VIEScore: Towards Explainable Metrics for Conditional Image Synthesis Evaluation

♠️Max Ku, ♠️Dongfu Jiang, ♠️Cong Wei, ♡Xiang Yue, ♠️Wenhu Chen
♠️University of Waterloo, ♡IN.AI Research {m3ku, dongfu.jiang, c58wei}@uwaterloo.ca, xiangyue@in.ai, wenhu.chen@uwaterloo.ca

ACL 2024

Code arXiv 🖼️ ImagenHub
VIEScore

Figure 1: Metrics in the future would provide not just the score but also the rationale, enabling the understanding of each judgment. Which method (VIEScore or traditional metrics) is “closer” to the human perspectives?

Abstract

In the rapidly advancing field of conditional image generation research, challenges such as limited explainability lie in effectively evaluating the performance and capabilities of various models. This paper introduces VIEScore, a Visual Instruction-guided Explainable metric for evaluating any conditional image generation tasks. VIEScore leverages general knowledge from Multimodal Large Language Models (MLLMs) as the backbone and does not require training or fine-tuning. We evaluate VIEScore on seven prominent tasks in conditional image tasks and found: (1) VIEScore (GPT4-o) achieves a high Spearman correlation of 0.4 with human evaluations, while the human-to-human correlation is 0.45. (2) VIEScore (with open-source MLLM) is significantly weaker than GPT-4o and GPT-4v in evaluating synthetic images. (3) VIEScore achieves a correlation on par with human ratings in the generation tasks but struggles in editing tasks. With these results, we believe VIEScore shows its great potential to replace human judges in evaluating image synthesis tasks.

How VIEScore works?

Traditional metrics lack task awareness! It also lacks reasoning ability. VIEScore tackles both downsides. In VIEScore, All input conditions, synthesized images, and rating instructions are fed together to the MLLM in one pass. Then we retrieve the score in any scale and the rationale.

MY ALT TEXT

Figure 2: Process of MLLM evaluation on one synthetic image in VIEScore. All input conditions, synthesized images, and rating instructions are fed together to the MLLM in one pass. Multi-concept image composition task is used here as an example.

To evaluate the effectiveness of our method, we used the 7 conditional Image generation tasks from ImagenHub and computed the correlation with human ratings collected in ImagenHub.

MY ALT TEXT

Figure 3: We study the correlation between MLLMs and human perspectives on rating images across all tasks in ImagenHub.

How effective is VIEScore with current state-of-the-art MLLMs?

But how well can Multimodal large language models access different tasks of conditional Image generation? We reported that the best model GPT4v’s performance is significantly better than the open-source models. Most open-source MLLMs failed to adapt to our VieScore except LLaVA.

MY ALT TEXT

Table 1: Correlations across all tasks with different backbone models. We highlight the highest correlation numbers in green. Visit our paper for insights and challenges in VIEScore.

How is Traditional Metrics correlating with human compare to VIEScore?

Looking into the details, we found that GPT4v achieves on par with human ratings on text-to-image task but it straggles on image editing tasks. We also compared with the traditional metrics.

Method Method-HumanSCcorr Method-HumanPQcorr Method-HumanOcorr
Text-guided Image Generation Model (5 models)
Human Raters - Unknown 0.5044 0.3640 0.4652
CLIP-Score -0.0817 -0.0114 -0.0881
VIEScore(GPT-4o0shot) 0.4989 0.2495 0.3928
VIEScore(GPT-4o1shot) 0.5124 0.0336 0.4042
VIEScore(Gemini-Pro0shot) 0.5123 0.1842 0.4356
VIEScore(Gemini-Pro1shot) 0.4757 0.2206 0.4326
VIEScore(GPT-4v0shot) 0.4885 0.2379 0.4614
VIEScore(GPT-4v1shot) 0.4531 0.1770 0.3801
VIEScore(LLaVA0shot) 0.1809 0.0306 0.1410
VIEScore(LLaVA1shot) 0.1789 -0.0020 0.1309
Mask-guided Image Editing Model (4 models)
Human Raters 0.5390 0.5030 0.4981
LPIPS -0.1012 0.0646 -0.0694
VIEScore(GPT-4o0shot) 0.5421 0.3469 0.4769
VIEScore(GPT-4o1shot) 0.5246 0.1272 0.4432
VIEScore(Gemini-Pro0shot) 0.4304 0.2839 0.3593
VIEScore(Gemini-Pro1shot) 0.4595 0.3170 0.4017
VIEScore(GPT-4v0shot) 0.4508 0.2859 0.4069
VIEScore(GPT-4v1shot) 0.4088 0.2352 0.3810
VIEScore(LLaVA0shot) 0.1180 -0.0531 0.0675
VIEScore(LLaVA1shot) 0.1263 -0.0145 0.1040
Text-guided Image Editing Model (8 models)
Human Raters 0.4230 0.5052 0.4184
LPIPS 0.0956 0.2504 0.1142
VIEScore(GPT-4o0shot) 0.4062 0.4863 0.3821
VIEScore(GPT-4o1shot) 0.3684 0.1939 0.3438
VIEScore(Gemini-Pro0shot) 0.2836 0.4291 0.2728
VIEScore(Gemini-Pro1shot) 0.2805 0.4657 0.2648
VIEScore(GPT-4v0shot) 0.2610 0.4274 0.2456
VIEScore(GPT-4v1shot) 0.2428 0.3402 0.2279
VIEScore(LLaVA0shot) 0.0448 0.0583 0.0273
VIEScore(LLaVA1shot) 0.0185 -0.0107 0.0258
Subject-driven Image Generation Model (4 models)
Human Raters 0.4780 0.3565 0.4653
DINO 0.4160 0.1206 0.4246
CLIP-I 0.2961 0.1694 0.3058
VIEScore(GPT-4o0shot) 0.4806 0.2576 0.4637
VIEScore(GPT-4o1shot) 0.4685 -0.0171 0.4292
VIEScore(Gemini-Pro0shot) 0.2906 0.1765 0.2851
VIEScore(Gemini-Pro1shot) 0.3486 0.2800 0.3342
VIEScore(GPT-4v0shot) 0.3979 0.1903 0.3738
VIEScore(GPT-4v1shot) 0.2757 0.2261 0.2753
VIEScore(LLaVA0shot) 0.0326 -0.0303 0.1219
VIEScore(LLaVA1shot) 0.1334 0.0858 0.1248
Subject-driven Image Editing Model (3 models)
Human Raters 0.4887 0.2986 0.4747
DINO 0.3022 -0.0381 0.3005
CLIP-I 0.2834 0.1248 0.2813
VIEScore(GPT-4o0shot) 0.4800 0.3734 0.3268
VIEScore(GPT-4o1shot) 0.3862 0.1273 0.3268
VIEScore(Gemini-Pro0shot) 0.2187 0.3148 0.2234
VIEScore(Gemini-Pro1shot) -0.0083 0.3181 0.0004
VIEScore(GPT-4v0shot) 0.3274 0.2960 0.1507
VIEScore(GPT-4v1shot) -0.0255 0.1572 -0.0139
VIEScore(LLaVA0shot) 0.0360 -0.0073 0.0168
VIEScore(LLaVA1shot) 0.0587 -0.0249 0.0309
Multi-concept Image Composition Model (3 models)
Human Raters 0.5927 0.5145 0.5919
DINO 0.0979 -0.1643 0.0958
CLIP-I 0.1512 -0.0963 0.1498
VIEScore(GPT-4o0shot) 0.4516 0.2751 0.4136
VIEScore(GPT-4o1shot) 0.4120 -0.0141 0.3523
VIEScore(Gemini-Pro0shot) 0.3557 0.1948 0.3314
VIEScore(Gemini-Pro1shot) 0.4151 0.1798 0.4131
VIEScore(GPT-4v0shot) 0.3209 0.3025 0.3346
VIEScore(GPT-4v1shot) 0.1859 0.1185 0.1918
VIEScore(LLaVA0shot) 0.1022 0.1194 0.1070
VIEScore(LLaVA1shot) 0.0828 0.0379 0.0293
Control-guided Image Generation Model (2 models)
Human Raters 0.5443 0.5279 0.5307
LPIPS 0.3699 0.4204 0.4133
VIEScore(GPT-4o0shot) 0.4972 0.4892 0.5439
VIEScore(GPT-4o1shot) 0.5544 0.3699 0.5238
VIEScore(Gemini-Pro0shot) 0.3254 0.3359 0.2960
VIEScore(Gemini-Pro1shot) 0.2677 0.4392 0.3240
VIEScore(GPT-4v0shot) 0.4360 0.4975 0.3999
VIEScore(GPT-4v1shot) 0.3892 0.4132 0.4237
VIEScore(LLaVA0shot) 0.2207 0.1060 0.1679
VIEScore(LLaVA1shot) 0.1121 0.0247 0.0416

Table 2: Correlations comparison of available methods. We highlight the best method and the correlation numbers closest to human raters. To conclude, VIEScore is the best metric in evaluating synthetic images across all tasks with high potential. DINO on the other hand proves to be an effective metric in Subject-Driven image generation and editing tasks.

MLLM struggles on rating image editing tasks

Why MLLMs struggle on rating image editing tasks? This marked the disability of MLLM’s evaluation on multiple images, as MLLMs are often confused with multiple images, such as failing to spot the difference between two images. But we believe VieScore will be a mainstream evaluation method in the near future, as in practice multiple image QA is an area coming with many applications and the MLLMs will be improved in the future.

MY ALT TEXT MY ALT TEXT

Citation

Please kindly cite our paper if you use our code, data, models or results: 

@misc{ku2023viescore,
                title={VIEScore: Towards Explainable Metrics for Conditional Image Synthesis Evaluation}, 
                author={Max Ku and Dongfu Jiang and Cong Wei and Xiang Yue and Wenhu Chen},
                year={2023},
                eprint={2312.14867},
                archivePrefix={arXiv},
                primaryClass={cs.CV}
            }