Organizations: Beijing Institute of Technology, Beijing, China · The Hong Kong Polytechnic University, Hong Kong, China · Li Auto, Beijing, China · Tsinghua University, Beijing, China
Image generation technology can synthesize condition-specific images to supplement real-world industrial anomaly data and enhance anomaly detection model performance. Existing generation techniques rarely account for the pose and orientation of industrial components in assembly, making the generated images difficult to utilize for downstream application. To solve this, we propose a novel image synthesis approach, called PostureObjectStitch, that achieves accurate generation to meet the requirement of industrial assembly. A condition decoupling approach is introduced to separate input multi-view images into high-frequency, texture, and RGB features. The feature temporal modulation mechanism adapts these features across diffusion model time-steps, enabling progressive generation from coarse to fine details while maintaining consistency. To ensure semantic accuracy, we introduce a conditional loss that enhances critical industrial elements and a geometric prior that guides component positioning for correct assembly relationships. Comprehensive experimental results on the MureCom dataset, our newly contributed DreamAssembly dataset, and the downstream application validate the outstanding performance of our method.
Figures & tables
Figure 1. Fantastic application of our proposed PostureObjectstitch in industrial anomaly generation considering assembly relationships. Compared to other methods, we simultaneously consider background retention, detail restoration, and assembly relationships. Enjoying the baseball game from the third-base seats. Ichiro Suzuki preparing to bat.
Figure 2. Overview of PostureObjectstitch. Given N reference images of a specific sample, PostureObjectstitch fine-tunes the model to generate high-quality anomaly images for industrial assembly scenarios. Our method has three key components: First, we design a condition decoupling strategy that splits input information into high frequency features, texture features, and RGB features to better suit industrial generation needs. Second, we propose a feature temporal modulation mechanism that allows different features to work properly across diffusion timesteps, enabling gradual generation from coarse to fine details. Finally, we introduce positional pose priors and conditional loss. The former ensures accurate component positioning and correct assembly relationships, while the latter improves generation quality of industrial information like text.
Figure 3. OCR auxiliary loss.
Figure 4. Pose and orientation prior fusion.
Figure 5. Dreamassembly dataset overview. Background: The background images are collected from real industrial environments on actual production lines. These backgrounds cover typical automotive assembly scenarios including assembly workstations, inspection platforms, and industrial equipment. Foreground: The foreground section presents 25 common key components in industrial assembly. Each component is captured under various poses including different rotation and tilt angles.
Figure 6. Visualization of different methods on the MureCom dataset. For each row, we present the background image with bounding box and a reference image, followed by the generation outputs of IC-custom, anydoor, Objectstitch, MureObjectstitch and our PostureObjectstitch method.
Method
CLIP-I ↑
DINO ↑
LPIPS ↓
SSIM ↑
ObjectStitch
0.7891
0.4831
0.0683
0.7636
MureObjectStitch
0.8139
0.5729
0.0694
0.7633
Anydoor
0.8118
0.6281
0.0676
0.8766
IC-Custom
0.7891
0.5092
0.0900
0.8452
Ours
0.8143
0.5885
0.0645
0.7703
Table 1. Quantitative comparison on MureCom dataset.
Method
CLIP-I ↑
DINO ↑
LPIPS ↓
SSIM ↑
ObjectStitch
0.7621
0.4103
0.0512
0.8268
MureObjectStitch
0.8033
0.5621
0.0411
0.8366
Anydoor
0.8789
0.7773
0.1677
0.8364
IC-Custom
0.7011
0.2258
0.4043
0.5458
Ours
0.8563
0.6651
0.0230
0.9821
Table 2. Quantitative comparison on our Dreamassembly dataset.
Figure 7. Visualization of different methods on our DreamAssembly dataset. For each row, we present the background image with mask and a reference image, followed by the generation outputs of IC-custom, anydoor, Objectstitch,MureObjectstitch and our PostureObjectstitch method.
Method
Acc
Precision
Recall
F1-score
ObjectStitch
0.6667
0.7345
0.6667
0.6991
MureObjectStitch
0.6800
0.7954
0.6800
0.7331
Anydoor
0.7200
0.7909
0.7200
0.7537
IC-Custom
0.2133
0.2055
0.2133
0.2093
Ours
0.8000
0.8437
0.8000
0.8211
Table 3. Quantitative comparison on YOLOv5 classification.
Method Components
Evaluation Metrics
TFD
TM
POP
OAL
CLIP-I ↑
DINO ↑
LPIPS ↓
SSIM ↑
0.8033
0.5621
0.0411
0.8366
✓
0.8212
0.6012
0.0422
0.8582
✓
✓
0.8401
0.6715
0.0485
0.8802
✓
✓
✓
0.8425
0.6612
0.0225
0.9815
✓
✓
✓
✓
0.8563
0.6651
0.0230
0.9821
Table 4. Ablation Study Results. TFD is target feature distanglement and TM is time-feature modulation. POP means pose and orientation prior. OAL means the OCR auxiliary loss.