sample.md

Sampling

All sampling commands must be run from within the submodule:

cd external/motion-diffusion-model

Model architecture arguments (--arch, --latent_dim, etc.) are loaded automatically from the checkpoint's args.json and should not be overridden manually.

1. Motion Synthesis

Generate motion sequences from a trained model conditioned on action IDs or text prompts.

python3 -m sample.generate \
  --model_path ./save/ntu60_kinemic/model005000.pt \
  --action_id 99 102 104 \
  --num_repetitions 5

Omitting --action_id automatically samples from the full set of action classes the model was trained on.

Argument reference

Argument	Default	Description
--model_path	required	Path to the .pt checkpoint file
--output_dir	auto	Output directory. If empty, created next to the checkpoint
--num_samples	9	Number of prompts/actions to sample
--num_repetitions	5	Repetitions per action/prompt
--guidance_param	2.5	Classifier-free guidance scale
--motion_length	2.5	Motion length in seconds
--sampling_stream	target	prior | target
--sampling_mode	single	single (one stream) | cycle (both streams alternating)
--action_id	[]	One or more action class IDs
--text_prompt	``	A single free-form text prompt
--input_text	``	Path to a .txt file with one text prompt per line
--action_file	``	Path to a .txt file with one action name per line
--unconstrained_sampling	flag	Sample without any conditioning

2. Build a Synthetic Dataset

Generate a full synthetic dataset mirroring the class and length distribution of the training data. Samples used during training are excluded; the rest is generated synthetically.

python3 -m sample.synth_dataset_generate \
  --model_path ./save/ntu60_kinemic/model005000.pt \
  --oversample 2.0

Argument reference

Argument	Default	Description
--model_path	required	Path to the .pt checkpoint file
--oversample	2.0	Scale factor applied to the number of real training samples
--guidance_param	2.5	Classifier-free guidance scale
--sampling_stream	target	Must be target for synthetic dataset generation

3. Inspect a Synthetic Dataset

Visualize the generated synthetic dataset as a grid of animated skeletons.

python3 -m sample.synth_dataset_inspect \
  --model_path ./save/ntu60_kinemic/model005000.pt

Argument reference

Argument	Default	Description
--model_path	required	Path to the .pt checkpoint used to generate the dataset
--num_rows	12	Rows in the visualization grid
--num_cols	5	Columns in the visualization grid