INTRODUCTIONThe emergent large language models (LLMs) such as ChatGPT and Bard, as well as their excellent open-source counterparts (e.g., LLaMA (Touvron et al., 2023), MPT (Team, 2023)), have served as powerful general-purpose assistants, which opens a new era for artificial intelligence (AI) from targeting specific tasks towards general intelligence. Following the advancements of LLMs, multimodality large language models (MLLMs), as represented by LLaVA (Liu et al., 2023b), MiniGPT-4 (Zhu et al., 2023), InstructBLIP (Dai et al., 2023), and Otter (Li et al., 2023a), have brought exciting progresses on the vision field as well. They are capable of providing robust general-level abilities on visual perception/understanding and can even seamlessly dialog and interact with humans through natural language. While such abilities of MLLMs have been explored and validated on several vision-language tasks such as image captioning (Chen et al., 2015), visual question answering (Antol et al., 2015), cross-modality grounding (Peng et al., 2023), and traditional vision tasks such as image classification or segmentation (Lai et al., 2023), most attention is paid to the high-level perception and understanding of visual contents. Meanwhile, the ability of MLLMs remains not clear on low-level visual perception and understanding, which play significant roles in image quality assessment (IQA) (Hosu et al., 2020;Fang et al., 2020) and its associated tasks on perceiving visual distortions (noises, blurs) (Su et al., 2021;Wu et al., 2023d) and other low-level attributes (color, lighting, composition, style, etc)  (Kong et al., 2016) that may relate to aesthetics and emotions of natural photos (Murray et al., 2012) and human preferences on emerging computer-2020). (2) global perception (e.g., sharpness of the whole picture) vs local content-related in-context perception (e.g., whether the red flower is in focus) (Li et al., 2019). With three types of questions and four quadrants of concerns, the proposed LLVisionQA dataset provides a holistic, diverse, and balanced benchmark for the perception ability on low-level visual attributes of MLLMs.For the description ability (A2), given that the output description is expected to be complex (without fixed formats), we propose the LLDescribe dataset by inviting experts to write long golden low-level descriptions (average 58 words per description) for 499 images, which serve as the reference texts for the single-modal GPT to evaluate MLLM output descriptions. The quality of MLLM descriptions is evaluated through three dimensions: completeness (punish missing information), preciseness (punish outputs controversial with reference), as well as relevance (punish outputs irrelevant to lowlevel attributes). With golden descriptions and the multi-dimensional evaluation process participated by GPT, we comprehensively evaluate the low-level description ability of MLLMs.Besides the two emerging language abilities, we also evaluate MLLMs on the traditional IQA task, a more abstract task that requires understanding on human opinions of low-level attributes, as follows:• Ability 3 (A3): Precise Assessment Aligned with Human Opinions. As depicted in Fig. 1(c), an MLLM should be able to predict quantifiable quality scores for images, which can be aligned with the human-rated mean opinion scores (MOS) on low-level visual appearances.For the assessment ability (A3), we utilize plenty of existing IQA databases (Hosu et al., 2020;Lin et al., 2019;Li et al., 2023c) that focus on various low-level appearances of images, to benchmark MLLMs within conventional IQA settings. Specifically, we notice that MLLMs encounter difficulties in providing sufficiently quantifiable outputs, whether instructed to directly rate with texts or provide numerical outputs. To solve this challenge, we propose to extract the softmax pooling result on the logits of the two most frequent tokens (good and poor) under the response template of MLLMs (Fig 1 (c)) as their quality predictions. Our studies prove that the proposed softmax-based strategy is generally better correlated with human perception than direct token outputs of MLLMs (via argmax), which bridges between these emergent MLLMs and the traditional IQA task settings. Under this strategy, we evaluate all MLLMs on their precise assessment ability by measuring the correlations between their predictions and human opinion scores in various IQA databases.In summary, we systematically explore the potential of MLLMs on three low-level visual abilities: perception, description, and assessment. The three realms compose into the proposed Q-Bench, a MLLM benchmark on low-level visual tasks. Our contributions can be summarized as three-fold:• We build a benchmark for MLLMs on low-level perception ability. To achieve this, we construct a first-of-its-kind balanced and comprehensive LLVisionQA dataset with 2,990 images with one low-level-related question-answer pair for each image. The LLVisionQA includes three question types and four quadrants of low-level concerns to ensure diversity. • We define a benchmark process to evaluate the low-level description ability of MLLMs, including an LLDescription dataset of 499 images with expert-labelled long golden quality descriptions, and a GPT-assisted evaluation to rate MLLM-descriptions in terms of completeness, preciseness, and relevance compared with golden descriptions. • To evaluate precise quality assessment ability, we propose a unified softmax-based quality prediction strategy for all MLLMs based on their probability outputs. With its effectiveness validated in our experiments, the proposed strategy sets up a bridge between generalpurpose MLLMs and traditional IQA tasks that requires quantifiable scores as outputs.2 CONSTRUCTING THE Q-BENCH
GENERAL PRINCIPLESFocusing on Low-level Visual Abilities of MLLMs. Unlike existing MLLM benchmarks (Li et al., 2023b;Liu et al., 2023c;Lu et al., 2023) that aim at all-round abilities, the tasks in Q-Bench are constrained with two basic principles: (1) Requiring perception and/or understanding on lowlevel attributes of images;(2) Not requiring reasoning (i.e. why) or outside knowledge (Marino et al., 2019). We adhere to the principles in designing the perception, description, and assessment tasks, making the proposed Q-bench a focused reflection on the low-level visual abilities of MLLMs.  Covering Diverse Low-level Appearances. To cover diverse low-level appearances, we collect multi-sourced images for each task, as depicted in Tab. 1. Among all images in the perception and description tasks, two-thirds are in-the-wild images directly collected from social media posts, smartphones or professional photography. The rest one-third images are collected after various artificial distortions, or via generative processes (CGI, AIGC). Furthermore, we employ k-means clustering for the low-level attribute indicators to certify that the sub-sampled images retain high diversity. In the assessment task, full images of 7 IQA datasets within all three source types are evaluated through traditional IQA metrics. The diverse and multiple sources of images morph the Q-bench into a holistic and balanced benchmark to fairly evaluate low-level-related abilities.
BENCHMARK ON LOW-LEVEL PERCEPTION ABILITYIn the first task of Q-Bench, we evaluate the low-level perception ability of MLLMs to examine whether they can answer simple natural queries related to low-level attributes. For this purpose, we first collect 2,990 images (I) from multiple sources (see Table 1) with diverse low-level concerns.Then, we collect one low-level-related question (Q), one correct answer to the question (C), and 1-3 candidate false answers (F) for each image. The 2,990 (I,Q,C,F) tuples compose into the LLVi-sionQA dataset (as illustrated in Fig. 2), the first visual question answering (VQA) dataset in the low-level computer vision field. Specifically, the questions in LLVisionQA cover four quadrants of distinct low-level concerns (in Sec. 2.2.1) and three question types (in Sec. 2.2.2). After constructing the dataset, the (I,Q,C,F) are together fed into MLLMs for evaluation, while their outputs are further examined by GPT to judge correctness (in Sec. 2.2.3). The details are elaborated as follows.
QUADRANTS FOR LOW-LEVEL VISUAL CONCERNSAxis 1: Distortions vs Other Low-level Attributes. The primary axis differentiates two categories of low-level perceptual attributes: 1) technical distortions (Su et al., 2021), seen as the low-level characteristics that directly degrade the quality of images (Ying et al., 2020), and 2) aesthetic-related other low-level attributes (Kong et al., 2016;Hou et al., 2023) which are discernible to human perception and evoke varied emotions. Several studies (Talebi & Milanfar, 2018;Ying et al., 2020;Guha et al., 2020) follow this paradigm and categorize them through a relative golden standard, that whether the attributes directly improve or degrade picture quality (Yes→Distortions; No→Others). Despite this standard, we also enumerate common types of distortions vs other low-level attributes as extra guidance for constructing the LLVisionQA dataset, as listed in Sec. A.1.2.Axis 2: Global Perception vs Local In-context Perception. In recent research on low-level vision, it is observed that human perceptions of low-level visuals often intertwine with higher-level contextual comprehension (Li et al., 2019;Wang et al., 2021;Wu et al., 2023a  
QUESTION TYPESIn the LLVisionQA dataset, we curate three question types, Yes-or-No, What, and How to simulate multiple query forms from humans. The details of the three question types are defined as follows.Type 1: Yes-or-No Questions. The fundamental type of questions is Yes-or-No, i.e., judgments. Specifically, we notice that some MLLMs especially prefer to respond with yes rather than no. To reduce such biases in our benchmark, though designing questions with answers as yes is easier, we ensure that around 40% of all judgments are with correct answers as no, via querying on contrastive low-level attributes or non-existing low-level attributes. We further measure the bias levels of different MLLMs and present a further de-biased evaluation among them, as discussed in Sec. A.3.2.Type 2: What Questions. Despite Yes-or-No judgments, the what questions are also a common type of queries in recent MLLM benchmarks such as Lu et al. (2023). In Q-bench, they classify low-level attributes in pictures (e.g., What distortion occurs in the image?), or associated context given specific low-level appearances (for in-context perception questions, e.g., Which object in the image is underexposed?). Unlike Yes-or-No questions, the What questions examine more comprehensive low-level attribute understanding of MLLMs, by requiring correct perception on multiple attributes.Type 3: How Questions. Despite the two common types, we also include a special type, the How questions, to cover non-extreme appearances (Wu et al., 2023d) of low-level attribute dimensions into our benchmark, as an extension to Yes-or-No questions. As shown in Fig. 2, we can query How is the clarity of the image? for the image with both clear and blurry areas, and answer with Medium. With this special question type, we broaden the Q-bench into finer-grained low-level perception.
GPT-ASSISTED EVALUATION PROCESSAfter constructing the LLVisionQA dataset, we feed it to multiple MLLMs to evaluate their abilities on low-level visual perception. The input format to query MLLMs is exemplified as follows: The correct and wrong answers are shuffled during the actual evaluation. Moreover, while traditional visual question answering (Antol et al., 2015;Marino et al., 2019) tasks typically employ traditional language metrics (BLEU-4, CIDEr) to compare performance, as observed by recent studies (Ye et al., 2023) and validated by us, most MLLMs cannot consistently provide outputs on instructed formats. Given the question above, different MLLMs may reply "A.", "High", "The clarity of the image is high.", "The image is of high clarity." (all correct), which are difficult to be exhaustively-This image has good lighting and vibrant colors, but there is overexposure on the surface and bottom of the petals. The focus is on the central flower, but it is not very sharp, while the texture of the stamen and petals is clear. The two flowers in the background are blurred with indistinct features. The background blur is severe, making it unrecognizable. The overall quality of the image is good.The clarity of this photo is not very high, it has very serious noise, and the pixel count is also relatively low. However, the colors in the image are vibrant and rich. The composition is very dense, with all hot air balloons, and the details are relatively clear. Therefore, the quality of this photo is average.Overall, this image is relatively clear. The content of the shark's head is presented well, and the details and texture are also well presented. However, there is a bit of noise, and the pixels are a bit too large. The motion blur made by rapid movement of another shark in the background also affects the presentation of the content of the front shark. Therefore, the quality of this image is generally poor.This photo was taken underwater. The overall image has a blue color cast. Although the composition is clear and the subject is distinct, the focus is not very good, making it difficult to recognize fine details of the subject. There is also overexposure in the upper right corner of the water's surface. Therefore, the quality of this photo is low.Overall, the focus of this image is somewhat poor. The details of the cow's face are presented very poorly, and the texture is very badly shown. The colors are okay, and the composition is also good. However, the content as a whole is too monotonous and lacks richness. Therefore, I personally believe that the overall quality of this image is very poor.The clarity of this photo is relatively high, with rich details in the facial features of the central figure. However, the top of the head is out of frame and not entirely complete. The background behind is reasonably blurred, with vibrant colors and high saturation.Therefore, this photo has high quality.In-the-wildIn-the-wild Artificially-distorted Artificially-distorted AI-generated CG-generated included under traditional metrics. To solve this problem, we design, validate, and employ a 5-round GPT-assisted evaluation process inspired by Liu et al. (2023c). Under this process, the question, correct answers, and MLLM replies are fed into GPT for evaluation (See Sec. A.2.1 for its details).
BENCHMARK ON LOW-LEVEL DESCRIPTION ABILITYIn the second task of Q-Bench, we evaluate the language description ability of MLLMs on low-level information. This task is a sibling task of image captioning (Chen et al., 2015;Young et al., 2014;Agrawal et al., 2019) that describes image content with natural language, with a specific concern on the low-level appearance of images. To evaluate this ability automatically, we first derive a golden low-level description dataset, denoted as LLDescribe (Sec. 2.3.1), including one long (average 40 words) golden description provided by experts for each of 499 images. With these golden text descriptions, we are able to measure the quality of output low-level descriptions from MLLMs with a single-modal GPT, under the three dimensions: completeness, preciseness, as well as relevance (Sec 2.3.2). The discussions of the golden descriptions and the evaluation process are as follows.
DEFINING Golden LOW-LEVEL DESCRIPTIONS FOR IMAGESFor the description ability, MLLMs should accurately and completely describe low-level visual information of images. Thus, the ground truths for these MLLMs are also built within a basic principle to cover as many low-level concerns as possible, so long as they are enumerated in Sec. 2.2.1 and occur in images. The resulting golden descriptions in LLDescribe have an average duration of 58 words, notably longer than common high-level image caption datasets (11 for Agrawal et al. (2019), 10 for Chen et al. ( 2015)). Similar to the LLVisionQA dataset for the perception task, the 499 images in LLDescribe dataset also include all 10 sources (as in Tab. 1) to cover images with diverse low-level appearances. The golden descriptions on different sources of images are depicted in Fig. 3.
EVALUATION WITH SINGLE-MODAL GPTRecent studies (Zheng et al., 2023) have proved single-modal GPT (OpenAI, 2023) to be a reliable evaluation tool for pure language tasks. Via the LLDescribe dataset, we convert the multi-modality problem into a text-only setting, by matching the MLLM outputs with the golden descriptions with single-modal GPT under three dimensions: (1) Completeness. More matched information with the golden description is encouraged.(2) Preciseness. The controversial information with the golden description is punished.(3) Relevance. More proportions of MLLM outputs should be related to low-level information, instead of others. Each dimension is scored among [0,1,2]. Similar as Sec. 2.2.3, we repeat 5 rounds for each single evaluation and collect the weighted average as the final score. The detailed settings for GPT to evaluate the three dimensions are in Sec. A.2.2. 
ITU-standard RescalingFigure 4: The proposed softmax-based quality assessment strategy for MLLMs. Instead of directly decoding tokens from the [SCORE TOKEN] position, the strategy extracts log probabilities (logits) of good and poor, and predicts quantifiable score via a softmax pooling between the two logits.
BENCHMARK ON PRECISE QUALITY ASSESSMENT ABILITYIn the third task, we benchmark the ability of MLLMs to provide quantitative assessment on the overall low-level appearance of images. Unlike the two tasks above, we utilize existing IQA datasets that are collected across a variety of low-level appearances to evaluate how MLLMs can predict quantitative quality scores aligned with human opinions. All the three types of IQA datasets (inthe-wild, generated, artificially-distorted) as mentioned in Sec. 2.1 are evaluated, to provide a broad range measurement of the assessment ability of MLLMs. Nevertheless, how to collect quantifiable quality scores from MLLMs remains challenging as their outputs only have weak measurability (Sec. 2.4.1). Noticing that MLLMs can provide probabilities of tokens, we employ softmax pooling on the logits of good and poor under a simple and direct prompt template, deriving into quantifiable quality score predicted by MLLMs (Sec. 2.4.2), as illustrated in Fig. 4. Details are as follows.
WEAK MEASURABILITY OF MLLM OUTPUTSIn Q-Bench, we aim to fairly compare the assessment ability between different MLLMs on diverse low-level appearances. Henceforth, our principle is to define a unified, simplest instruction that is applicable for all MLLMs on all IQA datasets. Under this principle, we conduct toy experiments on LLVisionQA on Shikra and LLaVA-v1, with two simple instruction strategies: (A) Direct Instruction, in which the prompt is designed as simple as "Rate the quality of the image". The top-frequency answers are good (78%), and poor (20%), with other outputs almost negligible. (B) Numerical Instruction, in which we specifically instruct numerical ratings, with the prompt: "Score the quality of the image from 1 to 5, with 1 as lowest and 5 as highest.". Under the numerical strategy, the top-frequency answers are 5 (84%), 1 (9%), and 3 (5%); though within the score range, the frequencies of scores 2 and 4 are both less than 1%. The toy experiments imply the weak measurability of MLLM outputs, given that the answers are statistically 1) biased towards positive, 2) biased towards extreme, and 3) with only two effective scales. Therefore, it is necessary to explore extended strategies for MLLMs to provide truly quantifiable outputs for low-level assessment.
A SOFTMAX-BASED EVALUATION STRATEGYGiven the above observations, we design the softmax-based evaluation strategy (Fig. 4) to reduce the negative impacts of the biases and lack of scales. To start with, we design our strategy within the Direct Instruction, which is more general and less biased than the Numerical Instruction. The strategy is based on the observation that two top-frequency outputs, good and poor, can be considered as anchors for better and worse human perception, and the Direct Strategy can be approximated into a binary classification problem on the [SCORE TOKEN] position, or technically, an argmax between the logits of good (x good SCORE TOKEN ) and poor (x poor SCORE TOKEN ) on this position. In our revised strategy, we modify the argmax into softmax to collect better quantifiable scores:q pred = e x good SCORE TOKEN e x good SCORE TOKEN + e x poor SCORE TOKEN(1)This simple and generally-applicable strategy enables us to collect quantifiable outputs (q pred ) from MLLMs with higher correlation to human ratings, as verified in our experimental analysis (Tab. 9). 
RESULTS AND OBSERVATIONS ON PERCEPTIONOpen-Source MLLMs. For a holistic examination on the perception ability of MLLMs, we evaluate the multi-choice correctness of MLLMs on different sub-categories of the LLVision dataset, which is equally divided as dev (Tab. 7, will be released) and test (Tab. 2, will keep private) subsets. We are glad that the majority of MLLMs can significantly outperform random guess on all sub-categories. Considering that all participating MLLMs are without any explicit training on lowlevel visual attributes, these results show strong potentials for these general-purpose models when further fine-tuned with respective low-level datasets. Among all MLLMs, the recently-released InternLM-XComposer-VL reaches the best accuracy on this question-answering task, followed by LLaVA-v1.5, QWen-VL and InstructBLIP (Flan-T5), which show rather close results. By achieving more than 60% accuracy on both subsets, these models show exciting potentials as robust low-level visual assistants in the future. Another key observation is that almost all methods perceive worse on distortions than other low-level attributes. One exception is LLaMA-Adapter-V2, which is the only MLLM that adopts multi-scale features as visual inputs. We also notice that all MLLMs prefer yes than no among Yes-or-No questions, as analyzed in Tab. 8; qualitative comparisons are illustrated in Fig. 10. For Kosmos-2, we specially adopt close-set inference for it, as discussed in Sec. A.2.1.GPT-4V vs Human. To evaluate the low-level perception abilities of the commercial MLLM, GPT-4V, we gauge its accuracy against human using the test subset of LLVision dataset. GPT-4V exhibits competitive performance and outperforms open-source MLLMs by a large margin (+9%), and on par accuracy with the Junior-level Human. Despite its prowess, there is still a way to go for GPT-4V before it can match the overall proficiency of the Senior-level Human (with experiences on low-level visual tasks, 8% better than GPT-4V). Furthermore, across all categories, the results show that GPT-4V, much like its open-source counterparts, faces challenges in recognizing distortions.
RESULTS AND OBSERVATIONS ON DESCRIPTIONFor the description ability, InternLM-XComposer-VL reaches best proficiency again, especially in terms of the relevance dimension. Nevertheless, in the perspective of the completeness and precision of the descriptions, even the best of all MLLMs cannot obtain an excellent score; on the contrary, almost all MLLMs reach an acceptable standard (0.8/2.0). In general, all MLLMs at present are only with relatively limited and primary ability to provide low-level visual descriptions. We also conduct a qualitative comparison for MLLM descriptions in Sec. A.3.3. 
RESULTS AND OBSERVATIONS ON ASSESSMENTTo measure the assessment ability, we evaluate the performance of 15 MLLMs on 7 IQA datasets that are with at least 1,000 images and 15 human ratings per image (itu, 2000). Primarily, we notice that the majority of MLLMs are notably better than NIQE on non-natural circumstances (CGI, AIGC, artificial distortions), showing their potential towards general-purpose evaluators on a broader range of low-level appearances. We also notice that without explicit alignment with human opinions during training, the most excellent MLLM, which is again InternLM-XComposer-VL, can already outperform CLIP-ViT-Large-14 by a large margin (20%), marking the dawn of MLLMs as robust quality evaluators. Furthermore, we also design a synonym ensemble (see Sec. A.2.3) strategy which can further generally improve IQA accuracy of MLLMs, whose results are analyzed in Sec. A.3.5. Despite their proficiency, current MLLMs are still less accurate in finer-grained situations (LIVE-FB, CGIQA-6K) for the assessment task, which could be enhanced in the future.
CONCLUSIONIn this study, we construct the Q-Bench, a benchmark to examine the progresses of MLLMs on lowlevel visual abilities. Anticipating these large foundation models to be general-purpose intelligence that can ultimately relieve human efforts, we propose that MLLMs should achieve three important and distinct abilities: accurate perception on low-level visual attributes, precise and complete language description on low-level visual information, as well as quantitative assessment on image quality. To evaluate the abilities, we collect two multi-modality benchmark datasets for low-level vision, and propose a unified softmax-based quantitative IQA strategy on MLLMs. Our evaluation proves that even without any low-level-specific training, several extraordinary MLLMs still have decent low-level abilities. Nevertheless, there is still a long way to go for MLLMs to be truly-reliable general low-level visual assistants. We sincerely hope that the observations found in the Q-Bench can inspire future MLLMs to enhance the low-level perception and understanding abilities. 
A APPENDIX A.1 MORE INFORMATION ON BENCHMARK DATASETS
A.1.1 SUBJECTIVE EXPERIMENTA total of eleven experts, each with professional skills and extensive experience in photography, are invited to participate in the subjective labeling experiment of Q-Bench. The subjective experiment takes place in a laboratory environment with standard indoor lighting. A Dell-4K monitor, which supports a resolution of 3840 × 2160, is used for displaying the interfaces. The screenshots of interfaces can be referred to in Fig. 5. Each expert annotates up to 30 images a day to avoid fatigue, and every annotation is carefully reviewed by at least three other experts before acceptance. In this way, we ensure the accuracy and rigor of the Q-Bench labels to the greatest extent possible. This, in turn, makes the performance testing capability of Q-Bench more precise and meaningful.A     Considering that it is still the MLLM with fewest parameters among the ten models, its results are decent at its model size. More importantly, they validate that our observation on the prompt failure is reasonable, and we will further delve deeper into this problem of MLLMs in our extended works.
Settings for GPT Evaluation:Given GPT's inherent variability, identical prompts can yield non-definitive responses. To address the impact of such situations on our evaluation, we've implemented a 5-round voting strategy. Under this approach, we pose the same prompt as defined in the following templates five times, taking the popular votes of GPT's answers to determine the final outcome. Our human analysis on a sample set confirms that the 5-round voting strategy improves GPT evaluation accuracy from 93.2% to 98.4%, reducing errors to only 1/4 compared with the single-round evaluation.Prompt Templates for GPT Evaluation:#System: You are a helpful assistant that grades answers related to image quality and aesthetics. There are a lot of special terms or keywords related to image processing and photography. You will pay attention to the context of 'quality evaluation' when grading.#User: Assuming you are a grader, you will now be provided with a question [question] and a set of options [options] with option [options[0]] being the correct answer. Additionally, there will be an answer [answer] provided by a respondent. Please determine whether the respondentś answer is correct considering the context of the question. Even if the word choice is not completely the same, you can decide based on the given options and see whether the one in the answer is close enough to the given correct answer, The result is 1 if the answer is correct and else the result is 0. Please only provide the result in the following format: Result:Examples for GPT Evaluation:(1) "Rephrased" Answers. (Fig. 6)Question: Which is the brightest part in this image?Choices: ['Capital letters E and S', ' ST','18','56'] MLLM Answer:Please rate score 2 for completely relevant, 1 for partly relevant, and 0 for totally irrelevant.Please only provide the result in the following format: Score:
5-Round GPT Answers:["Score: 2","Score: 1","Score: 1","Score: 2","Score: 1"] → Final Score: 1.4 In Algo. 1, we provide an example on how to evaluate image quality with MLLMs. The algorithm is simple with only 9 lines, and could be easily integrated with any new MLLMs (based on causal LLMs), so as to allow these models to quantitatively predict the quality of images.IQA Evaluation Strategy for CLIP-ViT-Large-14:In Tab. 4, we compare the IQA performance of MLLMs with CLIP-ViT-Large-14, the visual backbone of the majority of MLLMs. Attempting to understand whether the new language part (LLM) can do better than the original language part of CLIP, we try to compare between CLIP and MLLMs in a relatively aligned setting. Firstly, noticing that most MLLMs will resize images into 224 × 224 as their input sizes, we align this setting on CLIP, and ignore the strategies as proposed by (Wang et al., 2022). Secondly, same as the strategy on MLLMs, we also apply softmax pooling between good and poor, as in the CLIP's zero-shot classification format: a photo of good quality and a photo of poor quality. Besides the two alignments, similar as existing practices (Wang et al., 2022;Wu et al., 2023b;Zhang et al., 2023c), the quality scores of CLIP-ViT-Large-14 are obtained as follows:q pred,CLIP = e CosineSimilarity(f [IMAGE],f a photo of good quality ) e CosineSimilarity(f [IMAGE] ,f a photo of good quality ) + e CosineSimilarity(f [IMAGE] ,f a photo of poor quality ) (2) Special IQA Settings for Flan-T5-based InstructBLIP:For InstructBLIP (Dai et al., 2023) (Flan-T5-XL), different from the majority of LLaMA-based (or MPT-based Otter-v1) MLLMs, the two top-frequency tokens are high (89%) and low (8%) instead of the common good↔poor. Henceforth, based on our motivation to only modify the argmax into softmax and follow the default top-frequency output tokens of MLLMs, we replace the probabilities of good↔poor into those of high↔low in Eq. 1 for T5, defined as follows:q pred,T5 = e x high SCORE TOKEN e x high SCORE TOKEN + e x low SCORE TOKEN(3)As validated in our experiments (Tab. 10, the high↔low pair generally predicts better than good↔poor on majority of databases. The better performance on MLLM-specific top-frequency tokens by side validates the effectiveness of our methodology for MLLMs on IQA.Further Improving IQA Abilities of MLLMs with Synonym Ensemble:The quality assessment scores for the synonym ensemble strategy can be derived as:q pred = e t∈P t x t SCORE TOKEN e t∈P t x t SCORE TOKEN + e t∈N t x t SCORE TOKEN(4)where P indicates the positive token set (from good, fine, high, etc.), while N represents the negative token set (from poor, bad, low, etc.). The results of different P and N are listed in Tab. 11.
Special Validation Protocol for CGIQA-6K:The CGIQA-6K (Zhang et al., 2023b) dataset contains two separate sub-sets which consist of 3,000 game images and 3,000 movie images respectively, with different instructions for human annotators during its subjective experiments. Therefore, we validate the MLLMs' assessment performance on the two sub-sets individually and average the results for the final exhibition. The results of NIQE and CLIP-ViT-Large-14 are also obtained under the same protocol for a fair comparison. A.3 EXTENDED EXPERIMENTAL RESULTS
A.3.1 ARCHITECTURES OF DIFFERENT MLLMSAs compared in Tab. 6, the 15 variants of MLLMs as evaluated in the Q-Bench are with varying vision and language architectures, as well as the alignment strategies between the two modalities.It can be noticed that all MLLMs are combined with a version of CLIP Radford et al. (2021) and a large language model, which are generally connected under one among three strategies: direct project layers (MLP or linear layer), Q-Former (a transformer to abstract visual features into LLM tokens), or cross-attention (use visual features as conditions for text generation).
A.3.2 EXTENDED RESULTS FOR PERCEPTIONResults on the dev subset:In Tab. 7, we list the results on the dev subset of the LLVisionQA benchmark set for the lowlevel perception task. This subset is planned to be opened to public in the future. Therefore, the performance in it will only be taken as a reference. At present, all MLLMs as evaluated have not yet seen this subset, so it can be taken as a cross-validation with the test subset. From Tab. 7 and Tab. 2, we validate that MLLMs perform pretty similar between the two subsets, suggesting that LLVisionQA is a reliable and stable benchmark set for question answering on low-level vision.Radar Chart for Different MLLMs:In Fig. 9, we show the radar chart to compare the low-level perception abiliies among different MLLMs. Despite the observations as revealed in Sec. 3.1, we also notice two extra fun facts: 1) Adding the content context does not degrade the performance of MLLMs. On the contrary, MLLMs can answer better on in-context questions. This result validates the aforementioned conjectures that appropriate higher-level contexts as prompts may help improve the preciseness of low-level visual perception; 2) MLLMs have strong capabilities of answering what questions, suggesting potential reasoning abilities. In the future, we will excavate more interesting characteristics of MLLMs and try to improve their perception accuracy through better guidance based on these characteristics."Yes or No?": How Biased are MLLMs?In this section, we take a deeper analysis on the Yes-or-No judgment ability of MLLMs, that whether these models can get similar accuracy on questions that should be answered with Yes, as those should be replied as No. Sadly, we notice that all MLLMs have higher prediction accuracy on Yes-questions than No-questions, while some MLLMs are more very severe biased (e.g., IDEFICS-Instruct). Considering that our LLVisionQA dataset contains more (62%) Yes-questions than Noquestions (38%) and may introduce biases while comparing different MLLMs, we further compute a de-biased accuracy for all these methods, as the mean value of the accuracies on two types of questions, and present the respective de-biased rank for all participating MLLMs, as listed in Tab 8.User: Describe the quality, aesthetics and other low-level appearance of the image in details.
Shikra:The image features a portrait of a woman, with her hair parted to the side. Her facial features are realistic and clear, and she appears to be looking to the side with a serious expression. The overall quality of the image is good, with no obvious flaws or distortions in the depiction..
LLaVA-v1:The image is a pencil drawing of a woman with a serious expression. The quality of the drawing is quite good, with a clear and detailed representation of the woman's facial features and hair. The exposure appears to be well-balanced, and there is no significant noise level in the image. The aesthetics of the drawing are characterized by the use of muted colors, which give the image a somewhat somber and serious tone. The composition is well-balanced, with the woman's face taking up most of the space in the drawing. Overall, the image has a professional and polished appearance, with a focus on the woman's facial expression and her overall demeanor.
MiniGPT-4:The image is a portrait of a woman with long, straight black hair and a serious expression. The woman has a pale complexion and dark eyes, which are focused on the viewer. The woman's face is oval in shape and her features are well defined. The woman's nose is straight and her lips are thin and closed. The woman's eyes are large and her eyebrows are thick and straight. The woman's face is symmetrical and her facial expression is serious. The image is well lit and the lighting highlights the woman's features. The image is in color and the colors are muted and subdued. The image is of high quality and the details are well defined. The image is aesthetically pleasing and the overall appearance of the image is professional.
Kosmos-2:The image has a low-quality quality, as it is a painting of a woman with a long, dark hair. The quality of the painting is not as clear as it could be, as the woman's face is not fully visible. The lighting is not wellbalanced, and the overall atmosphere is not pleasant. The painting lacks a sense of style and feels unfinished.
LLaMA-Adapter-V2:The image is a black and white drawing of a woman with a blue background. The quality of the image is clear, with good exposure and no noticeable noise. The aesthetics of the image are simple and minimalistic, with focus on the woman's face and the blue background. The composition is balanced and wellframed, drawing attention to the subject. The style of the drawing is realistic, capturing the essence of the woman's facial features and expression. The overall feeling of the image is serene and contemplative, conveying a sense of calm and introspection.
InstructBLIP-T5:The image depicts a portrait of a woman with long, dark hair and a brown background. The image has a high level of quality and is well-executed in terms of color, contrast, and texture. It also has a strong contrast between the light and dark parts of the image, which creates a sense of depth and dimension. Overall, the image is well-executed in terms of quality, aesthetics, and other low-level appearance.
InstructBLIP-Vicuna:The image is a painting of a woman's face, depicted in a realistic style. The painting features a close-up view of the woman's face, focusing on her eyes, nose, and mouth. Her hair is styled in a simple yet elegant manner, emphasizing her facial features. The overall aesthetics of the painting convey a sense of realism and attention to detail, capturing the subtle nuances of the woman's facial expression and appearance.The painting also showcases the artist's skillful use of color and shading to create a lifelike representation of the subject.Otter-v1:. The image features a portrait of a woman with long hair, painted in an oil painting style. The woman's face is the main focus of the painting, and her features, such as her eyes, nose, and mouth, are welldefined. Her hair is described as dark, which adds depth and contrast to the overall composition. The painting appears to be a close-up of the woman, emphasizing her facial features and creating a sense of depth and detail.
IDEFICS-Instruct:The image is a drawing of a woman with long hair. The woman has a small nose and a small mouth. The hair is brown and the eyes are dark. The drawing is in shades of gray and the background is white. The woman's hair is long and flowing. The drawing is not very detailed and the woman's face is not very clear. The drawing is not very aesthetically pleasing and the woman's face is not very attractive.
mPLUG-Owl:The image is a portrait of a woman with a brownish-yellowish complexion, wearing a white shirt. The woman's face is the focal point of the painting, and her eyes are the most prominent feature. The painting has a soft, dreamy quality, with a warm, inviting atmosphere. The lighting is natural, creating a soft glow on the woman's face and hair. The composition is well-balanced, with the woman's face taking up most of the canvas, and the background providing a sense of depth and context. The style of the painting is realistic, capturing the woman's likeness and emotions in a realistic manner. Overall, the painting is aesthetically pleasing and evokes a sense of calm and serenity.Golden Description: This photo has relatively low clarity, making it difficult to discern the details of the brushstrokes in the center artwork. There is substantial loss of detail, and the overall color saturation is very low, giving it a dim and dull appearance. Therefore, the quality of this photo is low. User: Describe the quality, aesthetics and other low-level appearance of the image in details.
Shikra:The image presents a close-up view of a delicious dessert, which appears to be a piece of fruit or a scoop of ice cream with berries on top. The dessert is topped with whipped cream, giving it an appetizing and visually appealing appearance. In addition to the dessert, there are three red berries on top, adding a touch of color and natural sweetness to the dish. The overall quality of the image is good, with a clear representation of the dessert and the berries, making it an inviting and attractive scene for viewers.
LLaVA-v1:The image quality is high, with a clear and well-exposed shot of a delicious-looking dessert. The dessert is a fruit-based dish, featuring a slice of peach with whipped cream on top. The lighting in the image is bright and evenly distributed, which enhances the colors and overall appearance of dessert. The composition is visually appealing, with the fruit and whipped cream arranged in a pleasing manner. The aesthetics of the image are further enhanced by the use of natural light, which adds a warm and inviting atmosphere to the scene. Overall, the image is visually appealing, with a high level of quality and aesthetics that make it an enjoyable sight for viewers.
MiniGPT-4:The image shows a close-up view of a dessert dish consisting of sliced peaches topped with whipped cream and a sprig of mint. The peaches are fresh and ripe, with a vibrant color and a smooth, glossy texture. The whipped cream is smooth and fluffy, adding a creamy texture and a sweet flavor to the dish. The mint leaf adds a fresh and aromatic touch to the dish. The overall aesthetics of the image are pleasing and appetizing. The bright colors of the peaches and the whipped cream are visually appealing and make the dish look appetizing. The use of fresh ingredients and the attention to detail in the presentation of the dish make it look appetizing and well-made. The lighting in the image is soft and natural, which enhances the colors and texture of the dish. The quality of the image is high, with good resolution and clarity.
Kosmos-2:The fruit is displayed on a white plate, and the dessert is garnished with a sprig of mint and a few berries. The quality of the photo is clear, with no distracting elements or noise. The aesthetics of the dessert are pleasing, with the white plate and the fruit creating a visually appealing scene. The low-light conditions, such as the white background and the bright colors of the fruit, contribute to the overall pleasant appearance of the photograph.
LLaMA-Adapter-V2:The image has a high-quality, clear, and well-exposed photo of a blue and white dessert. The aesthetics of the image are pleasing, with the blue and white color scheme creating a visually appealing and appetizing presentation. The composition of the image is well-balanced, with the dessert placed prominently in the center of the frame, surrounded by a white background. The style of the image is clean and professional, showcasing the dessert in a way that highlights its beauty and appeal. The feeling of the image is inviting and enticing, making it an ideal choice for sharing on social media or as a visual representation of a delicious dessert.
InstructBLIP-T5:The image captures a close-up of a dessert consisting of an apple, a strawberry, and a piece of ice cream. The texture of the ice cream is smooth and creamy, with a hint of sweetness and a hint of tartness.There is also a red berry on top of the ice cream, which adds a touch of color and flavor to the dessert. Overall, the image has a high level of quality and aesthetic appeal.InstructBLIP-Vicuna: In the image, there is a delicious-looking dessert displayed on a white plate. The dessert consists of various fruits, such as peaches and berries, arranged in a visually appealing manner. The fruits are placed on top of each other, creating a colorful and eye-catching display. Additionally, the dessert is adorned with a sprig of fresh mint leaves, adding a touch of freshness and flavor to the dish. Overall, the image showcases a well-presented and visually appealing dessert, highlighting the quality and aesthetics of the dish.Otter-v1:. The image has a high-quality appearance with clear and sharp details, aesthetically pleasing colors, and a balanced composition.
IDEFICS-Instruct:The image is clear and well-exposed, with no visible noise. The composition is wellbalanced, and the colors are vibrant and appealing. The style is artistic, and the feeling is refreshing and invigorating.
mPLUG-Owl:The image is a close-up of a dessert with a slice of fruit on a plate. The dessert is presented in a dish with a white background, which creates a clean and visually appealing appearance. The fruit slice is placed on top of the dessert, adding a touch of color and contrast to the overall scene. The image is well-exposed, with no noticeable noise or distortion. The aesthetics of the image are enhanced by the use of a shallow depth of field, which focuses on the dessert and fruit slice, creating a sense of depth and drawing the viewer's attention to the subject. The image has a simple and elegant style, making it an appealing and visually pleasing representation of the dessert.
Golden Description:The overall clarity of this image is high, with rich and vibrant colors. The composition is exquisite and aesthetically pleasing. The texture details of the photographed fruit subject are abundant and clear, accurately capturing the reflection effect of light. Therefore, the quality of this image is very high. While text knowledge about image quality assessment should have been injected to them (e.g. a blurry image is a low quality image) during their pure-language training stages, we think this should not be regarded as data contamination for IQA, because the images cannot be seen by a language model. Instead, they are important knowledge for MLLMs to better link particular visual attributes (blur) to human opinions (quality), which motivates us to explore MLLMs for these tasks.
C LIMITATIONS AND DISCUSSIONSIn Section A.3.2, we observed that MLLMs frequently respond with 'yes' to Yes-or-No questions.It's worth noting that the current LLVisionQA dataset is skewed, with 62% of its questions being Yes-questions and only 38% being No-questions. This imbalance could introduce biases when comparing various MLLMs. To fully address this, we aim to balance the dataset by preparing a reversed version for each question in our subsequent work, ensuring a less biased evaluation.For the description task, we acknowledge that judging whether a description matches the gold description is a subjective process, which may not have absolute standard. Even when evaluated by humans, the scores rated for the MLLM descriptions are subject to individual differences. Though we have employed the 5-round GPT-assisted evaluation protocol, which could be the most reliable and reproducible way at present, it may still unavoidably contain hallucinations (from GPT). We will continue to explore how to design a more reliable evaluation protocol for the low-level visual description task in our follow-up works.While the proposed Q-Bench has offered a comprehensive evaluation on the low-level visual capabilities of MLLMs, it does not provide direct guidance on enhancing these capabilities. As our next steps, we intend to progressively scale up the LLDescribe and LLVisionQA datasets to eventually allow a reliable low-level visual instruction tuning process that can further improve the low-level abilities for MLLMs.Figure 2 :2Figure 2: A dataset card of LLVisionQA that evaluates the low-level perception ability of MLLMs. 2,990 (I,Q,C,F) tuples are collected to cover three question types and four quadrants of low-level visual concerns, providing an all-around evaluation of low-level visual perception for MLLMs.
#User: How is the clarity of the image? (Question) [IMAGE TOKEN] (Image) Choose between one of the following options: A. High (Correct) B. Medium(Wrong) C. Low(Wrong)
Figure 3 :3Figure 3: A dataset card of LLDescribe that evaluates the low-level description ability of MLLMs. 499 images from 10 diverse sources are labeled with golden descriptions, to serve as text references for single-modal GPT to evaluate the completeness, preciseness, and relevance of MLLM outputs.
Figure 6 :6Figure 6: Image of example (1).Figure 7: Image of example (2).Figure 8: Image of example (3).
Figure 7 :7Figure 6: Image of example (1).Figure 7: Image of example (2).Figure 8: Image of example (3).
Figure 8 :8Figure 6: Image of example (1).Figure 7: Image of example (2).Figure 8: Image of example (3).
Figure 12 :12Figure 12: A qualitative comparison for MLLM descriptions on an AI-generated image.
Figure 13 :13Figure 13: A qualitative comparison for MLLM descriptions on an in-the-wild photograph.


Table 1 :1Overview of the 10 diverse image source datasets in the Q-Bench, and the respective benchmark dataset size for each low-level ability among perception, descrption and assessment. The Corrupted COCO denotes COCO-Captions images corrupted byMichaelis et al. (2019).
Correct Answer (C): Poor(Axis 2)(Axis 1)(Axis 1)(1100)Yes-or-No(911)What(979)HowFurthermore, localized low-level appearances can deviate from their overall counterparts, as observed by Wu et al. (2022); Yinget al. (2021). Acknowledging these differences, we curate local in-context perception (Fig. 2 right)questions, that require MLLMs to grasp the content or other context to answer correctly, while otherquestions are categorized as global perception (Fig. 2 left). (More analysis in Sec. A.1.2.)). For instance, a clear sky might lack complex textures yet display exceptional clarity.
Table 2 :2Results on the test subset for the low-level Perception ability of MLLMs. MLLMs with top-3 performance in each sub-category and the overall LLVisionQA is emphasized with boldface.Sub-categoriesQuestion TypesQuadrants of Low-level ConcernsModel (variant)Yes-or-No↑What↑How↑Distortion↑Other↑In-context Distortion↑In-context Other↑Overall↑random guess50.00%28.48% 33.30%37.24%38.50%39.13%37.10%37.94%LLaVA-v1.5 (Vicuna-v1.5-7B)64.60%59.22% 55.76%47.98%67.30%58.90%73.76%60.07%LLaVA-v1.5 (Vicuna-v1.5-13B)64.96%64.86% 54.12%53.55%66.59%58.90%71.48%61.40%InternLM-XComposer-VL (InternLM)68.43%62.04% 61.93%56.81%70.41%57.53%77.19%64.35%IDEFICS-Instruct (LLaMA-7B)60.04%46.42% 46.71%40.38%59.90%47.26%64.77%51.51%Qwen-VL (QwenLM)65.33%60.74% 58.44%54.13%66.35%58.22%73.00%61.67%Shikra(Vicuna-7B)69.09%47.93% 46.71%47.31%60.86%53.08%64.77%55.32%Otter-v1 (MPT-7B)57.66%39.70% 42.59%42.12%48.93%47.60%54.17%47.22%InstructBLIP (Flan-T5-XL)69.53%59.00% 56.17%57.31%65.63%56.51%71.21%61.94%InstructBLIP (Vicuna-7B)70.99%51.41% 43.00%45.00%63.01%57.19%64.39%55.85%VisualGLM-6B (GLM-6B)61.31%53.58% 44.03%48.56%54.89%55.48%57.79%53.31%mPLUG-Owl (LLaMA-7B)72.45%54.88% 47.53%49.62%63.01%62.67%66.67%58.93%LLaMA-Adapter-V266.61%54.66% 51.65%56.15%61.81%59.25%54.55%58.06%LLaVA-v1 (Vicuna-13B)57.12%54.88% 51.85%45.58%58.00%57.19%64.77%54.72%MiniGPT-4 (Vicuna-13B)60.77%50.33% 43.00%45.58%52.51%53.42%60.98%51.77%GPT-4V (Close-Source Model)77.92%79.18% 62.68%70.58%73.03%74.66%77.95%73.36%Junior-level Human82.48%79.39% 60.29%75.62%72.08%76.37%73.00%74.31%Senior-level Human84.31%88.94% 72.02%79.65%79.47%83.90%87.07%81.74%3 RESULTS ON Q-BENCHIn Q-Bench, we evaluate 15 variants on 13 up-to-date popular and competitive open-source MLLMs, together with GPT-4V, under zero-shot settings. More results and analyses are appended in Sec. A.3.
Table 3 :3Results on the low-level Description ability of MLLMs. P i denotes frequency for score i.Dimensions Model (variant)P0Completeness P1 P2score↑P0Precision P1 P2score↑P0Relevance P1 P2score↑Sum.↑LLaVA-v1.5 (Vicuna-v1.5-7B)27.48% 54.74% 17.78% 0.90 30.51% 26.04% 43.45% 1.13 10.85% 60.34% 28.81% 1.183.21LLaVA-v1.5 (Vicuna-v1.5-13B)27.68% 53.78% 18.55% 0.91 25.45% 21.47% 53.08% 1.286.31% 58.75% 34.94% 1.293.47InternLM-XComposer-VL (InternLM) 19.94% 51.82% 28.24% 1.08 22.59% 28.99% 48.42% 1.261.05% 10.62% 88.32% 1.874.21IDEFICS-Instruct (LLaMA-7B)28.91% 59.16% 11.93% 0.83 34.68% 27.86% 37.46% 1.033.90% 59.66% 36.44% 1.333.18Qwen-VL (QwenLM)26.34% 49.13% 24.53% 0.98 50.62% 23.44% 25.94% 0.750.73% 35.56% 63.72% 1.633.36Shikra (Vicuna-7B)21.14% 68.33% 10.52% 0.89 30.33% 28.30% 41.37% 1.111.14% 64.36% 34.50% 1.333.34Otter-v1 (MPT-7B)22.38% 59.36% 18.25% 0.96 40.68% 35.99% 23.33% 0.831.95% 13.20% 84.85% 1.833.61Kosmos-28.76% 70.91% 20.33% 1.12 29.45% 34.75% 35.81% 1.060.16% 14.77% 85.06% 1.854.03InstructBLIP (Flan-T5-XL)23.16% 66.44% 10.40% 0.87 34.85% 26.03% 39.12% 1.04 14.71% 59.87% 25.42% 1.113.02InstructBLIP (Vicuna-7B)29.73% 61.47% 8.80%0.79 27.84% 23.52% 48.65% 1.21 27.40% 61.29% 11.31% 0.842.84VisualGLM-6B (GLM-6B)30.75% 56.64% 12.61% 0.82 38.64% 26.18% 35.18% 0.976.14% 67.15% 26.71% 1.212.99mPLUG-Owl (LLaMA-7B)28.28% 37.69% 34.03% 1.06 26.75% 18.18% 55.07% 1.283.03% 33.82% 63.15% 1.603.94LLaMA-Adapter-V230.44% 53.99% 15.57% 0.85 29.41% 25.79% 44.80% 1.151.50% 52.75% 45.75% 1.443.45LLaVA-v1 (Vicuna-13B)34.10% 40.52% 25.39% 0.91 30.02% 15.15% 54.83% 1.251.06% 38.03% 60.91% 1.603.76MiniGPT-4 (Vicuna-13B)34.01% 32.15% 33.85% 1.00 29.20% 15.27% 55.53% 1.266.88% 45.65% 47.48% 1.413.67
Table 4 :4Main evaluation results on the zero-shot Assessment ability of MLLMs, in comparison with NIQE and CLIP-ViT-Large-14, the visual backbone of most MLLMs. Metrics are SRCC/PLCC.Dataset Type Model / DatasetKONiQ-10kIn-the-wild SPAQ LIVE-FBLIVE-itwGenerated CGIQA-6K AGIQA-3K KADID-10K ArtificialAverageNIQE (Mittal et al., 2013)0.316/0.377 0.693/0.669 0.211/0.288 0.480/0.451 0.075/0.056 0.562/0.517 0.374/0.428 0.387/0.398CLIP-ViT-Large-140.468/0.505 0.385/0.389 0.218/0.237 0.307/0.308 0.285/0.290 0.436/0.458 0.376/0.388 0.354/0.368LLaVA-v1.5 (Vicuna-v1.5-7B)0.463/0.459 0.443/0.467 0.305/0.321 0.344/0.358 0.321/0.333 0.672/0.738 0.417/0.440 0.424/0.445LLaVA-v1.5 (Vicuna-v1.5-13B)0.448/0.460 0.563/0.584 0.310/0.339 0.445/0.481 0.285/0.297 0.664/0.754 0.390/0.400 0.444/0.474InternLM-XComposer-VL (InternLM) 0.564/0.615 0.730/0.750 0.360/0.416 0.612/0.676 0.243/0.265 0.732/0.775 0.546/0.572 0.541/0.581IDEFICS-Instruct (LLaMA-7B)0.375/0.400 0.474/0.484 0.235/0.240 0.409/0.428 0.244/0.227 0.562/0.622 0.370/0.373 0.381/0.396Qwen-VL (QwenLM)0.470/0.546 0.676/0.669 0.298/0.338 0.504/0.532 0.273/0.284 0.617/0.686 0.486/0.486 0.475/0.506Shikra (Vicuna-7B)0.314/0.307 0.320/0.337 0.237/0.241 0.322/0.336 0.198/0.201 0.640/0.661 0.324/0.332 0.336/0.345Otter-v1 (MPT-7B)0.406/0.406 0.436/0.441 0.143/0.142 -0.008/0.018 0.254/0.264 0.475/0.481 0.557/0.577 0.323/0.333Kosmos-20.255/0.281 0.644/0.641 0.196/0.195 0.358/0.368 0.210/0.225 0.489/0.491 0.359/0.365 0.359/0.367InstructBLIP (Flan-T5-XL)0.334/0.362 0.582/0.599 0.248/0.267 0.113/0.113 0.167/0.188 0.378/0.400 0.211/0.179 0.290/0.301InstructBLIP (Vicuna-7B)0.359/0.437 0.683/0.689 0.200/0.283 0.253/0.367 0.263/0.304 0.629/0.663 0.337/0.382 0.389/0.446VisualGLM-6B (GLM-6B)0.247/0.234 0.498/0.507 0.146/0.154 0.110/0.116 0.209/0.183 0.342/0.349 0.127/0.131 0.240/0.239mPLUG-Owl (LLaMA-7B)0.409/0.427 0.634/0.644 0.241/0.271 0.437/0.487 0.148/0.180 0.687/0.711 0.466/0.486 0.432/0.458LLaMA-Adapter-V20.354/0.363 0.464/0.506 0.275/0.329 0.298/0.360 0.257/0.271 0.604/0.666 0.412/0.425 0.381/0.417LLaVA-v1 (Vicuna-13B)0.462/0.457 0.442/0.462 0.264/0.280 0.404/0.417 0.208/0.237 0.626/0.684 0.349/0.372 0.394/0.416MiniGPT-4 (Vicuna-13B)0.239/0.257 0.238/0.253 0.170/0.183 0.339/0.340 0.252/0.246 0.572/0.591 0.239/0.233 0.293/0.300
Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition, 2023a.Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonableeffectiveness of deep features as a perceptual metric. In CVPR, pp. 586-595, 2018.Zicheng Zhang, Wei Sun, Tao Wang, Wei Lu, Quan Zhou, Qiyuan Wang, Xiongkuo Min, GuangtaoZhai, et al. Subjective and objective quality assessment for in-the-wild computer graphics images.ACM TOMM, 2023b.Zicheng Zhang, Wei Sun, Yingjie Zhou, Haoning Wu, Chunyi Li, Xiongkuo Min, and XiaohongLiu. Advancing zero-shot digital human quality assessment through text-prompted evaluation,2023c.Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang,Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica.Judging llm-as-a-judge with mt-bench and chatbot arena, 2023.Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.
.1.2 MORE DETAILS ON LLVISIONQA The Enumeration on Distortions and Other Low-level Attributes:Distortions: Blurs [lens blur (out-of-focus), motion blur, zoom blur, gaussian blur, glass blur], Noises [gaus-sian noise, speckle noise, pepper noise], Artifacts [compression artifact, transmission error], Exposure Issues[under-exposure, over-exposure], Miscellaneous Artificial Distortions [pixelate, color-diffusion, jitter, etc]Other low-level attributes: Color [color style, color vividity], Lighting [bright, dim], Composition [Sym-metrical, Rule-of-Thirds], Visual Styles [animation, realism, computer-generated, AI-generated], PhotographicMethods [background bokeh (shallow DOF), high contrast, motion blur (on fast-moving objects), etc]Relationship between In-context Questions and Global Questions:Distortions: Is this image blurred? → In-context Distortions: Is the tree in the image blurred? Distortions: How is the clarity of the image? → In-context Distortions: How is the clarity of the man's face? Other low-level: Is this image colorful? → In-context Other: Is the house colorful? Other low-level: How is brightness of the image? → In-context Other: Which is the darkest object?
Table 5 :5Perplexity-based close-set evaluation compared with normal evaluation on LLVisionQA; after eliminating the prompt failures, the results of Kosmos-2 significantly improved.Sub-categoriesQuestion TypesQuadrants of Low-level ConcernsModel (variant)Yes-or-No↑ What↑How↑ Distortion↑ Other↑In-context In-context Distortion↑ Other↑Overall↑ #↓random guess50.00%28.18% 33.30%37.54%38.49%38.70%36.50%37.87% -⋆⋆ Kosmos-2 (normal)58.20%29.13% 34.22%38.10%44.30%40.93%44.20%41.47% ✗⋆⋆ Kosmos-2 (close-set)61.48%37.13% 40.76%40.04%50.88%45.30%58.15% 47.26% ✓
Algorithm 1 Pytorch-style Pseudo Code for Softmax-based Strategy for IQA with MLLMsfrom PIL import Imagefrom my_mllm_model import Model, Tokenizer, embed_image_and_textmodel, tokenizer = Model(), Tokenizer()prompt = "##User: Rate the quality of the image.\n" \"##Assistant: The quality of the image is"good_idx, poor_idx = tokenizer(["good","poor"]).tolist()A.2.3 EVALUATION DETAILS FOR ASSESSMENT ABILITYExample Pseudo Code for MLLMs on IQA:
Table 7 :7Results on the dev subset for the low-level Perception ability of MLLMs. MLLMs with top-3 performance in each sub-category and the overall LLVisionQA is emphasized with boldface.Sub-categoriesQuestion TypesQuadrants of Low-level ConcernsModel (variant)Yes-or-No↑ What↑How↑ Distortion↑ Other↑In-context In-context Distortion↑ Other↑Overall↑random guess50.00%27.86% 33.31%37.89%38.48%38.28%35.82%37.80%LLaVA-v1.5 (Vicuna-v1.5-7B)66.36%58.19% 50.51%49.42%65.74%54.61%70.61%58.66%LLaVA-v1.5 (Vicuna-v1.5-13B)65.27%64.38% 56.59%56.03%67.13%61.18%67.35%62.14%InternLM-XComposer-VL (InternLM)69.45%65.27% 60.85%61.67%70.14%56.91%75.10%65.35%IDEFICS-Instruct (LLaMA-7B)56.18%44.69% 44.02%42.80%54.17%44.74%56.33%48.70%Qwen-VL (QwenLM)63.09%58.19% 56.39%50.58%62.73%57.89%73.88%59.40%Shikra (Vicuna-7B)65.64%47.35% 49.09%48.83%59.49%50.00%64.08%54.65%Otter-v1 (MPT-7B)57.09%40.71% 39.55%42.22%49.31%44.08%52.65%46.35%InstructBLIP (Flan-T5-XL)67.64%59.96% 55.98%56.23%65.51%58.22%69.39%61.47%InstructBLIP (Vicuna-7B)71.64%52.65% 43.81%48.64%62.50%55.59%64.90%56.72%VisualGLM-6B (GLM-6B)60.18%54.20% 46.25%51.75%54.40%53.62%57.14%53.78%mPLUG-Owl (LLaMA-7B)66.0%54.87% 44.02%51.36%55.09%54.28%65.71%55.38%LLaMA-Adapter-V266.18%59.29% 52.13%57.39%56.25%63.16%64.90%59.46%LLaVA-v1 (Vicuna-13B)54.00%53.10% 55.38%48.64%54.63%55.59%63.27%54.18%MiniGPT-4 (Vicuna-13B)55.82%50.22% 40.37%42.02%48.38%51.97%61.22%49.03%