Quantitative Measures of Task-Oriented Creativity in Popular Generative Vision Models Aditi Ramaswamy, Hana Chockler, Melane Navaratnarajah Abstract Creativity of generative AI models has been a subject of scientific debate in the last years, without a conclusive answer. In this paper, we study creativity from a practical perspective and introduce quantitative measures that help the user to choose a suitable AI model for a given task. We evaluated our measures on a number of popular img2img generation models and the results suggest that our measures conform to the intuition. Figure 1: Example Flux chains, generated from a single seed image apple_pie_139.png (the leftmost image), with different input temperatures. Introduction Generative AI has become the center of a series of heated debates, one of the biggest ones centered around the concept of human creativity. Some researchers argue that creativity is derived from an innately human “sociocultural context”, thus excluding AI models [ Oppenlaender, 2022 , Kaufman et al., 2019 , Wingström et al., 2022 ] . However, other definitions focus on purely behavioral requirements, such as “domain-relevant skills”, “creativity-relevant processes”, and “extrinsic motivation”, all of which can be exhibited by generative AI models [ Amabile, 1983 ] . Related to this, Margaret Boden proposes an AI-relevant cognitive dimension of creativity separate from emotion, focusing on behaviors such as novelty and association of ideas [ Boden, 1998 ] . Boden also discusses “exploratory” creativity, the “generation of novel ideas” within a space, such as an AI model responding to a given task by creating new content within the parameters of the task. These process-oriented definitions suggest that some generative models can exhibit a form of task-oriented creative behavior in response to prompts, thus forming the motivation for this paper. Quantifying this task-oriented creative behavior could provide valuable insights for both the users who interact with these models and the model developers. If a user is presented with clear information about the ways in which a given generative model may exhibit creative behavior in response to their queries, they can make an informed choice about whether they wish to use it. The same information could also pinpoint specific areas of change which those who develop generative models can use to tweak their training data and model architecture to better suit their intended purpose. The focus of this paper is answering the following research question: “Can we define mathematical measures quantifying useful aspects of creative behavior of popular img2img generation models?” We restrict our attention to the aspects of creativity that can be quantified and used to choose a generative model most suitable for a specific purpose. For example, a generative model suitable for producing images for a physics textbook would probably have very different creativity-related characteristics than one suitable for generating illustrations for a fantasy novel. Using [ Hauhio, 2024 ] ’s idea of uncontrollability based on artifact spaces, as well as the novel , useful , and typical criteria laid out by [ Peeperkorn et al., 2024 ] , we propose to characterize outputs of models according to three criteria, whose formal quantitative definitions we present in the definitions section: (1) satisfaction of prompt requirements, (2) cohesion between output artifacts, and (3) diversity of output artifacts. We introduce an iterative process involving construction of a chain for measuring these with a high level of granularity and analyze the results through both statistical and visual means. The chain construction is inspired by the telephone game and includes a repeated generation of new outputs based on the previous ones. This process is repeated for a predefined number of steps, with the resulting chain analyzed for the concepts introduced above. To the best of our knowledge, while previous papers have described aspects of creative behavior, this is the first attempt at mathematically measuring useful aspects of creativity in popular img2img generation models. The chain construction is illustrated in Figure 1 , which also shows how outputs with low creativity scores vs high ones look like. Specifically, the creativity score of the top chain is 0 , as it does not satisfy the prompt requirements at all; the creativity score of the bottom chain is 0.46 , as, though faithful to the prompt, it does not introduce any new elements; the creativity score of the middle chain is 0.7 , as it is both faithful to the prompt and introduces new elements that are cohesive with the ones in the original image. We show experimental results for evaluating a number of popular image generators with a variety of input temperatures. Input temperature, which influences the distribution from which output features are selected, has been proposed as a metric for measuring creative behavior in generative models, but this has engendered debate, as a high input temperature can easily lead to an unwanted and excessive level of hallucination. Our measures capture hallucinations, as well as simply copying the seed image; both are ranked low on creativity. Due to the lack of space, we only present the summary results and a small number of illustrative images. The code, full experimental results, and full sets of images and chains can be downloaded from https://figshare.com/s/d88d4966b606163d02fc . Background Here, we provide a detailed breakdown of the concepts and definitions on which we base our experimental setup. Computational Creativity and Temperature A high value of temperature, a parameter used to determine “the proportion by which the output probability distribution should be flattened” according to [ Peeperkorn et al., 2024 ] , has been suggested to be a strong influence on creativity, as it increases the chance of more unique, low-probability elements being selected for representation in the model’s output [ Chen and Ding, 2023 ] . This has been described as “uncontrollability” by [ Hauhio, 2024 ] , and “hallucination” in other contexts pertaining to creative behavior in models, such as [ Lee, 2023 , Wu et al., 2023 ] . These studies posit that temperature is a strong influence on creative behavior. However, other studies argue that using high temperature alone to measure creativity does not allow for any differentiation between (1) high-quality images that use given text or image input prompts as inspiration, and (2) random images produced with little to no influence from the input prompts [ Peeperkorn et al., 2024 ] . While a high temperature makes it likelier for a model to add low-probability elements to its output, there is no guarantee that those elements will be seen as high-quality by the user. Given this disagreement, as well as the fact that prominent papers like [ Peeperkorn et al., 2024 ] and [ Chen and Ding, 2023 ] focus on large language models, we have chosen to probe further into the effects of temperature on img2img models as a novel contribution to the literary body. We also note a gap in the literature where img2img model behavior is concerned. Prior studies on creativity in generative AI models have tended to deal with large language models or text2img generation [ Hauhio, 2024 , Gao et al., 2024 ] . The creative behavior of img2img generation, however, is largely unstudied. Defining Facets of Creative Behavior in Generative Vision Models We use the term “artifact”, derived from [ Hauhio, 2024 ] ’s usage as well as the common linguistic meaning, to refer to a core feature of a text or image, such as “apple pie” in the textual prompt “a slice of apple pie”, or a segment of an image labeled “pie” by an object detection tool. Each prompt and output can therefore be understood as a set of artifacts. [ Hauhio, 2024 ] proposes multiple artifact spaces, the relationship between which can be used to classify models’ behavior as creative. C , referred to as the “conceptual set” by [ Hauhio, 2024 ] , is the set of artifacts specified in the user’s input prompt to a generative model, and V is the set of artifacts a user would find valuable given that this user has a particular goal \mathbb{R}_{A} in using the generative model. Uncontrollability, in this paper, refers to the idea that a given generative model, when fed C , will use a strategy that produces a valuable artifact set V that contains extraneous artifacts not specified in C as well as those in \mathbb{R}_{A} . This expansion of valuable artifacts from strictly the ones referred to in the prompt can, according to this paper, be understood as a basis for creative behavior. Definitions Using [ Hauhio, 2024 ] ’s idea of uncontrollability based on artifact spaces, as well as the novel , useful , and typical criteria laid out by [ Peeperkorn et al., 2024 ] , we propose to characterize outputs of models according to three criteria, whose formal quantitative definitions we present below: input prompt requirement satisfaction , cohesion , and diversity . Satisfaction of Prompt Requirements , the idea that an output contains (all) the elements specified in the user’s input, builds off [ Peeperkorn et al., 2024 ] ’s statement that an output must be typical of its class, as well as [ Hauhio, 2024 ] ’s “concept set” C . We build cohesion from [ Peeperkorn et al., 2024 ] ’s assertion that an output must also be useful to be considered creative, by working off the idea that an output with elements that are generally more closely related to each other has more use cases than an output with disjoint elements. cohesion is also derived from [ Boden, 1998 ] ’s “association of ideas”: how well does a generative model build strong associations between artifacts within its output in response to a given prompt? For diversity , we build off [ Hauhio, 2024 ] ’s idea of valuable artifacts that were not specified by the input prompt, but are added spontaneously by the generative model, as well as [ Peeperkorn et al., 2024 ] ’s and [ Boden, 1998 ] ’s ideas that an output must contain novel elements to be considered creative. Specifically, Boden mentions “transformational” creativity as the idea that new elements can be introduced into a space through transformation, a behavior directly linked to the ability of img2img models to generate novel output elements using the framework of an image input. Our measure differs from prior diversity metrics like Fréchet Inception Distance (FID) [ Heusel et al., 2017 ] in that it solely focuses on individual detectable artifacts within an image, rather than an image’s overall structure. Contrary to FID and similar measures, our goal is to optimize the measure for images that differ notably from the original image while still containing elements (artifacts) from the original. In the definitions below we use the following notations: M is the generative model, A_{O} is the artifact set of a given output image O_{M} (generated by M ), and A_{I} is the artifact set of the input prompt I , which can be an image or a text. The cosine similarity measure, which is used throughout these definitions, is a way of measuring the similarity between the vector projections of two artifacts in semantic embedding space—in other words, a way to mathematically compare the meanings behind textual representations of artifacts. It will be henceforth denoted as the function \theta or as the dot product \cdot , depending on context. For a set A of artifacts, the embedding of each artifact a\in A in semantic space is calculated using the embedding function emb , denoted as emb(a)=\hat{a} [ Matsuki et al., 2019 ] . Definition 1 ( Satisfaction of Prompt Requirements ) An image is said to be satisfying requirements if A_{I}\subseteq A_{O} . In other words, all artifacts in I are present in O_{M} . Definition 2 ( Cohesion ) Cohesiveness C(O_{M}) of an output image O_{M} is defined wrt the input I . Given the set A_{I} of the artifacts in I and the set A_{O} of the artifacts in O_{M} , let P be the set of all pairs (a,b) , where a\in A_{O}\setminus A_{I} and b\in A_{I} . Furthermore, let M_{N}=\max_{(c,d)\in A_{O}\setminus A_{I}}((\hat{c})\cdot(\hat{d})) . Then, cohesiveness of O_{M} is defined as C(O_{M})=\frac{{\sum}_{a:(a,b)\in P}{\max_{b:(a,b)\in P}((\hat{a})\cdot(\hat{b}))}+M_{N}}{|P|+1}. In other words, the cohesiveness of O_{M} is the mean value of the maximal similarities between new artifacts and seed artifacts and the maximal similarity between new artifacts. We note that the scalar product (\hat{a})\cdot(\hat{b}) of the representations of a and b in the semantic embedded space increases with increasing the similarity. As 0\leq(\hat{a})\cdot(\hat{b})\leq 1 , also 0\leq C(O_{M})\leq 1 . Definition 3 ( Diversity ) Diversity D(O_{M}) of an output image O_{M} is defined as the proportion of new artifacts in the image, if O_{M} contains all artifacts of I . In other words, D(O_{M}) is \frac{|A_{O}\setminus A_{I}|}{|A_{O}|} if A_{I}\subseteq A_{O} and is 0 otherwise. While diversity and cohesiveness are not independent concepts, an output image O_{M} can have high diversity and low cohesiveness if it contains many new artifacts with low similarity to the original ones, and it can also have high cohesiveness and low diversity if there are very few new artifacts and they are very similar to the original ones. Methodology In this section, we outline our methodology for quantitative evaluation of creativity of image generation models. Chain Construction As [ Hauhio, 2024 ] points out, image generation models are often used iteratively, with users modifying their input prompts based on the artifacts present in the generated output. In this paper, we simulate this iterative approach, which allows us to measure the afore-discussed aspects of creative behavior in a way that is more pertinent to real-world use-cases. To our knowledge, an iterative approach like ours has not been utilized for experimentation in prior computational creativity studies. We base our approach on the well-known game telephone (or Chinese whispers ). As in this game, each chain in our experiment begins with a “ground truth”, or seed image. This image is then processed into an input for an image generation model, and the output of this is and fed back into the image generation model; this process repeats for x iterations, where x is a parameter chosen by the user. One step of the process is illustrated in the flowchart in Figure 2 . Figure 2 : The flowchart depicting a single step in the chain process The specific mechanisms of the input processing mentioned above are determined by the type of chain. Our main focus is img2img generative models, hence our experimental setup evaluates how strongly an image input influences aspects of creativity, compared to textual input. We implement three types of chains, as shown in Table 1 . Generative image models can be used with text input prompts, image input prompts, or a combination of both, which is reflected in our experimental setup. For chains with text input, text2img , we auto-generate a caption for the prior image in the chain and feed that into the image generation model in the next step. For chains with image input ( img2img ), we pass the prior image in the chain directly into the image generation model. Table 1: Chain Types in Experimental Process Chain Type Image Seed Text Seed img_only Previous image None cap_only None Caption of Previous Image img_cap Previous Image Caption of Previous Image Based on the earlier discussion of the effects of temperature and uncontrollability on creative behavior, we consider the temperature as a parameter of the image inputs to the img2img pipelines. Each of the img2img pipeline types is executed with multiple levels of temperature for the image input, to determine whether different temperatures have statistically significant effects on the requirement satisfaction , cohesion , diversity , and overall creativity ranking of the outputs of a given generation model. Measures for Creativity Using the conceptual definitions of requirement satisfaction , cohesion , and diversity provided earlier, we can define mathematical measures for each of these aspects in the context of the chains, so that we can quantify the creative behaviors exhibited by the generator used within a given chain. The satisfaction factor measure provides the backbone for the other two, as we want to focus on images that exhibit a higher-than-random level of requirement satisfaction for the input prompt. We expand and quantify Definition 1 to produce this measure for a chain of images rather than a single comparison between two images. For a given chain, we quantify satisfaction factor as the normalized longest unbroken sequence of the chain wherein, at each step, the seed artifact set, or a set of artifacts that approximate the seed artifact set within a cosine similarity threshold t , is a subset of the artifact set of the image generated for that step. We then multiply this proportion of the chain by the average \theta score for the set of pairs consisting of one artifact from the seed set and its closest match in the generated image’s artifact set, to account for the fact that the generated image may not be perfectly faithful. For example if the seed artifact set is [“apple pie”, “horse”], and the generated image for which satisfaction factor is calculated has the artifact set [“apple cake”, “horse”], the proportion of the chain that satisfies the input prompt as described in the previous paragraph would be multiplied by \frac{\theta(\mbox{``apple pie''},\mbox{``apple cake''})+\theta(\mbox{``horse''},\mbox{``horse''})}{2} to differentiate it between a generated image that contains the exact seed artifact set. The later in the chain the seed artifacts or their approximations still appear, the higher the satisfaction factor value is going to be, although it also depends on how closely the artifacts in the generated image approximate the chain image. Let l be the maximal length of the chain, O_{n} the image produced at step n of the chain, where 1\leq n\leq l , A_{S} the textual set of artifacts present in the chain’s seed image, A_{n} the textual set of artifacts present in O_{n} , and t the threshold of approximation between seed and generated image artifacts: Let K , for 0\leq K\leq l , be the number of consecutive steps from the beginning of a chain for which the following holds: for all k\leq K and a_{s}\in A_{S} , there exists a_{k}\in O_{k} , where (\hat{a_{k}})\cdot(\hat{a_{s}})\ \geq\ t. (1) A chain’s Requirements Satisfaction Factor is then defined as RS=\frac{K}{l}\times mean((\hat{a_{k}})\cdot(\hat{a_{s}})). (2) In other words, we extract artifacts from the generated image in each step of the chain. We then represent each artifact in the current image as a vector in semantic embedding space, and calculate the cosine similarity between these embeddings and the embeddings of the artifacts in the seed image. K is the index of the last step, where where all artifacts fall within the similarity threshold t of their counterparts in the seed image. The chain’s Requirements Satisfaction Factor is defined as the product of K normalized by the overall chain length l with the mean value of the similarities between artifacts of the generated images and their closest counterparts in the seed image. Consider the image O_{K} (the K -st image in the chain). As it is the last image in the generated chain that satisfies Equation 1 , we use it to analyze cohesion and diversity, rather than calculating those for the whole chain. Let A_{K} be the artifact set of O_{K} , and A_{S} once again represent the artifact set of the seed image. A chain’s Cohesion Factor B_{R} (using B for bond, to differentiate from CR , the creativity score as defined later) can be calculated using Definition 2 . We set P to be the set of all pairs (a,b) , where a\in A_{K}\setminus A_{S} and b\in A_{K}\cap A_{S} : m_{1}={\sum}_{a:(a,b)\in P}({\max_{b:(a,b)\in P}((\hat{a})\cdot(\hat{b}))}) (3) Let M_{N} be as in Definition 2 , with A_{O}=A_{K} and A_{I}=A_{S} . Then, B_{R} is defined as: B_{R}=\begin{cases}\frac{M_{N}+\sum_{m\in m_{1}}m}{|m_{1}|+1}\text{ if }|A_{K}|\geq 2\\ 0\text{ if }|A_{K}|<2\end{cases} (4) Breaking this equation down step-by-step, the cohesion factor is calculated by first creating a set of all new artifacts in O_{K} , that is, A_{K}\setminus A_{S} . Then we calculate the similarity of these artifacts to the artifacts brought from the seed image, in Equation 3 . M_{N} stands for the similarity between the new artifacts themselves. The B_{R} score is the mean value of these, where the similarity between the new artifacts has a lower weight than the similarity between the new artifacts and the seed image ones. We account for instances where A_{K} has less than 2 artifacts by automatically setting B_{R} to be 0 , as cohesion requires at least two objects. The diversity score captures the fraction of A_{K} consisting of labels that are not found in A_{S} , indicating how many extraneous objects were added to O_{K} , and is based on Definition 3 : D_{R}=\begin{cases}\frac{|A_{K}\setminus A_{S}|}{|A_{K}|}\text{ if }A_{S}\subseteq A_{K}\\ \text{ else }0.\end{cases} (5) Each of these quantify individual behaviors for a given chain. In order to measure overall creativity, we combine these measures to an overall creativity ranking score, CR , where 0\leq CR\leq 1 . For a given chain: CR=RS\times\frac{B_{R}+D_{R}}{2}. (6) The chain under evaluation is a parameter to all scores we introduce in this section; it is omitted for brevity. Intuitively, the creativity score can be higher than 0 only if the generated images satisfy the prompt requirements; this prevents us from awarding a high creativity score to hallucinating models. On the other hand, a mechanistic copying of the seed image is not creative either. Indeed, while RS=1 for a mechanistic copy, both B_{R} and D_{R} are 0 , hence resulting in the creativity score 0 as well. For models that satisfy the prompt requirements, while adding new elements to the generated images, both the diversity between the new elements and their cohesion with the seed image elements, as well as between themselves, affect the creativity score. If we are given a set X of generative models, we can use the means of RS , B_{R} , D_{R} , and CR across the total number of chains for all models in X to rank them in terms of specific aspects of creative behavior as per our earlier discussion. Of course, since creativity is inherently subjective and may vary across users and contexts, this ranking may not be universally applicable. Extensions or modifications to these definitions can be explored to better fit specific domains or to align with different interpretations of creativity. Evaluation Experimental Setup For our experiments we selected sets of simple seed images focusing on a single, easy-to-detect subject, as these would be the easiest to measure requirement satisfaction , cohesion , and diversity , since A_{S} would be 1 . Since many object detectors are trained on datasets that prominently feature food images, we pulled our 999 seed images from the publicly-available Food-101 dataset [ Bossard et al., 2014 ] . We selected three food categories at random: “apple pie”, “donuts”, and “pizza”, and randomly selected 333 images from each. Input prompts I were constructed from each of these seed images. We evaluated our measures on three popular open-source image generation models from Hugging Face: Stable-Diffusion 3 3 [ Esser et al., 2024 ] , Kandinsky 2.2 [ Razzhigaev et al., 2023 ] , and Flux [ BlackForestLabs, 2023 ] . The nature of the input prompt I depends on the type of chain as shown in Table 1 . For the former two chain types, we pass the seed image into the image generation model using the appropriate img2img pipeline. For the latter two chain types, we generate a text caption for the seed image using KOSMOS model from Hugging Face [ Peng et al., 2023 ] . The rest of the chain is constructed using the same principle: for img2img chains, the next entry is produced by passing in an I consisting of the previous image, while for text2img chains I contains the KOSMOS caption for the previous image. For chains with a combination of both caption and image input, both the previous image and its KOSMOS caption are passed in. In our experiments, we set the maximal length of the chain to 10 to avoid computational explosion, while still producing meaningful results. For temperature, we used the strength parameter as described in the Hugging Face img2img pipeline documentation [ Liu, 2023 ] . The documentation claims that strength influences the creativity of the pipeline’s output, a claim we aim to test with our experiments. Functionally, strength determines the amount of noise added to the seed image during the generative process, meaning that a lower value of strength adds less noise and therefore produces an output that more closely resembles the seed image, while a higher value of strength adds more noise to the input image and thus produces an output that is visually farther from the seed image. We executed both the img_only and img_cap pipelines for all three img2img models with three different strength levels: 0.3 , 0.6 , and 0.9 , thus encompassing an equally spaced range from low to high. Since strength controls the number of generation steps in img2img pipelines and is not present in text2img pipelines, for the sake of accurate comparison we executed each of the cap_only pipelines with three generation step counts: 15 , corresponding to the number of generation steps used by the strength of 0.3 pipelines; 30 , for the strength of 0.6 ; and 45 , corresponding to the strength 0.9 . In the rest of this paper, we use “temperature” and “strength” interchangeably. To compute the measures of requirement satisfaction , cohesion , diversity , and creativity ranking defined earlier, we used two object detectors, GroundingDINO and DETR , to extract artifacts in textual form from images in the chain, where the initial image is the seed image, and for A_{S} we simply use a text representation of the subject of the seed image: apple pie, donut, or pizza. For word embedding and semantic similarity calculations between the artifacts, we used the spaCy library. We selected 0.65 as the value of t for satisfaction factor , after running preliminary tests to account for incorrect object detection results. We used paired T-tests to determine any statistically significant differences between results for different strengths, chain types, and image generation models, on the same dataset of 999 images. The reasoning behind this choice is that finding statistically significant differences between different chain configurations would yield information about the impact of image vs text inputs on our creativity measures, as well as allow us to compare the relative performance of different generative models. We used SciPy ’s ttest_rel function to perform these tests. Presentation of Results (a) At strength 0.3. (b) At strength 0.6. (c) At strength 0.9. Figure 3: Example img_cap Kandinsky 2.2 chains for various strengths, generated from a single seed image apple_pie_184.png . From left to right: the seed image, followed by steps 1- n of the generated chain where n is the last image that satisfies the prompt requirements. The values of RS , B_{R} , and D_{R} were calculated for the chains generated for each of the 27 possible combinations of image generation model, strength, and chain type. GroundingDINO and DETR were used on each generated image to extract artifact labels from them in text form. Figure 3 is an example of how our creativity measures manifest visually in a set of img_cap Kandinsky 2.2 chains at different strengths. The chain pictured in Figure 3(a) was generated with strength 0.3 . The satisfaction factor value of this chain is 0.68 , meaning that while the chain remains unbroken due to the average cosine similarity staying above the threshold of 0.65 , the closest artifact found in the last image, “cake”, only has a similarity of 0.68 to the seed artifact, “apple pie”. Accordingly, the chain values of cohesion factor , diversity factor , and creativity ranking as calculated from the last image in this chain are B_{R}=0.3;D_{R}=0.67;CR=0.48 , indicating a medium creativity score, due to relatively high diversity (“syrup”, “dining table”, and “cup” were added in) and low to medium cohesion between the artifacts. The satisfaction factor value of the chain shown in Figure 3(b) , generated with strength 0.6 , is 0.9 , indicating that the subject remains “apple pie” until step 9 out of 10 . Using the 9 th image (the rightmost pictured), the chain’s scores are B_{R}=0.45;D_{R}=0.75;CR=0.54 , indicating a higher creativity than the previous chain, due to an increased number of added labels, particularly a person’s hand pouring syrup. The satisfaction factor of the chain shown in Figure 3(c) is also 0.9 , with the closest artifact match being “apple”, but because no other labels were identified by either GroundingDINO or DETR in the image, D_{R}=0.0 . A single subject yields a cohesion factor value of 0 , hence B_{R}=0.0 as well. Therefore, CR=0.0 , indicating a chain that adequately satisfies the prompt requirements but, unlike the first two, does not contain any novel elements. Given the standard threshold of p<0.05 to indicate a statistically significant difference between sets of results, almost all the models showed statistically significant differences. The tables and figures below show averages and comparisons of these measures across img_cap chains for each combination of image generation models with different strength. Values of p that are extremely close to 0 are represented as 0+\epsilon . Note that the mean values are rounded to 2 decimal points, meaning that some means that look very close may in fact belong to statistically different datasets. Table 2: Paired T-Test results for CR in img_cap chains with strength 0.3 model 1 model 2 mean 1 mean 2 P-value Flux Kandinsky 2.2 0.18 0.19 0.5007 Flux Stable-Diffusion 3 0.18 0.1 0+\epsilon Kandinsky 2.2 Stable-Diffusion 3 0.19 0.1 0+\epsilon Table 3: Paired T-Test results for CR in img_cap chains with strength 0.6 model 1 model 2 mean 1 mean 2 P-value Flux Kandinsky 2.2 0.2 0.17 0.0220 Flux Stable-Diffusion 3 0.2 0.1 0+\epsilon Kandinsky 2.2 Stable-Diffusion 3 0.17 0.1 0+\epsilon Table 4: Paired T-Test results for CR in img_cap chains with strength 0.9 model 1 model 2 mean 1 mean 2 P-value Flux Kandinsky 2.2 0.22 0.19 0.0011 Flux Stable-Diffusion 3 0.22 0.13 0+\epsilon Kandinsky 2.2 Stable-Diffusion 3 0.19 0.13 0+\epsilon In summary, as shown by Figures 3 , 4 , 5 , and Tables 2 , 3 , 4 , 5 , 6 , and 7 , applying our defined measures to the chains generated from our experimental process yields statistical significant differences between the various combinations we used. In the next section we analyze what those differences mean in the context of our definitions of Satisfaction of Prompt Requirements , Cohesion , and Diversity . Analysis Our first observation is that the average creativity ranking, CR , across all the models regardless of temperature was relatively low, ranging from 0.0 to 0.24 . Manual examination showed that these low scores were due to multiple factors: early deviation from the prompt requirements, satisfaction of prompt requirements but low diversity, or the generation of unclassifiable objects such as glitchy patterns or irregular geometric shapes. None of the image generation models we tested performed too well in terms of creativity. However, our statistical tests and manual observations demonstrated significant differences in the scores between chain types and models, as detailed below. As can be seen in Tables 5 , 6 , and 7 , the paired T-tests demonstrate significant statistical differences between the results in all the img_cap and cap_only chains, supporting the claim that input images have a measurable impact on model outputs. Furthermore, for Flux and Stable-Diffusion 3 , there are statistically significant differences in CR values for img_cap chains between strength 0.3 and 0.9 , with CR values for higher-temperature chains being higher on average. This lends support to our original hypothesis that a higher image input temperature increases the creativity of an img2img model’s output. Table 5: Paired T-Test Results for CR in Flux chains strength img_cap mean cap_only mean P-value 0.3 0.18 0.24 0+\epsilon 0.6 0.2 0.27 0+\epsilon 0.9 0.22 0.27 0+\epsilon Table 6: Paired T-Test results for CR in Kandinsky 2.2 chains strength img_cap mean cap_only mean P-value 0.3 0.19 0.23 0+\epsilon 0.6 0.17 0.25 0+\epsilon 0.9 0.19 0.23 0+\epsilon Table 7: Paired T-Test results for CR in Stable-Diffusion 3 chains strength img_cap mean cap_only mean P-value 0.3 0.1 0.19 0\ +\ \epsilon 0.6 0.1 0.21 0\ +\ \epsilon 0.9 0.13 0.2 0\ +\ \epsilon Notably, all of the img_only chains demonstrated statistically significant differences from the other types of chains, with the average CR values for img_only chains being significantly lower than the img_cap and cap_only averages across all generation models and strengths. Only Flux img_only chains for strength 0.3 and 0.6 have average CR values above 0.0 (see Figure 4 for an illustration). The data in Tables 5 , 6 , and 7 and the statistically significant differences between img_cap chains of different temperatures further confirm our hypothesis that the temperature of image inputs directly influences the overall creativity of the output. However, the low average creativity scores of img_only chains also indicate that captions are important for image generation models, even the img2img ones we tested. The analysis below takes into account the granular measures Requirements Satisfaction Factor ( RS ), Cohesion Factor ( B_{R} ), and Diversity Factor ( D_{R} ), focusing on img_only and img_cap chains. The values of RS in Flux decrease sharply as strength increases, with RS=0.66 for strength 0.3 and RS=0.0 for strength 0.9 . For other too models, Stable-Diffusion 3 3 and Kandinsky 2.2 , average RS is close to 0 across all strengths. For RS we first looked at img_only chains, observing that in Flux , RS decreases drastically as strength increases, starting at 0.66 for strength = 0.3 and ending up at 0.0 for strength = 0.9; while for the other two models, average RS values are extremely low, approaching 0, across all strengths. This observation is supported by the visual examination of randomly sampled images created within img_only chains for each generator and is illustrated by Figure 4 , which shows a series of images generated from the same seed image, apple_pie_027.png in Figure 4(a) , at the first step of each of their respective img_only chains, with varying strengths. The extremely low requirement satisfaction is evident here: only the Flux chain at strength 0.3 shows an image resembling an apple pie. (a) Seed image apple_pie_027.png Figure 4 : Example images generated at step 1 of their respective img_only chains, all seeded with the same image. Flux in the top row, Kandinsky 2.2 in the middle row, and Stable-Diffusion 3 in the bottom row; strength 0.3 in the leftmost column, 0.6 in the middle column, and 0.9 in the rightmost column. img_cap chains, by contrast, have high average RS values across all three models, in the 0.7-0.85 range. These results indicate that the models we tested rely heavily on textual captions for correct and creative image generation, with the partial exception of Flux exhibiting some ability to generate faithful images from image input. At the same time, img_cap chains for all three models demonstrate a statistically significant difference in RS value between strengths 0.3 and 0.9 , indicating that image input strength affects the model’s ability to satisfy its original prompt even with textual captions. This claim is further supported by the fact that all three models show statistically significant differences between img_cap and img_only chains for strengths 0.6 and 0.9 , with Stable-Diffusion 3 showing differences between strengths 0.3 and 0.6 as well. For img_cap chains created with Flux and Stable-Diffusion 3 , strength 0.9 results in the highest average RS values. The most significant effect of strength is observed in Stable-Diffusion 3 chains, and the least significant one in Flux . Figure 5 shows the first image in each chain generated from the same seed image in Figure 4(a) , also organized by generator and strength level. Unlike the img_only chains, all of these initial images satisfy the prompt requirement by including the artifact “apple pie”. Figure 5 : Example images generated at step 1 of their respective img_cap chains, all seeded with the same image. Flux in the top row, Kandinsky 2.2 in the middle row, and Stable-Diffusion 3 in the bottom row; strength 0.3 in the leftmost column, 0.6 in the middle column, and 0.9 in the rightmost column. Adding to our observation that image input strength has a direct relationship with our creativity measures, average B_{R} scores for img_cap chains across all three models show a statistically significant increase between chains generated with strength 0.3 and those generated with strength 0.9 . img_only chains, on the other hand, only show statistically significant differences for Stable-Diffusion 3 , with all three models hovering around an average B_{R} of 0.14 regardless of strength. This disparity indicates that textual input is particularly important for cohesion. D_{R} , like B_{R} , is informed by RS : the diversity factor for a given chain is automatically set to 0.0 if the prompt requirements are not satisfied within the chain. img_cap chains for both Flux and Kandinsky 2.2 display no statistically significant differences between each other, and have average values around 0.25, but Stable-Diffusion 3 differs from both, with a significantly lower average D_{R} around 0.13. In contrast, all three models produced statistically significant differences between their img_only chains, but only Flux chains have average D_{R} values significantly higher than 0.0. Unlike the other measures, diversity does not have a straightforwardly direct relationship with image input strength, but seems to heavily rely on textual input. To summarize, our results show that image input strength has a measurable effect on the satisfaction factor , cohesion factor , and overall creativity ranking of the outputs produced by a given img2img generation model, while diversity appears to be less influenced by it. However, in order to produce outputs that score well in terms of creativity ranking, textual input is also crucial for all three models; out of the three that we tested, Flux seems to be the best at interpreting image-only inputs through a creative lens. Limitations Apart from computational resources, a significant limitation of our approach is its reliance on existing popular object detection and word similarity calculation tools. These models are prone to occasionally making decisions that do not seem intuitive to humans, such as placing two seemingly unrelated words close together in semantic embedding space, or labeling one object with a similar but incorrect designation (e.g. labeling a slice of apple pie as a slice of cake). There is currently no way around this limitation, but as the outputs of such tools more closely approach human-level understanding, the quality of our measures will improve accordingly. Additionally, in order to improve the visual accuracy and explainability of our proposed measures, future research could incorporate human observation as a factor in validating the scores output. We also acknowledge that the simplicity of our seed image dataset, although easy to experiment on and ideal for observing changes, could impact creativity scores. Seed images with a large artifact set may potentially increase D_{R} , for example, or make it harder for generated images to satisfy prompt requirements. However, exploring the creativity across diverse domains and more complex datasets subjects is a crucial next step. Conclusions and Future Research In this paper, we suggest a set of measures to quantitatively evaluate creativity in img2img generative models. Our experimental results visually and statistically demonstrate that the image input temperature (also referred to as “strength”) plays a statistically significant and measurable role in manifestations of creative behavior of img2img generative models. Moreover, textual input is also crucial for creative behavior; lack of textual input leads to generated images that bear very little resemblance to the input image. We can, therefore, answer the research question posed in the introduction affirmatively: our metrics provide a fine-grained score of creative behavior that conforms relatively well to human intuition and can be used to determine which model and at which strength is most suitable for a particular task. In the future, some existing metrics, such as LPIPS [ Zhang et al., 2018 ] for requirement satisfaction, could be incorporated into our creativity ranking to improve the granularity of our measures. Further directions for this research include exploring why some generative models score differently, and whether these metrics can help tailor future model architecture and training techniques to specific purposes.