JUICE-2B-SigLIP2

Text-to-image generation and question answering (SFT model of the paper).

JUICE is one flow-matching Transformer (MM-RiT d28, 2.1 B parameters) that generates images and text together in the representation spaces of frozen encoders: SigLIP2 patch states for images and Flan-T5-Small token states for text.

• Code: https://github.com/Ced3-han/JUICE-Code
• Paper: https://arxiv.org/abs/2610.28801
• All JUICE models: https://huggingface.co/collections/CedLJH/juice
• Weights: fast EMA weights (decay 0.9996) at step 15,000 of the SFT.
• Image space: siglip2; inference defaults of the paper in config.json.

Usage

Install the code from https://github.com/Ced3-han/JUICE-Code, then:

from juice.pipeline import JuicePipeline

pipe = JuicePipeline.from_pretrained("CedLJH/JUICE-2B-SigLIP2", device="cuda")
images = pipe.text_to_image("a red bicycle next to a blue door", num_images=4, seed=0)
answers = pipe.answer(["photo.jpg"], ["What color is the door?"])

Results of the paper

GenEval	DPG-Bench	VQAv2 test-dev	VQAv2 (1k)	POPE	GQA	MMBench
0.866	85.27	73.10	71.2	84.0	58.6	66.0

DPG-Bench: official scoring rule (1,064 prompts). VQAv2 test-dev: score of the EvalAI server (test-dev phase).

Text-to-image: 25 steps, guidance 5 (the defaults of config.json).

The repository guides (docs/EVAL_T2I.md, docs/EVAL_VQA.md) give the full tables and the commands that reproduce them.

License

The weights are under CC BY-NC 4.0. The code downloads the frozen encoders and decoders separately; they keep their own licenses (see THIRD_PARTY_LICENSES.md in the code repository).

Citation

@article{li2026juice,
  title   = {{JUICE}: Unified Multimodal Modeling in Pretrained Representation Spaces},
  author  = {Li, Jiahan and Su, Huangyuan and Tong, Shengbang and Yang, Shusheng and
             Guo, Ruihan and Huang, Suozhi and Zhao, Hexu and Sharma, Pratyusha and
             Yang, Sherry and Gu, Jiatao and Fergus, Rob and Cho, Kyunghyun},
  journal = {arXiv preprint arXiv:2610.28801},
  year    = {2026}
}