Original ImageNet-256 IBQ tokenizer with raw and EMA weights and Gestalt token mapping.

Contains encoder, index-propagation quantizer, decoder and quantization projections. Default variant is EMA, matching the original evaluation configuration. Gestalt visual IDs equal IBQ IDs + 126356 inside IMAGE delimiters; visual IDs overlap the tail of the text vocabulary. This is the existing IBQ codec reused by Gestalt; the Gestalt diffusion/language backbone is not included.

Source and conversion

Model/weights: TencentARC/IBQ-Tokenizer-16384, fixed revision ef2200997f97118050c5b17dde48e8b78690f866. Native code: TencentARC/SEED-Voken, fixed revision c1544efff959ba18d453089071307a7115698551.

All selected tensor names, shapes, dtypes and payload bytes are preserved, except the explicit IBQ EMA shadow-name mapping. See extraction.json, provenance.json and tensor/conversion mappings. This repository contains component weights and a local native loader, not an AutoModel.from_pretrained checkpoint.

Load

Install requirements.txt in a compatible environment, then:

from huggingface_hub import snapshot_download
import sys
import torch
root = snapshot_download("gatilin/ibq_tokenizer_16384_gestalt")
sys.path.insert(0, root)
from component import load_model
model = load_model(root, dtype=torch.float32, device="cpu")

The model is returned in evaluation mode. Run inference under torch.inference_mode(). Inputs are native tensors/patches, not arbitrary PIL images; inspect component.py and the included validation script for the exact layout. CPU float32 was tested; CUDA/MPS and full end-to-end benchmarks were not tested.

Validation

See validation.json for the executed component tests, shapes and limitations. Source payload equality and native strict loading are required before release. Component tests do not establish full-model generation or understanding quality. Original stochastic decoding settings and normalization conventions must be retained.

License and attribution

Original upstream notices and license files are retained. Reused third-party components keep their original terms. This extraction does not claim authorship of the underlying trained model.

Use variant="raw" for non-EMA weights. to_gestalt_ids/from_gestalt_ids only apply within explicit image spans; they are not global text token mappings. The linked Gestalt checkpoint is marked as continual pretraining, not the final SFT model.