Two input spaces
Semantic and reconstructive patch embeddings view the same image through complementary projections.
THE METHOD
Patch reparameterization keeps meaning and pixels in view.
One shared backbone
A frozen pretrained vision transformer preserves its semantic prior while the lightweight reconstructive path learns detail.
One unified model
Qwen3 with Mixture-of-Transformers consumes the visual representation for understanding, editing, and generation.
VISUAL GALLERY
Examples produced by the release checkpoint.
Deterministic samples include their prompt and seed. More generation, reconstruction, and editing examples will be added with the stable release.
Complex text-to-imageDrag or scroll →
Instruction-based image editingInput → UniSpace
All examples above are copied directly from the retained 0012000 benchmark output. No hand-picked external generations are used.
PAPER RESULTS
Selected release checkpoints.
Numbers shown here are the paper results. A full 2,212-image GenEval rerun matches every breakdown; a complete 737-image ImgEdit rerun scores 4.25 overall versus 4.28 in the paper.
Clean-cluster status — GenEval: 0.98 / 0.92 / 0.69 / 0.88 / 0.83 / 0.73, overall 0.84, exact paper match. ImgEdit: 737/737 judge responses, overall 4.25 versus 4.28. Public artifact hashes and the remaining reproduction logs will accompany the stable release.
ENCODER RESULTS
Patch-reparameterized visual encoders
ImageNet-1K reconstruction at 256 × 256
| Encoder | Training data | Downsample | PSNR ↑ | SSIM ↑ | rFID ↓ |
|---|---|---|---|---|---|
| PR-SigLIP2 | ImageNet-1K | 16 | 29.64 | 0.87 | 0.18 |
| PR-DINOv2 | ImageNet-1K | 14 | 30.84 | 0.90 | 0.14 |
| PR-Qwen-ViT | Web data | 16 | 30.16 | 0.88 | 0.17 |
ImageNet class-conditional generation with DiTDH-XL
| Encoder | CFG | gFID ↓ | sFID ↓ | IS ↑ | Precision ↑ | Recall ↑ |
|---|---|---|---|---|---|---|
| PR-SigLIP2 | No | 4.42 | 6.66 | 190.4 | 0.72 | 0.65 |
| PR-SigLIP2 | Yes | 2.80 | 6.05 | 248.2 | 0.78 | 0.61 |
| PR-DINOv2 | No | 2.10 | 5.39 | 217.2 | 0.78 | 0.64 |
| PR-DINOv2 | Yes | 1.87 | 4.89 | 274.3 | 0.82 | 0.60 |
Multimodal understanding
| Encoder | POPE | GQA | TextVQA | MM-Vet | MMBench | MME-S | MME-P | Avg. ↑ |
|---|---|---|---|---|---|---|---|---|
| PR-SigLIP2 | 84.80 | 61.10 | 54.60 | 25.50 | 65.20 | 1741.1 | 1447.2 | 64.37 |
| PR-Qwen-ViT | 86.57 | 63.13 | 63.68 | 30.28 | 69.07 | 1874.5 | 1522.7 | 68.94 |
UNISPACE RESULTS
Unified generation and editing
Text-to-image generation
| Benchmark | Breakdown | Scores | Overall ↑ |
|---|---|---|---|
| GenEval | Single / Two / Count / Color / Position / Attribute | 0.98 / 0.92 / 0.69 / 0.88 / 0.83 / 0.73 | 0.84 |
| OneIG EN | Align / Text / Reason / Style / Diversity | 0.860 / 0.937 / 0.311 / 0.467 / 0.233 | 0.561 |
| OneIG ZH | Align / Text / Reason / Style / Diversity | 0.807 / 0.881 / 0.276 / 0.455 / 0.244 | 0.533 |
| DPG-Bench | Global / Entity / Attribute / Relation / Other | 84.80 / 92.26 / 90.00 / 94.97 / 88.80 | 86.49 |
Image editing
| Benchmark | Breakdown | Scores | Overall ↑ |
|---|---|---|---|
| ImgEdit | Add / Adjust / Extract / Replace / Remove / Background / Style / Hybrid / Action | 4.53 / 4.38 / 3.61 / 4.67 / 4.42 / 4.23 / 4.55 / 2.70 / 4.47 | 4.28 |
| GEdit EN | Semantic consistency / Perceptual quality | 8.287 / 7.055 | 7.407 |
| GEdit ZH | Semantic consistency / Perceptual quality | 8.270 / 6.998 | 7.382 |
| GEdit paper aggregate | Semantic consistency / Perceptual quality | 8.253 / 7.032 | 7.399 |
MODEL FAMILY
Three reproducible release lines.
TOKENIZER + DiT
PR-SigLIP2
Shared SigLIP2 backbone, semantic and reconstruction-aware patch embeddings, unified checkpoint, and ImageNet DiT.
Release status →TOKENIZER + DiT
PR-DINOv2
The complete DINOv2 reconstruction and generation line, released alongside PR-SigLIP2.
Release status →UNDERSTAND + EDIT + GENERATE
UniSpace MoT
Qwen3-8B, Qwen3Unified visual tokenizer, and task-specialized MLP experts with shared attention.
Release status →RELEASE STATUS
Paper now. Code and reproducibility artifacts next.
Project overview, paper metrics, selected qualitative results, and release status.
Training and inference code, public model weights, hashes, benchmark commands, and clean-cluster reproduction logs.
The preview repository deliberately distinguishes original paper measurements from results reproduced using public download paths.