We propose MorphSIG, a training-free framework that reformulates subject-driven image generation (SIG) as pseudo-video feature transport. MorphSIG contains two core modules: Progressive Decoupled Anchoring, which uses frequency-domain decoupling and structural scrambling during early denoising to break structural locking and enable a gradual transition from free composition to identity alignment; and Pseudo-flow Guided Feature Transport, which constructs a latent-space pseudo-flow field to mimic Image-to-Video temporal coherence and transport subject semantics precisely. To address stylistic homogeneity and the lack of ground truth in existing benchmarks, we further introduce a benchmark with diverse stylized subjects. Extensive experiments show that MorphSIG significantly improves pose diversity and subject fidelity over baselines without additional training. Overall, MorphSIG bridges static generation and dynamic propagation, providing a cost-effective paradigm for high-fidelity subject consistency.
Figures & tables
Figure 1: Qualitative comparison between MorphSIG and the baseline.
Figure 2: Comparison of different SIG paradigms. SC (Subject Consistency), and Cost (computational complexity).
Figure 3: Overview of the proposed framework. MorphSIG consists of two modules: Progressive Decoupled Anchoring (PDA) and Warping Feature Transport (WFT). PDA suppresses structural shortcut copying by decoupling and progressively restoring reference anchors during denoising, enabling flexible composition. WFT estimates pseudo-flow to warp and transport reference features into the generation space, improving identity consistency under large pose and geometry variations.
Figure 4: Analysis of High-frequency Shortcut in T&I2I and Temporal Feature Transport in I2V.
Figure 5: Synergistic Analysis of Spectral Decoupling and Structural Permutation.
Figure 6: Feature Energy Analysis.
Methods
Venue
CSD ↑
CIDS ↑
PA ↑
AES ↑
CP ↓
INC ↑
Cross
Self
Cross
Self
Scene
Shot
CI
IA
Avg.
DreamO Mou et al. (2025)
SIGGA’25
0.489
0.584
0.542
0.643
2.99
2.21
1.72
1.65
2.14
5.31
0.533
9.21
DiptychP Shin et al. (2025)
CVPR’25
0.394
0.462
0.417
0.536
2.73
2.04
1.61
1.50
1.97
4.60
0.524
13.18
XVerse Chen et al. (2025a)
NIPS’25
0.229
0.461
0.342
0.531
2.96
2.37
1.89
1.73
2.23
4.92
0.519
13.67
UNO Wu et al. (2025d)
ICCV’25
0.513
0.576
0.528
0.632
3.08
2.52
2.07
1.91
2.39
4.94
0.542
8.27
OmniGen2 Wu et al. (2025a)
arXiv’25
0.446
0.546
0.460
0.587
3.12
2.46
1.80
1.64
2.27
5.19
0.531
9.87
Table 1: Quantitative Results of Various SIG Methods on ViStoryBench Zhuang et al. (2025) . CI: Character Interaction, IA: Individual Action; The first , second , and third values are highlighted.
Methods
Venue
DreamSim ↓
PA ↑
CLIP-T ↑
CLIP-I ↑
CDS ↑
AES ↑
VRAM ↓
GT
Ref
Scene
CI
IA
Avg.
DreamO Mou et al. (2025)
SIGGA’25
0.273
0.596
3.04
2.43
2.07
2.51
0.361
0.834
0.373
6.00
9.09
DiptychP Shin et al. (2025)
CVPR’25
0.459
0.651
2.72
2.16
2.45
2.44
0.332
0.696
0.339
5.24
30.01
XVerse Chen et al. (2025a)
NIPS’25
0.339
0.660
2.95
2.53
2.12
2.53
0.358
0.767
0.366
6.09
37.39
UNO Wu et al. (2025d)
ICCV’25
0.315
0.604
3.21
2.59
2.28
2.69
0.345
0.802
0.366
5.64
35.09
OmniGen2 Wu et al. (2025a)
arXiv’25
0.287
0.607
3.28
2.67
2.41
2.78
0.363
0.822
0.380
5.47
9.92
Table 2: Quantitative Results of Various SIG Methods on SIGBench. VRAM: Using GPU memory during inference. The first , second , and third values are highlighted.
Figure 7: Qualitative Comparison on ViStoryBench Across Diverse Characters and Styles.
Figure 8: Qualitative Comparison on Multi-Character Story Generation in ViStoryBench.
Figure 9: Qualitative Comparison on SIGBench with Ground Truth.
Method
DreamO
DiptychP
XVerse
UNO
OmniGen2
Scone
UMO
MOSAIC
FLUX.2-4B
Ours
Percent (%) ↑
2.79
1.41
0.62
10.19
3.03
12.84
15.95
4.48
22.75
26.31
Table 3: Human preference comparison among different methods.
CIDS ↑
INC ↑
PA ↑
AES ↑
w/o PDA
0.614
13.41
2.37
5.30
w/o SP
0.556
13.64
2.72
5.33
w/o SD
0.618
13.57
2.39
5.36
High-pass only
0.485
12.81
1.92
5.05
Full PDA
0.629
13.88
2.58
5.39
Table 4: The ablation study of PDA.
CIDS ↑
INC ↑
PA ↑
AES ↑
w/o WFT
0.541
13.47
2.64
5.25
w/o MNN
0.578
13.32
2.47
5.18
w/o HF-WTA
0.601
13.65
2.52
5.14
w/o LF-SLERP
0.584
13.30
2.34
5.19
ADD
0.561
12.93
2.16
5.05
Full WFT
0.629
13.88
2.58
5.39
Table 5: The ablation study of WFT.
Figure 10: Qualitative Comparison of Ablation Study.
Figure 11: Ablation analysis. (a) Effect of neighborhood size K in Local Consensus. (b, c) Sensitivity of PDA and WFT scheduling ranges.
Method
DreamO
DiptychP
XVerse
UNO
OmniGen2
Scone
UMO
MOSAIC
FLUX.2-4B
Ours
Training-free
×
✓
×
×
×
×
×
×
–
✓
Latency (s)
76.9
169.4
126.2
82.6
171.7
98.5
84.3
154.7
115.6
130.1
VRAM (GB)
9.09
30.01
37.39
35.09
9.92
30.43
33.78
35.79
12.21
13.44
Table 6: Comparison of computational cost on a single NVIDIA A100 GPU.