drift-generator/grn-safetensor

GRN 2B text-to-image — safetensors conversion

An unofficial conversion of the GRN (Generative Refinement Networks) 2B text-to-image checkpoint and its HBQ tokenizer from PyTorch pickles to safetensors. The weights are ByteDance's; nothing was retrained or modified.

• Official weights: bytedance-research/GRN
• Official code: MGenAI/GRN
• Paper: arXiv 2604.13030

For anything about the model itself — usage, training, results, citation — see the official pages above.

Files

File	Converted from	Notes
grn-t2i-2b-bf16.safetensors	GRN_T2I_2B_FSA_251600.pth	the transformer, cast to BF16
grn-hbq-tokenizer.safetensors	HBQ_image_video_tokenizer_64dim_M4_20260626.ckpt	the tokenizer (EMA weights), as stored
grn_convert.py		the script used

Tensor names are unchanged. The text encoder (umT5-XXL) is not included; use the one from the official repository or any standard umT5-XXL encoder.

Conversion

Done on the CPU with torch.load(..., weights_only=True), keeping tensors only:

python grn_convert.py GRN_T2I_2B_FSA_251600.pth grn-t2i-2b-bf16.safetensors bf16
python grn_convert.py HBQ_image_video_tokenizer_64dim_M4_20260626.ckpt grn-hbq-tokenizer.safetensors

License

MIT, as the original release.