---

name: haar_high_frequency_enhancement description: Extract Haar high-frequency components and compute DDPM-PA high-frequency preservation and detail losses.

Haar High-Frequency Enhancement

Use this skill when implementing the DDPM-PA high-frequency branch. It decomposes images with a Haar transform, forms the paper's high-frequency representation LH + HL + HH, computes high-frequency pairwise preservation, and computes high-frequency MSE against target training images.

Do not substitute ordinary pixel MSE for this module. DDPM-PA separates image-level relative-distance preservation from high-frequency detail preservation.

Inputs

• Source predicted clean images, adapted predicted clean images, and target clean images.
• Images represented as nested lists shaped [batch, channels, height, width] with even height and width.
• A callable or script for pairwise KL loss, or the included fallback implementation.

Outputs

• High-frequency component tensors for each input batch.
• Lhf: pairwise KL preservation loss over high-frequency components.
• Lhfmse: mean squared error between adapted and target high-frequency components.

Workflow

1. Validate even spatial dimensions and matching tensor shapes.
2. For each 2x2 block, compute Haar components equivalent to LL, LH, HL, and HH filters.
3. Return hf = LH + HL + HH for each block/channel/sample.
4. Compute pairwise KL between source and adapted high-frequency tensors.
5. Compute MSE between adapted and target high-frequency tensors.
6. Record component energies for mechanism checks.

Validation

Run:

python scripts/haar_hf.py --smoke
python -m pytest tests

Tests confirm constant images have zero high-frequency response, checkerboards have positive response, and losses are finite.

Limitations

This portable script implements a one-level Haar transform for small recovery tensors. Full training code may use framework convolutions, but should preserve the same component definitions and loss separation.