unitok_large_visual_codec / preprocessor_config.json
gatilin's picture
Publish pinned visual components with exact source payloads and native numerical validation
7bfa997 verified
Raw History Blame Contribute Delete
524 Bytes
{
"reference": "https://github.com/FoundationVision/UniTok/blob/bb8012e8e8d7be66928bedd80506c56a55ea6305/utils/data.py",
"reference_function": "build_transforms: preprocess_val",
"resize_shortest_edge": 256,
"resize_interpolation": "PIL Lanczos",
"center_crop": [256, 256],
"convert_to_rgb": true,
"to_tensor_scale": "uint8 / 255",
"normalization": "tensor * 2 - 1",
"layout": "NCHW",
"notes": "This describes the native deterministic validation transform, not every downstream MLLM prompt processor."
}