---

name: score_sde_family description: Define VE, VP, and sub-VP SDE contracts plus reverse-time SDE and probability-flow ODE dynamics for score-based generative modeling.

Score SDE Family

Use this skill when an experiment needs the paper's forward SDE perturbation, prior density, reverse-time SDE, or probability-flow ODE semantics. Do not use it to claim full image-generation quality by itself.

Inputs

• SDE type: ve, vp, or subvp.
• Hyperparameters: sigma_min, sigma_max, beta_min, beta_max, num_steps.
• State values as scalars or lists and continuous time t in (0, 1].
• Optional score function value for reverse dynamics.

Outputs

• Forward drift and diffusion.
• Marginal mean and standard deviation.
• Prior log-density.
• Reverse drift and diffusion, including probability-flow zero diffusion.

Workflow

1. Instantiate an SDE with scripts/sde_family.py.
2. Use marginal_prob to perturb clean data for denoising score matching.
3. Use reverse_drift_diffusion with probability_flow=False for stochastic reverse SDE checks.
4. Use probability_flow=True for deterministic ODE checks and likelihood-style recovery.
5. Record SDE type, time, drift, diffusion, and marginal statistics in recovery logs.

Validation

Run:

python tests/test_sde_family.py
python scripts/sde_family.py --sde vp --x 1.0 --t 0.5 --score -0.25

Limitations

This skill implements small deterministic numeric contracts for recovery and testing. Full neural image models, JAX/Flax code, and large checkpoint loading are outside this skill.