notesum.ai

Published at December 4

DynamicControl: Adaptive Condition Selection for Improved Text-to-Image Generation

cs.CV

Released Date: December 4, 2024

Authors: Qingdong He¹, Jinlong Peng¹, Pengcheng Xu², Boyuan Jiang¹, Xiaobin Hu¹, Donghao Luo¹, Yong Liu¹, Yabiao Wang¹, Chengjie Wang¹, Xiangtai Li³, Jiangning Zhang⁴

Aff.: ¹Youtu Lab, Tencent; ²Western University; ³Nanyang Technological University; ⁴Zhejiang University

Arxiv: http://arxiv.org/pdf/2412.03255v1

Condition

(Metric)

Canny

(F1 Score

\uparrow

)

Hed

(SSIM

\uparrow

)

Openpose

(mAP

\uparrow

)

Depth

(RMSE

\downarrow

)

Seg. Mask

(mIoU

\uparrow

)

Dataset

T2I Model

MultiGen-20M

ADE20K

COCO-Stuff

ControlNet [63]

SDXL

40.01

T2I-Adapter [35]

SDXL

28.03

63.89

39.76

T2I-Adapter [35]

SD1.5

23.66

60.17

48.40

12.60

Gligen [29]

SD1.4

26.92

0.5641

69.88

38.82

23.77

Uni-ControlNet [65]

SD1.5

27.31

0.6912

72.71

40.66

19.39

UniControl [39]

SD1.5

30.83

0.7967

75.87

39.17

25.45

ControlNet [63]

SD1.5

34.66

0.7622

32.56

27.47

Cocktail [16]

SD1.5

35.22

0.8152

78.82

35.90

36.55

29.68

ControlNet++ [27]

SD1.5

37.04

0.8097

28.32

43.64

34.56

Ours

SD1.5

39.26

0.8376

82.63

23.21

48.56

37.78