llama.cpp
9cec449e - vlib-video: PixelDiffFilter 64x64 thumbnail MSE pre-VLM filter

Commit
59 days ago
vlib-video: PixelDiffFilter 64x64 thumbnail MSE pre-VLM filter Port of mlx-vlm-continuous/watchdawg/adaptive.py:PixelDiffFilter (lines 111-150). Cheap MSE-on-thumbnails check to skip near-identical frames before they reach the VLM encoder. Nearest-neighbour downsample (vs. MLX's PIL bilinear) keeps the dependency surface zero — the difference is negligible for an MSE-based skip heuristic. MLX default threshold (0.005) is the recommended starting point for screen-capture sources. Tests (test-vlib-pixel-diff-filter): first-frame, identical-frame, ~10% strip perturbation, single-pixel change. All green.
Author
Parents
Loading