PaddleOCR-VL-1.6-MLX-8bit / preprocessor_config.json
dikeoulias's picture
preprocessor_config.json: preprocessor/processor config: raise vision max_pixels 1,003,520 -> 4,064,256 (6x6 stage of the 28px/token grid: 5,184 merged-token budget; A4 lands ~5,100; min_pixels unchanged 112,896). Small-text OCR tuning for the zoom-4.0 render lane; lower stages selectable client-side via render scale / crop pixel budget. Deliberate divergence from upstream PaddlePaddle/PaddleOCR-VL-1.6 and mlx-community 1.5/1.6 defaults (1,003,520). Measured on M-series: full-page wall ~5x vs 1.0MP cap, crop lane unaffected, server RSS ~2.2 GB.
b4c87a6 verified
Raw History Blame
641 Bytes
{
"auto_map": {
"AutoImageProcessor": "image_processing_paddleocr_vl.PaddleOCRVLImageProcessor",
"AutoProcessor": "processing_paddleocr_vl.PaddleOCRVLProcessor"
},
"do_convert_rgb": true,
"do_normalize": true,
"do_rescale": true,
"do_resize": true,
"image_mean": [
0.5,
0.5,
0.5
],
"image_processor_type": "PaddleOCRVLImageProcessor",
"image_std": [
0.5,
0.5,
0.5
],
"max_pixels": 4064256,
"merge_size": 2,
"min_pixels": 112896,
"patch_size": 14,
"processor_class": "PaddleOCRVLProcessor",
"resample": 3,
"rescale_factor": 0.00392156862745098,
"temporal_patch_size": 1
}