gatilin/nld-uni-image-tokenizer

NLD-Uni discrete image tokenizer

Community extraction of the encoder.vqvae.* tensors from NVIDIA NLD-Uni at revision 5db755a2eb0ab84654e3dcf80c7d0245bfe36a50. It contains the image encoder, vector-quantization codebook and image decoder. The separate packed image bridge, RADIO understanding encoder and diffusion language backbone are not included; this is not a standalone text-to-image model.

Tensor names, shapes, dtype and values are retained; see tensor_mapping.json and extraction.json. The bundled vq.py and vq_model.py are from the original checkpoint repository. The CPU smoke test is documented in validation.json.

Loading

Install PyTorch, numpy, accelerate, Pillow and safetensors. The bundled component.load_model(root) loads the weights. Input is RGB, normalized to [-1, 1]; model.encode(x) yields quantized features and token IDs, and model.decode(quantized_features) returns reconstructed pixels. Preserve spatial token arrangement for downstream image generation.

License

Distributed under the bundled NVIDIA License. Section 3.3 limits use to non-commercial research or evaluation (subject to the stated NVIDIA exception). Retain all source notices.