Score SDE Reverse Sampling

Use this skill when you need an executable reduced sampler that follows the Score SDE reverse-time mechanism. It consumes SDE drift/diffusion callbacks and a score function; it does not depend on the original paper repository.

Do not use this skill to claim full image generation quality. It is a compact sampler for validating reverse SDE and predictor-corrector mechanics.

Inputs

• Initial state from a prior or supplied latent value.
• Descending times from T to eps.
• Drift and diffusion callbacks from a forward SDE.
• Score callback.
• Predictor mode: Euler-Maruyama or probability flow.
• Corrector settings: number of Langevin steps, step size, optional noise seed.

Outputs

• Final state and trajectory.
• Score evaluation count.
• Predictor/corrector update diagnostics.
• Flags indicating stochastic or deterministic operation.

Workflow

1. Build a SamplerConfig.
2. Call run_sampler with callbacks for drift, diffusion, and score.
3. Use probability_flow=True for deterministic ODE-style updates.
4. Use corrector_steps > 0 to add Langevin correction before each predictor step.
5. Save the trajectory and score-call count as recovery evidence.

Validation

Run:

python scripts/reverse_sampling.py --self-test
python tests/test_reverse_sampling.py

The tests check deterministic probability-flow behavior, finite stochastic updates, and corrector movement under a controlled score.

Limitations

The sampler is intentionally low-dimensional and standard-library only. Full PC sampling for images requires neural model replicas, device mapping, denoising options, and dataset scalers outside this skill.