---

name: guided_score_combination description: Compute classifier-free guided diffusion score estimates from conditional and unconditional predictions.

Guided Score Combination

Use this skill when a sampler needs Ho and Salimans classifier-free guidance: guided = (1+w)*conditional - w*unconditional. It is not a classifier-gradient implementation.

Inputs

• Scalar or equal-length vector conditional epsilon estimate.
• Scalar or equal-length vector unconditional epsilon estimate.
• Guidance strength w, usually non-negative for fidelity-increasing guidance.

Outputs

• Guided epsilon estimate with matching shape.
• Audit metadata preserving w and formula.

Workflow

1. Validate shapes and numeric values.
2. Apply (1+w) * eps_cond - w * eps_uncond.
3. For w=0, confirm the output equals the conditional prediction.
4. Log formula usage for recovery source evidence.

Validation

Run the included tests or the Distiller skill validator.

Limitations

The formula depends on a denoiser that was trained with conditioning dropout; by itself it does not train a model.