Work

Layerwise Weight Statistics and Pell-Inspired Quantization for SDXL: An Exploratory Technical Note on Calibration-Free Weight Quantization

Year: 2026
Type: article
Abstract: This technical note explores a weight-statistics-based approach to neural network quantization, centered on a layerwise amplitude-to-dispersion statistic defined as the ratio of the maximum absolute weight to its standard deviation. Preliminary measurements across ResNet-18, BERT-base, and Stable Diffusion XL (SDXL) suggest one robust observation: this statistic exhibits substantially lower empirical dispersion in trained feed-forward and convolutional families than in cross-attention families. This differential behavior may inform family-specific quantization strategies. Building on this, we investigate a weight-only quantization heuristic incorporating an exploratory arithmetic anchor derived from the Pell norm form sequence in ℤ[√2] (D5 = 839). We evaluate condition-number-guided routing strategies to preserve highly anisotropic attention projections. A preliminary SDXL evaluation (100 PartiPrompts) reports semantic alignment, structural similarity, and memory consumption. We emphasize that the amplitude statistic is not derived from first principles, and the Pell-derived anchor is presented openly as an exploratory implementation rather than a theoretically established optimum. We welcome alternative arithmetic anchors and independent investigation of these hypotheses. Research status: Exploratory technical note. Not peer-reviewed.
Language: English
Open Access status: green
OpenAlex ID: W7233749807
FWCI:
Cites:
Cited by:
Related to: 10