UniSpace

Official model repository for UniSpace: Patch-Reparameterized Vision Encoders for Unified Representation and Scalable Multimodal Modeling.

Model preview. Checkpoints, model configuration, and inference examples are coming soon. This repository is published now as the canonical model release location accompanying the paper.

Links

• Code
• Project page
• Paper: coming soon

Model family

• PR-SigLIP2 — patch-reparameterized SigLIP2 encoder
• PR-DINOv2 — patch-reparameterized DINOv2 encoder
• PR-Qwen-ViT — patch-reparameterized Qwen vision encoder
• UniSpace — unified visual understanding, generation, and editing model

Selected results

Patch-reparameterized encoders

Encoder	PSNR ↑	SSIM ↑	rFID ↓
PR-SigLIP2	29.64	0.87	0.18
PR-DINOv2	30.84	0.90	0.14
PR-Qwen-ViT	30.16	0.88	0.17

UniSpace

Benchmark	Overall ↑
GenEval	0.84
DPG-Bench	86.49
OneIG EN	0.561
OneIG ZH	0.533
ImgEdit	4.28
GEdit EN	7.407
GEdit ZH	7.382

Release status

• Canonical model repository
• Paper metrics and release links
• Model checkpoints
• Configuration and preprocessing files
• Inference examples
• Reproduction logs and artifact hashes

Citation

Citation metadata will be added with the public paper link.

License

The repository shell is released under the MIT License. Checkpoint-specific licenses and third-party notices will be included with each released artifact.