File size: 33,950 Bytes
67a8d88 ad2f241 67a8d88 ad2f241 67a8d88 ad2f241 bbfd692 8436989 bbfd692 8436989 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 bbfd692 ad2f241 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 |
---
base_model:
- ByteDance-Seed/BAGEL-7B-MoT
datasets:
- Ryann829/Scone-S2I-57K
language:
- en
license: apache-2.0
pipeline_tag: text-to-image
tags:
- subject-driven
library_name: transformers
---
<p align="center">
<img src="assets/logo.png" alt="Scone" width="400"/>
</p>
<h3 align="center">
Scone: Bridging Composition and Distinction in Subject-Driven Image Generation </br>via Unified Understanding-Generation Modeling
</h3>
<p align="center">
<a href="https://arxiv.org/abs/2512.12675"><img alt="Build" src="https://img.shields.io/badge/arXiv-Paper-b31b1b.svg"></a>
<a href="https://huggingface.co/papers/2512.12675"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Paper&color=blue"></a>
<a href="https://huggingface.co/Ryann829/Scone"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Model&color=green"></a>
<a href="https://huggingface.co/datasets/Ryann829/Scone-S2I-57K"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Data&color=yellow"></a>
<a href="https://huggingface.co/datasets/Ryann829/SconeEval"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Benchmark&color=yellow"></a>
<a href="https://github.com/Ryann-Ran/Scone"><img alt="Build" src="https://img.shields.io/github/stars/Ryann-Ran/Scone"></a>
</p>
><p align="center">
> <span style="color:#137cf3; font-family: Gill Sans;">
> Yuran Wang<sup>1,2</sup><strong>*</strong> Bohan Zeng<sup>1,2</sup><strong>*</strong> Chengzhuo Tong<sup>1,2</sup> Wenxuan Liu<sup>1</sup> Yang Shi<sup>1,2</sup><br>Xiaochen Ma<sup>1</sup> Hao Liang<sup>1</sup> Yuanxing Zhang<sup>2</sup> Wentao Zhang<sup>1</sup><strong>†</strong>
> </span>
> <br>
> <span><sup>1</sup>Peking University <sup>2</sup>Kling Team, Kuaishou Technology</span>
> <br>
> <span><strong>*</strong> Equal contribution, <strong>†</strong> Corresponding author</span>
></p>
# 📢 News
- 2025.12.16: The [paper](https://arxiv.org/abs/2512.12675), [training code](https://github.com/Ryann-Ran/Scone?tab=readme-ov-file#-train), [inference and evaluation code](https://github.com/Ryann-Ran/Scone?tab=readme-ov-file#-inference-and-evaluation), [model weight](https://huggingface.co/Ryann829/Scone), [training data](https://huggingface.co/datasets/Ryann829/Scone-S2I-57K), [SconeEval benchmark](https://huggingface.co/datasets/Ryann829/SconeEval) are now released.
# 📖 Introduction
Subject-driven image generation has recently gained significant attention, with the focus evolving from single-subject to multi-subject generation, incorporating more input images. Existing methods can process two or more input images and combine subjects based on instructions, showcasing potential for more complex composition tasks.
However, existing works primarily focus on expanding subject combinations while neglecting the ability to distinguish target subjects in complex contexts. As shown in Figure 1.(a), although current models can combine multiple subjects, they may fail to identify and generate the correct target subject when a reference image contains multiple candidates, leading to problems such as subject omissions (none of the candidate subjects appear) or errors (misidentification of the target subject).
Real-world images often involve interference and intricate details, which further limit practical performance.
Thus, we emphasize examining the input subjects themselves, focusing on the model’s ability to ***distinguish the target subject within complex contexts and leverage this information for generation***.
<figure style="text-align: center; border: none; margin: auto;">
<img src="assets/problem.png" width="512" alt="The distinction problem and challenges."/>
<figcaption><b>Figure 1. The distinction problem and challenges.</b></figcaption>
</figure>
* We propose the **Scone** (**S**ubject-driven **co**mposition and distinctio**n** **e**nhancement) model, which supports multi-subject composition and excels in subject distinction in complex contexts. Experiments show our Scone ranks first among open-source models on OmniContext benchmark.
* We introduce the **understanding bridge strategy**, which transforms the understanding expert into a semantic bridge, enabling early multimodal alignment and attention-based semantic filtering to guide the generation expert, enhancing subject distinction and semantic fidelity without adding extra parameters.
* We develop **SconeEval**, a challenging benchmark with three difficulty levels, to evaluate performance on subject-driven image generation tasks from both composition and distinction perspectives.
# 🔧 Environment setup
```bash
git clone https://github.com/Ryann-Ran/Scone.git
cd Scone
conda create -n scone python=3.10 -y
conda activate scone
pip install -r requirements.txt
pip install flash_attn==2.5.8 --no-build-isolation
```
# 🔥 Train
## Data and base model preparation
1. Download our **22K refined single-candidate data** and **35K multi-candidate data** from [Scone-S2I-57K](https://huggingface.co/datasets/Ryann829/Scone-S2I-57K). The 70K base single-canididate data are sampled from open-source datasets like [X2I](https://huggingface.co/datasets/yzwang/X2I-subject-driven), [MUSAR-Gen](https://huggingface.co/datasets/guozinan/MUSAR-Gen), [UNO-1M](https://huggingface.co/datasets/bytedance-research/UNO-1M), and [Echo-4o-Image](https://huggingface.co/datasets/Yejy53/Echo-4o-Image). Please refer to the dataset links for more details.
```bash
cd Scone
# pip install -U huggingface_hub
hf download Ryann829/Scone-S2I-57K --repo-type=dataset --local-dir ./datasets/Scone-S2I-57K
```
2. Organize the data hierarchy as follows:
```
Scone-S2I-57K
├── parquet_data
│ ├── scone_single_candidate_base/
│ ├── scone_single_candidate_refined/
│ └── scone_multi_candidate/
└── parquet_info
├── scone_single_candidate_base.json
├── scone_single_candidate_refined.json
└── scone_multi_candidate.json
```
3. Replace each `your_data_path` placeholder with your **actual absolute path** in:
* Parquet information files: `./datasets/Scone-S2I-57K/parquet_info/*.json`
* Dataset information file: `./data/dataset_info.py`
4. Download the checkpoint of our base model [BAGEL](https://github.com/ByteDance-Seed/Bagel) from [HuggingFace](https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT):
```bash
cd Scone
# pip install -U huggingface_hub
hf download ByteDance-Seed/BAGEL-7B-MoT --local-dir ./ckpts/BAGEL-7B-MoT
```
> - **Note**: To avoid out-of-memory (OOM) issues, we disable the EMA update strategy originally used in BAGEL. All our training processes are conducted on 8 Nvidia A800 GPUs.
> - The usage of semantic mask in the understanding bridge strategy is controlled by the training argument `--use_semantic_mask`.
## Stage I: Composition training
For Step 1, please use base single-candidate data for 1 epoch (~30 hours):
```bash
bash scripts/train_stage1_step1.sh # 🔥 Und., Gen.
```
For Step 2, please use refined single-candidate data for 1 epoch (~15 hours) and replace `model_path` in the script with your Step 1 checkpoint :
```bash
bash scripts/train_stage1_step2.sh # 🔥 Und., Gen.
```
## Stage II: Distinction training with understanding bridge strategy
For Step 1, please use refined sinlgle-candidate data and multi-candidate data for 1k steps (~5 hours) and replace `model_path` in the script with your Stage 1 Step 2 checkpoint:
```bash
bash scripts/train_stage2_step1.sh # 🔥 Und. ❄️ Gen.
```
For Step 2, please use refined sinlgle-candidate data and multi-candidate data for 1k steps (~5 hours) and replace `model_path` in the script with your Stage 2 Step 1 checkpoint:
```bash
bash scripts/train_stage2_step2.sh # 🔥 Und., Gen.
```
# 🔍 Inference and Evaluation
## Scone model preparation
Download the Scone model checkpoint from [HuggingFace](https://huggingface.co/Ryann829/Scone):
```bash
# pip install -U huggingface_hub
hf download Ryann829/Scone --local-dir ./ckpts/Scone
```
## Single case inference
Run the inference script:
```bash
bash scripts/inference_single_case.sh
```
**Example Output:** (Images sampled at 1024x1024 resolution with seed 1234, except for GPT-4o and Gemini-2.5-Flash-Image APIs)
<table>
<tr>
<th style="width: 150px; text-align: center;">Ref. 1</th>
<th style="width: 150px; text-align: center;">Ref. 2</th>
<th style="width: 200px; text-align: center;">Instruction</th>
<th style="width: 150px; text-align: center;">Scone (Ours)</th>
<th style="width: 150px; text-align: center;">GPT-4o</th>
<th style="width: 150px; text-align: center;">Gemini-2.5-Flash-Image</th>
<th style="width: 150px; text-align: center;">UNO</th>
<th style="width: 150px; text-align: center;">Qwen-Image-Edit-2509</th>
<th style="width: 150px; text-align: center;">BAGEL</th>
<th style="width: 150px; text-align: center;">OmniGen2</th>
<th style="width: 150px; text-align: center;">Echo-4o</th>
</tr>
<tr>
<td style="width: 150px; text-align: center;"><img src="./example_images/image_object.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_images/image_character.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 200px; text-align: center;">The man from image 2 holds the object which has a blue-and-red top in image 1 in a coffee shop.</td>
<td style="width: 150px; text-align: center;"><img src="./example_results/Scone.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/GPT-4o.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/Gemini-2.5-Flash-Image.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/UNO.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/Qwen-Image-Edit-2509.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/BAGEL.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/OmniGen2.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
<td style="width: 150px; text-align: center;"><img src="./example_results/Echo-4o.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
</tr>
</table>
# 📊 Performance
## OmniContext benchmark
<table border="1" style="border-collapse: collapse; width: 100%;">
<thead>
<tr>
<th rowspan="2">Method</th>
<th colspan="2" style="text-align: center;">SINGLE ↑</th>
<th colspan="3" style="text-align: center;">MULTIPLE ↑</th>
<th colspan="3" style="text-align: center;">SCENE ↑</th>
<th rowspan="2" style="text-align: center;">Average ↑</th>
</tr>
<tr>
<th style="text-align: center;">Character</th>
<th style="text-align: center;">Object</th>
<th style="text-align: center;">Character</th>
<th style="text-align: center;">Object</th>
<th style="text-align: center;">Char. + Obj.</th>
<th style="text-align: center;">Character</th>
<th style="text-align: center;">Object</th>
<th style="text-align: center;">Char. + Obj.</th>
</tr>
</thead>
<tbody>
<tr>
<td colspan="10" style="background-color: #ffefe6; text-align: center; font-weight: bold; font-style: italic;">Closed-Source Model</td>
</tr>
<tr>
<td>Gemini-2.5-Flash-Image</td>
<td>8.79</td>
<td><strong>9.12</strong></td>
<td>8.27</td>
<td>8.60</td>
<td>7.71</td>
<td>7.63</td>
<td>7.65</td>
<td>6.81</td>
<td>8.07</td>
</tr>
<tr>
<td>GPT-4o<sup>*</sup></td>
<td><strong>8.96</strong></td>
<td>8.91</td>
<td><strong>8.90</strong></td>
<td><strong>8.95</strong></td>
<td><strong>8.81</strong></td>
<td><strong>8.92</strong></td>
<td><strong>8.40</strong></td>
<td><strong>8.44</strong></td>
<td><strong>8.78</strong></td>
</tr>
<tr>
<td colspan="10" style="background-color: #e0eef9; text-align: center; font-weight: bold; font-style: italic;">Generation Model</td>
</tr>
<tr>
<td>FLUX.1 Kontext [dev]</td>
<td>8.07</td>
<td>7.97</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>UNO</td>
<td>7.15</td>
<td>6.72</td>
<td>3.56</td>
<td>6.46</td>
<td>4.90</td>
<td>2.72</td>
<td>4.89</td>
<td>4.76</td>
<td>5.14</td>
</tr>
<tr>
<td>USO</td>
<td>8.03</td>
<td>7.55</td>
<td>3.32</td>
<td>6.10</td>
<td>4.56</td>
<td>2.77</td>
<td>5.38</td>
<td>5.09</td>
<td>5.35</td>
</tr>
<tr>
<td>UniWorld-V2</td>
<td>8.45</td>
<td><strong>8.44</strong></td>
<td>7.87</td>
<td>8.22</td>
<td><strong>7.95</strong></td>
<td><strong>5.36</strong></td>
<td>7.47</td>
<td><strong>6.98</strong></td>
<td>7.59</td>
</tr>
<tr>
<td>Qwen-Image-Edit-2509</td>
<td><strong>8.56</strong></td>
<td>8.41</td>
<td><strong>7.92</strong></td>
<td><strong>8.37</strong></td>
<td>7.79</td>
<td>5.23</td>
<td><strong>7.70</strong></td>
<td>6.86</td>
<td><strong>7.60</strong></td>
</tr>
<tr>
<td colspan="10" style="background-color: #E6E6FA; text-align: center; font-weight: bold; font-style: italic;">Unified Model</td>
</tr>
<tr>
<td>BAGEL</td>
<td>7.00</td>
<td>7.04</td>
<td>5.32</td>
<td>6.69</td>
<td>6.74</td>
<td>3.94</td>
<td>5.77</td>
<td>5.73</td>
<td>6.03</td>
</tr>
<tr>
<td>OmniGen2</td>
<td>8.17</td>
<td>7.63</td>
<td>7.26</td>
<td>7.03</td>
<td>7.56</td>
<td>7.02</td>
<td>6.90</td>
<td>6.64</td>
<td>7.28</td>
</tr>
<tr>
<td>Echo-4o</td>
<td><strong>8.34</strong></td>
<td>8.27</td>
<td>8.13</td>
<td><strong>8.14</strong></td>
<td>8.11</td>
<td><strong>7.07</strong></td>
<td>7.73</td>
<td><strong>7.77</strong></td>
<td>7.95</td>
</tr>
<tr>
<td><strong>Score (Ours)</strong></td>
<td><strong>8.34</strong></td>
<td><strong>8.52</strong></td>
<td><strong>8.24</strong></td>
<td><strong>8.14</strong></td>
<td><strong>8.30</strong></td>
<td>7.06</td>
<td><strong>7.88</strong></td>
<td>7.63</td>
<td><strong>8.01</strong></td>
</tr>
</tbody>
</table>
> - *: GPT-4o responded to 365~370 test cases out of the total 409 cases due to OpenAI safety restrictions.
> - To mitigate randomness, we perform 3 rounds of sampling at 1024x1024 resolution, scoring 3 times per round, yielding 9 group results. The final score is the average of these results.
## SconeEval benchmark
<table border="1" style="border-collapse: collapse; width: 100%;">
<thead>
<tr>
<th rowspan="3">Method</th>
<th colspan="2">Composition ↑</th>
<th colspan="4">Distinction ↑</th>
<th colspan="4">Distinction & Composition ↑</th>
<th colspan="3">Average ↑</th>
</tr>
<tr>
<th>Single</th>
<th>Multi</th>
<th colspan="2">Cross</th>
<th colspan="2">Intra</th>
<th colspan="2">Cross</th>
<th colspan="2">Intra</th>
<th rowspan="2">COM</th>
<th rowspan="2">DIS</th>
<th rowspan="2">Overall</th>
</tr>
<tr>
<th>COM</th>
<th>COM</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
</tr>
</thead>
<tbody>
<tr>
<td colspan="14" style="background-color: #ffefe6; text-align: center; font-weight: bold; font-style: italic;">Closed-Source Model</td>
</tr>
<tr>
<td>Gemini-2.5-Flash-Image</td>
<td>8.87</td>
<td>7.94</td>
<td>9.12</td>
<td><strong>9.15</strong></td>
<td>9.00</td>
<td>8.50</td>
<td>8.27</td>
<td><strong>8.87</strong></td>
<td>8.17</td>
<td>8.85</td>
<td>8.56</td>
<td>8.84</td>
<td>8.70</td>
</tr>
<tr>
<td>GPT-4o<sup>*</sup></td>
<td><strong>8.92</strong></td>
<td><strong>8.51</strong></td>
<td><strong>9.18</strong></td>
<td>8.55</td>
<td><strong>9.45</strong></td>
<td><strong>9.01</strong></td>
<td><strong>8.83</strong></td>
<td>8.49</td>
<td><strong>8.99</strong></td>
<td><strong>9.56</strong></td>
<td><strong>8.98</strong></td>
<td><strong>8.90</strong></td>
<td><strong>8.94</strong></td>
</tr>
<tr>
<td colspan="14" style="background-color: #e0eef9; text-align: center; font-weight: bold; font-style: italic;">Generation Model</td>
</tr>
<tr>
<td>FLUX.1 Kontext [dev]</td>
<td>7.92</td>
<td>-</td>
<td>7.93</td>
<td>8.45</td>
<td>6.20</td>
<td>6.11</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>USO</td>
<td>8.03</td>
<td>5.19</td>
<td>7.96</td>
<td>8.50</td>
<td>7.14</td>
<td>6.51</td>
<td>5.10</td>
<td>6.25</td>
<td>5.07</td>
<td>5.57</td>
<td>6.41</td>
<td>6.71</td>
<td>6.56</td>
</tr>
<tr>
<td>UNO</td>
<td>7.53</td>
<td>5.38</td>
<td>7.27</td>
<td>7.90</td>
<td>6.76</td>
<td>6.53</td>
<td>5.27</td>
<td>7.02</td>
<td>5.61</td>
<td>6.27</td>
<td>6.31</td>
<td>6.93</td>
<td>6.62</td>
</tr>
<tr>
<td>UniWorld-V2<br>(Edit-R1-Qwen-Image-Edit-2509)</td>
<td>8.41</td>
<td><strong>7.16</strong></td>
<td>8.63</td>
<td>8.24</td>
<td><strong>7.44</strong></td>
<td>6.77</td>
<td>7.52</td>
<td>8.03</td>
<td><strong>7.70</strong></td>
<td><strong>7.24</strong></td>
<td><strong>7.81</strong></td>
<td>7.57</td>
<td>7.69</td>
</tr>
<tr>
<td>Qwen-Image-Edit-2509</td>
<td><strong>8.54</strong></td>
<td>6.85</td>
<td><strong>8.85</strong></td>
<td><strong>8.57</strong></td>
<td>7.32</td>
<td><strong>6.86</strong></td>
<td><strong>7.53</strong></td>
<td><strong>8.13</strong></td>
<td>7.49</td>
<td>7.02</td>
<td>7.76</td>
<td><strong>7.65</strong></td>
<td><strong>7.70</strong></td>
</tr>
<tr>
<td colspan="14" style="background-color: #E6E6FA; text-align: center; font-weight: bold; font-style: italic;">Unified Model</td>
</tr>
<tr>
<td>BAGEL</td>
<td>7.14</td>
<td>5.55</td>
<td>7.49</td>
<td>7.95</td>
<td>6.93</td>
<td>6.21</td>
<td>6.44</td>
<td>7.38</td>
<td>6.87</td>
<td>7.27</td>
<td>6.74</td>
<td>7.20</td>
<td>6.97</td>
</tr>
<tr>
<td>OmniGen2</td>
<td>8.00</td>
<td>6.59</td>
<td>8.31</td>
<td>8.99</td>
<td>6.99</td>
<td>6.80</td>
<td>7.28</td>
<td>8.30</td>
<td>7.14</td>
<td>7.13</td>
<td>7.39</td>
<td>7.81</td>
<td>7.60</td>
</tr>
<tr>
<td>Echo-4o</td>
<td><strong>8.58</strong></td>
<td><strong>7.73</strong></td>
<td>8.36</td>
<td>8.33</td>
<td>7.74</td>
<td>7.18</td>
<td>7.87</td>
<td>8.72</td>
<td>8.01</td>
<td>8.33</td>
<td>8.05</td>
<td>8.14</td>
<td>8.09</td>
</tr>
<tr>
<td><strong>Scone (Ours)</strong></td>
<td>8.52</td>
<td>7.40</td>
<td><strong>8.98</strong></td>
<td><strong>9.73</strong></td>
<td><strong>7.97</strong></td>
<td><strong>7.74</strong></td>
<td><strong>8.20</strong></td>
<td><strong>9.25</strong></td>
<td><strong>8.21</strong></td>
<td><strong>8.44</strong></td>
<td><strong>8.21</strong></td>
<td><strong>8.79</strong></td>
<td><strong>8.50</strong></td>
</tr>
</tbody>
</table>
> - *: GPT-4o responded to 365~370 test cases out of the total 409 cases due to OpenAI safety restrictions.
> - To mitigate randomness, we perform 3 rounds of sampling at 1024x1024 resolution, scoring 3 times per round, yielding 9 group results. The final score is the average of these results.
## SconeEval benchmark
<p align="center">
<img src="assets/logo_sconeeval.png" alt="SconeEval Benchmark" width="400"/>
</p>
To evaluate a model’s ability to distinguish and generate the referred subject in complex visual contexts, we introduce a new benchmark, **SconeEval**. It contains 409 test cases across character, object, and scene combinations and subject distinction, with 19 case types in Figure 2(a) and 6 subtasks in Figure 2(b), providing a comprehensive evaluation of a model’s ability to distinguish and utilize subject features.
Unlike traditional benchmarks that emphasize visual fidelity or text alignment, SconeEval focuses on cross-modal reasoning from complex contexts involving reference images and instructions, which requires deciding *whom* to generate when multiple candidates appear within or across images.
SconeEval includes three progressively challenging tasks, as shown in Figure 2(c): composition, distinction, and distinction & composition. In the composition task, each reference image contains a subject, and one or more images correspond to single or multiple generated subjects. In the distinction task, each reference image contains multiple subjects, and the model generates one target subject. The distinction & composition task integrates both settings, where each reference image contains multiple subjects and multiple images are used for multi-subject generation. Tasks involving distinction include cross-category and intra-category cases, indicating whether candidate subjects in a reference image belong to the same category.
<figure style="text-align: center; border: none; margin: auto;">
<img src="assets/sconeeval.png" width="1024" alt="Overview of our SconeEval benchmark."/>
<figcaption><b>Figure 2. Overview of our SconeEval benchmark.</b></figcaption>
</figure>
### 📊 LeaderBoard
<table border="1" style="border-collapse: collapse; width: 100%;">
<thead>
<tr>
<th rowspan="3">Method</th>
<th colspan="2">Composition ↑</th>
<th colspan="4">Distinction ↑</th>
<th colspan="4">Distinction & Composition ↑</th>
<th colspan="3">Average ↑</th>
</tr>
<tr>
<th>Single</th>
<th>Multi</th>
<th colspan="2">Cross</th>
<th colspan="2">Intra</th>
<th colspan="2">Cross</th>
<th colspan="2">Intra</th>
<th rowspan="2">COM</th>
<th rowspan="2">DIS</th>
<th rowspan="2">Overall</th>
</tr>
<tr>
<th>COM</th>
<th>COM</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
<th>COM</th>
<th>DIS</th>
</tr>
</thead>
<tbody>
<tr>
<td colspan="14" style="background-color: #ffefe6; text-align: center; font-weight: bold; font-style: italic;">Closed-Source Model</td>
</tr>
<tr>
<td>Gemini-2.5-Flash-Image</td>
<td>8.87</td>
<td>7.94</td>
<td>9.12</td>
<td><strong>9.15</strong></td>
<td>9.00</td>
<td>8.50</td>
<td>8.27</td>
<td><strong>8.87</strong></td>
<td>8.17</td>
<td>8.85</td>
<td>8.56</td>
<td>8.84</td>
<td>8.70</td>
</tr>
<tr>
<td>GPT-4o<sup>*</sup></td>
<td><strong>8.92</strong></td>
<td><strong>8.51</strong></td>
<td><strong>9.18</strong></td>
<td>8.55</td>
<td><strong>9.45</strong></td>
<td><strong>9.01</strong></td>
<td><strong>8.83</strong></td>
<td>8.49</td>
<td><strong>8.99</strong></td>
<td><strong>9.56</strong></td>
<td><strong>8.98</strong></td>
<td><strong>8.90</strong></td>
<td><strong>8.94</strong></td>
</tr>
<tr>
<td colspan="14" style="background-color: #e0eef9; text-align: center; font-weight: bold; font-style: italic;">Generation Model</td>
</tr>
<tr>
<td>FLUX.1 Kontext [dev]</td>
<td>7.92</td>
<td>-</td>
<td>7.93</td>
<td>8.45</td>
<td>6.20</td>
<td>6.11</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>USO</td>
<td>8.03</td>
<td>5.19</td>
<td>7.96</td>
<td>8.50</td>
<td>7.14</td>
<td>6.51</td>
<td>5.10</td>
<td>6.25</td>
<td>5.07</td>
<td>5.57</td>
<td>6.41</td>
<td>6.71</td>
<td>6.56</td>
</tr>
<tr>
<td>UNO</td>
<td>7.53</td>
<td>5.38</td>
<td>7.27</td>
<td>7.90</td>
<td>6.76</td>
<td>6.53</td>
<td>5.27</td>
<td>7.02</td>
<td>5.61</td>
<td>6.27</td>
<td>6.31</td>
<td>6.93</td>
<td>6.62</td>
</tr>
<tr>
<td>UniWorld-V2<br>(Edit-R1-Qwen-Image-Edit-2509)</td>
<td>8.41</td>
<td><strong>7.16</strong></td>
<td>8.63</td>
<td>8.24</td>
<td><strong>7.44</strong></td>
<td>6.77</td>
<td>7.52</td>
<td>8.03</td>
<td><strong>7.70</strong></td>
<td><strong>7.24</strong></td>
<td><strong>7.81</strong></td>
<td>7.57</td>
<td>7.69</td>
</tr>
<tr>
<td>Qwen-Image-Edit-2509</td>
<td><strong>8.54</strong></td>
<td>6.85</td>
<td><strong>8.85</strong></td>
<td><strong>8.57</strong></td>
<td>7.32</td>
<td><strong>6.86</strong></td>
<td><strong>7.53</strong></td>
<td><strong>8.13</strong></td>
<td>7.49</td>
<td>7.02</td>
<td>7.76</td>
<td><strong>7.65</strong></td>
<td><strong>7.70</strong></td>
</tr>
<tr>
<td colspan="14" style="background-color: #E6E6FA; text-align: center; font-weight: bold; font-style: italic;">Unified Model</td>
</tr>
<tr>
<td>BAGEL</td>
<td>7.14</td>
<td>5.55</td>
<td>7.49</td>
<td>7.95</td>
<td>6.93</td>
<td>6.21</td>
<td>6.44</td>
<td>7.38</td>
<td>6.87</td>
<td>7.27</td>
<td>6.74</td>
<td>7.20</td>
<td>6.97</td>
</tr>
<tr>
<td>OmniGen2</td>
<td>8.00</td>
<td>6.59</td>
<td>8.31</td>
<td>8.99</td>
<td>6.99</td>
<td>6.80</td>
<td>7.28</td>
<td>8.30</td>
<td>7.14</td>
<td>7.13</td>
<td>7.39</td>
<td>7.81</td>
<td>7.60</td>
</tr>
<tr>
<td>Echo-4o</td>
<td><strong>8.58</strong></td>
<td><strong>7.73</strong></td>
<td>8.36</td>
<td>8.33</td>
<td>7.74</td>
<td>7.18</td>
<td>7.87</td>
<td>8.72</td>
<td>8.01</td>
<td>8.33</td>
<td>8.05</td>
<td>8.14</td>
<td>8.09</td>
</tr>
<tr>
<td><strong>Scone (Ours)</strong></td>
<td>8.52</td>
<td>7.40</td>
<td><strong>8.98</strong></td>
<td><strong>9.73</strong></td>
<td><strong>7.97</strong></td>
<td><strong>7.74</strong></td>
<td><strong>8.20</strong></td>
<td><strong>9.25</strong></td>
<td><strong>8.21</strong></td>
<td><strong>8.44</strong></td>
<td><strong>8.21</strong></td>
<td><strong>8.79</strong></td>
<td><strong>8.50</strong></td>
</tr>
</tbody>
</table>
> - *: GPT-4o responded to 365~370 test cases out of the total 409 cases due to OpenAI safety restrictions.
> - To mitigate randomness, we perform 3 rounds of sampling at 1024x1024 resolution, scoring 3 times per round, yielding 9 group results. The final score is the average of these results.
### Inference
Download the data:
```bash
# pip install -U huggingface_hub
hf download Ryann829/SconeEval --repo-type=dataset --local-dir ../SconeEval
```
Run the script:
```bash
bash scripts/inference_sconeeval.sh
```
### Evaluation
Use GPT-4.1 to evaluate the quality of the generated images and calculate the final score. Please ensure your API key is configured before running the script.
```bash
bash eval/s2i/sconeeval/eval.sh
```
# 🚀 Updates
- [x] Release paper
- [x] Release training code
- [x] Release inference and evaluation code
- [x] Release model weight
- [x] Release training data
- [x] Release SconeEval benchmark
# 🚰 Citation
If you find Scone helpful, please consider giving the repo a star ⭐.
If you find this project useful for your research, please consider citing our paper:
```bibtex
@misc{wang2025sconebridgingcompositiondistinction,
title={Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling},
author={Yuran Wang and Bohan Zeng and Chengzhuo Tong and Wenxuan Liu and Yang Shi and Xiaochen Ma and Hao Liang and Yuanxing Zhang and Wentao Zhang},
year={2025},
eprint={2512.12675},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.12675},
}
```
# 💪 Acknowledgements
This project builds upon the following repositories:
* [BAGEL](https://github.com/ByteDance-Seed/Bagel)
* [OmniContext](https://github.com/VectorSpaceLab/OmniGen2)
Special thanks to these original projects and open-source datasets for their valuable contributions. |