---

name: diffusion_sampler_proxy description: Run a tiny reverse-diffusion-like sampler that exercises classifier-free guided score calls at each step.

Diffusion Sampler Proxy

Use this skill for bounded recovery of classifier-free guidance when full image diffusion models are unavailable. It consumes conditional and unconditional predictions and calls the guided score formula every step.

Inputs

• Initial scalar noise values.
• Condition/class target prototype.
• Guidance strength list and step count.
• Callable guided-score implementation.

Outputs

• Generated scalar proxy samples.
• Per-step trace containing conditional prediction, unconditional prediction, and guided prediction.

Workflow

1. Initialize samples from deterministic noise.
2. At every step estimate conditional and unconditional residuals toward class and global means.
3. Combine estimates with classifier-free guidance.
4. Update the sample and retain trace records.

Validation

Tests verify that stronger guidance moves samples closer to the class prototype and that trace records include both denoiser evaluations.

Limitations

This is a reduced mechanism check, not a photographic diffusion sampler.