on Reddit.
50%