Table of Links
1.3 Our Work and Contributions and 1.4 Organization
-
Related Work
-
Prosecutor Design
-
OS2a: Objective Service Assessment for Mobile AIGC
-
OS2A on Prosecutor: Two-Phase Interaction for Mobile AIGC
-
Implementation and Evaluation
7.1 Implementation and Experimental Setup
7.2 Prosecutor Performance Evaluation
5.3 Subjective Experience of AIGC Outputs
Besides the objective KPIs, clients also care about whether the generated AIGC outputs can meet their personal requirements, which we coin as the experiences of the AIGC service outcome. Considering the subjectivity in artwork assessment, clients evaluate the AIGC outputs using their own standards and preferences. Nonetheless, regardless of which kind of model is used, the final output should be converted to a boolean judgment, whose value is 0 or 1. Value 1 means that the client receives a satisfying AIGC output. In contrast, value 0 indicates that the received output suffers from flaws, e.g., low resolution, distortion, and mismatched art style. Then, an intuitive way is to average the judgment items from all the clients and calculate the reputation score for every MASP. However, recall that these clients hold different levels of strictness. We should calibrate the judgment from numerous heterogeneous clients and generate a fair overall reputation for each MASP. To reach this goal, we use WMSL [16] to calibrate the judgment, as shown in Algorithm 2.
5.3.1 Local Opinion
5.3.2 Reference Opinion and Calibration
5.3.3 Opinion Fusion and Reputation
![Fig. 6: The perceivable experience of AIGC output quality. Images (a) and (b) are trained on AnyLoRA-Checkpoint model using the same prompt: Arcane style, 1 girl, pink hair, long hair, one braid, white shirt, coat, yellow eyes, looking at viewer, city street. However, the model for Image (b) is fine-tuned by Arcane Style LoRA. Image (c) shows some references of Arcane style from Google. This example is from [55].](https://hackernoon.imgix.net/images/fWZa4tUiBGemnqQfBGgCPf9594N2-th93240.png?auto=format&fit=max&w=1920)
5.3.4 Discussion: Rationale of Reputation
Authors:
(1) Yinqiu Liu, School of Computer Science and Engineering, Nanyang Technological University, Singapore ([email protected]);
(2) Hongyang Du, School of Computer Science and Engineering, Nanyang Technological University, Singapore ([email protected]);
(3) Dusit Niyato, School of Computer Science and Engineering, Nanyang Technological University, Singapore ([email protected]);
(4) Jiawen Kang, School of Automation, Guangdong University of Technology, China ([email protected]);
(5) Zehui Xiong, Pillar of Information Systems Technology and Design, Singapore University of Technology and Design, Singapore ([email protected]);
(6) Abbas Jamalipour, School of Electrical and Information Engineering, University of Sydney, Australia ([email protected]);
(7) Xuemin (Sherman) Shen, Department of Electrical and Computer Engineering, University of Waterloo, Canada ([email protected]).
This paper is
