Sketchar: AI Co-Creation for Game Characters
- Sketchar is a web-based human–AI co-creation tool for game character design that streamlines conceptual prototyping, visual reference generation, and designer-illustrator collaboration.
- It employs a multi-step workflow that integrates narrative inputs, LLM-driven keyword extraction, and DALLE-generated images to support iterative design refinement.
- Empirical studies show that Sketchar enhances design communication and creativity, particularly benefiting non-artistic designers, despite current image detail limitations.
Sketchar is a web-based, human–AI co-creation tool for game character design that supports conceptual prototyping, visual reference generation, and communication between designers and illustrators. In “Sketchar: Supporting Character Design and Illustration Prototyping Using Generative AI,” the system is presented as a workflow in which designers provide narrative and attribute-level inputs, receive structured character profiles and DALLE-generated reference images, iteratively refine them through human-in-the-loop interaction, and package the outputs for later collaboration with illustrators (Ling et al., 17 Aug 2025).
1. Problem setting and design rationale
The system is motivated by communication problems in game character design. The paper states that designers tend to communicate in text, while illustrators must translate concepts into visuals, often causing “misinterpretation, resulting in repetitive cycles of changes that require significant time and resource investment” (Ling et al., 17 Aug 2025). Formative interviews with 10 professionals describe this gap in concrete terms. One respondent reported “noticeable challenges in communication between the art team and the game designers” due to “difference in thought processes,” while another emphasized that subtle details are “challenging to articulate verbally,” leading to “distortion or loss of information” if illustrators do not infer them correctly. A designer summarized the problem directly: “I have an awesome image of a character in my head, but I don’t know how to express it.”
Within that framing, Sketchar positions generative AI as an intermediate representational layer. The paper argues that such systems can produce reference images from narrative inputs and can use LLMs to structure text so that designers can refine keywords and profiles. It further reports that GenAI can assist when designers “fail to find the highlights and inspiration” or when character stories become “increasingly difficult” (Ling et al., 17 Aug 2025).
The design is organized around three explicit goals. DG1 is to enhance keyword-driven character image and profile generation for efficiency and coherence. DG2 is to construct sharing channels as “communication bridges” between designers and illustrators. DG3 is to brainstorm character properties and inspire creative ideas using GenAI. These goals define Sketchar not merely as an image generator, but as a workflow instrument for ideation, refinement, and handoff.
2. Technical architecture and generation pipeline
The reported architecture consists of a React front end, a Flask back end, OpenAI API integration for ChatGPT (GPT-3.5) and DALLE, and an MSDB database for storage and management (Ling et al., 17 Aug 2025). The system architecture caption also mentions CLIP. The paper does not specify sampler settings, guidance scales, or comparable low-level image generation hyperparameters; instead, control is exerted through prompt structure, iterative editing, and artifact management.
Input begins with free-form narrative and structured fields. The first-layer inputs are “name,” “role details,” “background story,” and “game type,” with optional “render style.” Prompt engineering imposes explicit constraints such as “no more than 150 words,” and it uses fields including “name, age, dressing style, weapon, background story.” The generation pipeline is hierarchical. First, ChatGPT summarizes the designer inputs into “integrated information.” Second, the LLM generates keywords from that summary. Third, DALLE consumes the keywords together with render style and role details to generate images.
The image subsystem is described as DALLE or DALL·E 2 used as the text-to-image engine. In the human-in-the-loop stage, each generation produces “five reference images” for character depiction. The text subsystem produces a structured character profile summarized by ChatGPT, with refined keywords that can be edited and reused in later iterations (Ling et al., 17 Aug 2025).
Two auxiliary modules extend the core pipeline. A role agent uses JSON as fine-tuning content to define personality and memory structure for LLM interactions. A family tree module built with React Flow allows users to “build relationships by dragging lines” and to “fill in the relationship name below the card.” These additions broaden the system from single-character generation toward narrative and relational organization.
3. Interaction workflow and produced artifacts
The workflow is described as a sequence of seven steps. Step 1 is initialization, where the designer enters design intentions. Step 2 is result generation, in which ChatGPT provides “prompt suggestions and potential images.” Step 3 is modification and regeneration. Step 4 is detailed editing, allowing direct changes to appearance, keywords, personality, and background narrative. Step 5 presents the result as an “ID card.” Steps 6–7 support sharing, conversing with generated characters, and constructing lineage trees (Ling et al., 17 Aug 2025).
This structure embeds human-in-the-loop editing at several levels. Designers can select among the five generated reference images, revise prompt content, adjust textual profiles, and rerun the generation process. The resulting artifacts include a structured character profile, a set of DALLE-generated reference images, an “ID card” presentation, downloadable character files for sharing, and relationship maps that connect multiple character nodes.
The conversational subsystem is designed so that AI answers preserve personality and context derived from the generated profile. The relationship subsystem supports cross-character organization rather than isolated outputs. In practical terms, Sketchar therefore combines prompt authoring, iterative image selection, structured profile editing, artifact packaging, and lightweight narrative simulation into a single interface.
The paper also emphasizes that outputs are meant to travel. Designers can save the ID card, download files, and share them with collaborators. In the reported workflow, this makes the generated profile and selected references available for later use by illustrators in subsequent design stages (Ling et al., 17 Aug 2025).
4. Empirical studies and reported results
The paper reports a mixed-method program consisting of formative interviews, a qualitative study, a quantitative study, and expert evaluation (Ling et al., 17 Aug 2025).
| Study | Participants | Main reported result |
|---|---|---|
| Formative interviews | 10 professionals | Identified communication friction and difficulty expressing design intentions visually |
| Qualitative study | 13 game designers | Reported refinement of design details and usefulness in real-world workflows |
| Quantitative study | 17 participants | Total CSI significantly higher than baseline, with |
| Expert evaluation | 5 illustrators | Average 6.84/10 images per set judged “relevant valuable references” |
The formative interviews involved 10 professionals, 3 female and 7 male, all of whom had studied game design or worked in game companies. The method used online interviews, audio recording, and thematic analysis with open coding by three researchers followed by a consensus codebook. These interviews established the communication and ideation problems that motivated the system.
The qualitative study involved 13 game designers in China, 5 female and 8 male. Their backgrounds included 5 from small studios, 3 from prominent companies, 2 indie practitioners, and 3 students. More than 50% had more than 3 years of experience; 6 had strong artistic proficiency and 7 did not. The tasks required designing a master character for a cyberpunk open-world game and a master character for a platformer anime game. The section text reports 25 minutes for Task 1 without Sketchar, 5 minutes of Sketchar instructions, 25 minutes for Task 2 with Sketchar, and then a semi-structured interview. The accompanying figure reports “20 minutes” for Task 1, “25 minutes” for Task 2, and a “10-minute semi-structured interview.” The study used observations and semi-structured interviews, again analyzed thematically by three researchers.
The quantitative study involved 17 participants, 9 males, aged 21–27, with Mean and SE . Nine had art experience of at least one year and eight did not; all had at least one game development experience as character designers. The design was a within-subject comparison of a Baseline sketching-only workflow versus a Sketchar-assisted workflow, with randomized order. After each task, participants completed the Creativity Support Index across six dimensions on a 0–10 scale. Because the data were non-normal, the authors used a two-way nonparametric ANOVA (Scheirer–Ray–Hare), which found no significant interaction (). A Wilcoxon signed-rank test showed the Sketchar condition had significantly higher paired CSI totals than baseline (). Reliability analysis found “Cronbach’s alpha of each CSI factor was greater than 0.7.” A Wilcoxon rank-sum test further showed that total CSI under Sketchar was significantly higher for participants without art skills than for those with art skills (). At the factor level, Expressiveness () and Results Worth Effort () were significantly higher for the non-art group under Sketchar. The paper also reports that the Collaboration score was significantly higher under Sketchar than baseline, while baseline had higher Enjoyment, possibly because of familiarity with traditional workflows (Ling et al., 17 Aug 2025).
The expert evaluation involved five illustrators. For five distinct prompts with ten images each, the experts identified an average of 6.84/10 images per set as “relevant valuable references” suitable for professional workflows. The paper treats this as evidence that the outputs can be incorporated into real-world pipelines.
5. Reported effects on design practice
The central qualitative finding is that the co-created reference images “fostered refinement of design details and can be incorporated into real-world workflows” (Ling et al., 17 Aug 2025). Designers reported that Sketchar “can help with basic character analysis and style generation,” especially by “swiftly organizing appearance, weaponry, and personality details.” One participant stated that the generated documents could “serve as the basis for my further detailed design” and could “inspire me to think about the whole character’s setting.”
The system appears to be particularly valuable for designers without strong drawing ability. The abstract states that designers without artistic backgrounds found the workflow to be more expressive and worthwhile. Participant testimony reinforces this: “Since I don’t know how to represent my design in the form of art, I thought generating some simple reference images would help me communicate my design.” Another reaction, “Something is better than nothing,” captures the way the tool substitutes structured visual proxies for absent manual illustration skill.
The paper also reports concrete handoff benefits. Designers used images to articulate specifics to illustrators, as in the statement: “I can inform the illustrator that I desire the face and hairstyle depicted in this picture while requesting adjustments …” The discussion of workflow transfer is explicit: designers package edited profiles and selected reference images into an ID card and downloadable files, then share them with illustrators. The next step described in the paper is that illustrators use the outputs as blueprints, reducing ambiguity (Ling et al., 17 Aug 2025).
At the same time, the study distinguishes between designer groups. Participants with strong art backgrounds demanded higher fidelity. One such response was: “The generated image is a little bit easy and lacks many details compared with the standard of my workflow.” Another participant noted that DALLE outputs can be “hard to utilize in pragmatic workflows” unless detail is sufficient. The system’s benefits are therefore not reported as uniform; rather, they are stronger for non-art designers and more conditional for skilled visual practitioners.
6. Limitations, ethics, and relation to adjacent sketch-centered research
The paper identifies several limitations. First, image detail and control remain imperfect. Some designers required multiple iterations; one reported example needed “5 iterations to achieve an outcome with distinct facial tattoos.” The paper attributes this to DALL·E 2 limitations in fine-grained visual features. Second, collaboration in the study was asynchronous. Although the system is intended to bridge designers and illustrators, actual synchronous co-creation between the two groups was not observed; the CSI Collaboration result therefore reflects perceived support rather than directly measured live team dynamics. Third, the participant pool was drawn from mainland China, where access constraints limited prior exposure to ChatGPT and DALL·E, so the paper identifies cross-cultural validation as an open issue (Ling et al., 17 Aug 2025).
Ethical issues are treated explicitly. The paper notes skepticism among artists regarding generative AI, especially around copyright, ownership, and training data provenance. It also documents representational bias: when prompting for indigenous characters such as Māori, DALLE produced stereotyped outputs. The proposed safeguards include cultural counselor review for marginalized or Indigenous representations and warnings when counselor review is not available. The system’s use of prompt history and versioned ID cards is presented as a transparency mechanism for handoff and traceability (Ling et al., 17 Aug 2025).
Future work follows directly from these limitations. The authors propose exploring “Stable Diffusion or DALL-E 3” for higher detail and control, studying synchronous designer–illustrator collaboration, conducting broader cross-cultural evaluations, and directly comparing Sketchar against ChatGPT plus DALL·E configurations. They also suggest investigating whether structured outputs improve designers’ document organization and visual expression skills over time.
Within the broader literature, similarly sketch-centered systems address different problem classes. “Sketch-a-Classifier” studies photo classifier synthesis from a user’s free-hand sketch without category-specific training photos (Hu et al., 2018). “CustomSketching” introduces sketch concept extraction for fine-grained sketch-based image synthesis and editing in Stable Diffusion v1.5 (Xiao et al., 2024). “RealitySketch” embeds responsive graphics and visualizations in AR through dynamic sketching (Suzuki et al., 2020). “Query by Semantic Sketch” formulates image and video retrieval by painting concept maps that encode what is where (Rossetto et al., 2019). “A Sketch Is Worth a Thousand Words: Image Retrieval with Text and Sketch” addresses image retrieval using both a sketch and a text query (Sangkloy et al., 2022). Against that background, Sketchar is specifically oriented toward early-stage game character ideation, reference generation, and interdisciplinary communication rather than zero-shot classification, AR binding, or media retrieval.