Cosplay: Culture, Craft, and Computation

• Cosplay is a fan-driven practice where enthusiasts embody fictional characters through meticulously crafted costumes, performative art, and digital engagement.
• Craft production in cosplay emphasizes DIY ethics, precision in costume design, and adherence to community norms that ensure screen-accuracy and creative authenticity.
• Recent research integrates computational methods like image-to-image translation and 3D human generation to enhance costume design and audience-interactive wearable systems.

Searching arXiv for the papers on arXiv and closely related cosplay research to ground the article. Cosplay is a Japlish portmanteau of “Costume” and “Play,” and denotes a practice in which participants dress as characters from manga, games, and anime while embodying fictional personas through clothing, wigs, makeup, props, and performance. In Japan it is described as a female-dominated niche subculture of extreme fans and mavens, while in a broader economic framing it has grown from its origins at fan conventions into a billion-dollar global dress phenomenon. Recent scholarship treats cosplay simultaneously as a subcultural learning system, a site of intensive craft production, and a target for computational methods including image-to-image translation, wearable interaction systems, and layered 3D human generation (Matsuura et al., 2015, Tango et al., 2020).

1. Cultural definition and social scope

Cosplay is first and foremost an expression of love for anime, manga, and their characters. The practice extends beyond simple dress-up: cosplayers rehearse signature poses or scenes, select wigs and styling methods that match hairstyle profiles, and construct or modify props and accessories to achieve screen-accuracy. In Japan, the majority of active cosplayers are female, predominantly high-school and college students or young adults in their twenties; the ethnographic study by Matsuura and Okabe examined ten female informants aged 18–25, with 2–8 years of cosplay experience, recruited by snowball sampling during fieldwork conducted from August 2011 to December 2013 (Matsuura et al., 2015).

Cosplay is organized through both online and offline venues. Large-scale gatherings, including theme-park cosplay parties and comic-market fanzine fairs, provide stages for display, photography by peers or audiences, and informal after-event parties. Dedicated SNSs for cosplayers, including “Cure,” Twitter, and specialized blogs, function as persistent infrastructures for exchange: users upload high-resolution photos of works-in-progress, pattern tutorials, painting steps, and wig-styling sequences. This combination of event-based display and networked documentation makes cosplay a public yet highly technical form of fan labor.

A plausible implication is that cosplay should be understood less as a single event and more as a distributed production-and-performance pipeline. Costume making, photography, online circulation, and peer evaluation are not ancillary to the practice; they are constitutive of it.

2. Craft production, accuracy regimes, and normative order

A core ethic of the community is DIY: sewing, painting, and crafting one’s own outfits and accessories, even though ready-made costumes exist. Cosplayers are highly conscious of quality standards for costumes, makeup, and accessories. Attention to character accuracy includes correct fabrics, precise color-matching, and weathering or texturing techniques. Prop-making often involves wood, foam, resin or 3D-printed parts, and paint and finish are treated as critical. One informant stated, “If you don’t paint it properly, it doesn’t match the character” (Matsuura et al., 2015).

The tolerance regime is exacting. Weapons, jewelry, belts, and small ornaments are constructed or modified to achieve screen-accuracy, and tolerances of millimeters in shape or hue are not uncommon. Makeup emphasizes enlarged eyes using colored contacts, false lashes, and precise contouring. Hair work is similarly constrained by character-specific shape profiles. These practices position cosplay within a broader category of vernacular precision fabrication, where aesthetic fidelity is governed by collectively enforced standards rather than by formal certification.

Cosplay also operates through explicit and tacit rules. Written taboos include no unapproved photography and no commercial flogging, while unwritten norms govern courtesy during photoshoots and behavior at events. The interview analysis reported categories including DIY Ethics, Shared Rules and Codes of Conduct, Peer Review, Rejecting Commercial/Mainstream Cosplay, Reciprocal Learning, Sharing Cosplay Knowledge on SNSs, Learning from Others’ Digital Data, SNSs as Scaffolding System, and Standardization and Creating New Tasks (Matsuura et al., 2015).

This suggests that cosplay’s material culture is inseparable from its normative order. Craftsmanship is evaluated not only by technical execution but also by conformity to community-specific rules of respect, authorship, and participation.

3. Peer-based learning and distributed scaffolding

Matsuura and Okabe frame cosplay as a “peer-based, reciprocal learning environment” and relate it to connected learning by Ito et al. In this environment, members both teach and learn from one another without a formal instructor. Photo-sharing on “Cure” or Twitter triggers waves of retweets that signal peer approval and simultaneously function as distributed tutorials for newcomers. An interview example states: “You can see how to make the weapon on Twitter. … When I make this weapon, I’m really careful with the painting process” (Matsuura et al., 2015).

The paper’s central theoretical move is to expand Bruner’s classical notion of scaffolding. In the classical account, a more knowledgeable other interacts directly with a learner and provides temporary support until independence is reached. In cosplay, by contrast, scaffolding is described as emerging from the entire ecology of peers, SNS archives, and events. The study formalizes this relation with the shorthand

$$\text{ConnectedLearning} = \mathcal{I}\cap\mathcal{P}\cap\mathcal{S},$$

where $\mathcal{I}$ denotes interest-powered practice, $\mathcal{P}$ peer-supported relationships, and $\mathcal{S}$ shared purpose and public recognition. It also provides a dynamical sketch of skill acquisition: $$\frac{dU_i}{dt} = \alpha\sum_{j\in\mathrm{Peers}_i}\mathrm{Interaction}_{ij}(t) + \beta\,\mathrm{ResourceAvailability}_i(t),$$ where $U_i$ is cosplayer $i$’s skill level, $\mathrm{Interaction}_{ij}=1$ when $i$ uses $j$’s tutorial, and $\mathcal{I}$0 indexes the richness of online archives available to $\mathcal{I}$1 (Matsuura et al., 2015).

Methodologically, the study used ethnographic fieldwork, participant observation, digital audio recordings of informal post-event conversations, and semi-structured interviews. The transcripts were analyzed via SCAT, a four-step coding process attributed to Otani (2008). The resulting account treats cosplay not merely as fandom but as a decentralized pedagogical system in which public documentation, critique, and iteration sustain both standardization and innovation.

4. Computational synthesis and digital representations of cosplay garments

A major computational treatment of cosplay appears in “Anime-to-Real Clothing: Cosplay Costume Generation via Image-to-Image Translation,” which addresses the problem of translating animated costume imagery into plausible real garments. Because cosplay items can be significantly diverse in their styles and shapes, the paper begins with large-scale dataset construction and calibration. Web crawling used queries of the form “cosplay costume A B,” where $\mathcal{I}$2 is an anime title sampled from MyAnimeList and AnimeRecODB and $\mathcal{I}$3 is one of 40 online cosplay shop names, producing approximately 1 TB of images. An active-learning pipeline manually labeled about 300 images as “good” or “bad,” fine-tuned an ImageNet-pretrained VGG-16 classifier on 2,760 train / 109 val samples, iteratively retrained it to a final 3,052 train / 215 val set, and discarded negatives. Bounding boxes for character and clothing were annotated on 1,059 images to train an SSD detector; near-duplicates were removed with DupFileEliminator at a 90 % similarity threshold; and position calibration used Waifu2X, mirror-padding, Gaussian-blurred margins, and a FashionAI key-point detector based on Cascaded Pyramid Network. The final paired dataset contained 35,633 $\mathcal{I}$4 image pairs, with 32,608 training pairs and 3,025 testing pairs (Tango et al., 2020).

The proposed GAN uses a U-Net generator with skip connections and coarse-to-fine progressive-growing training. It combines a domain-pair discriminator $\mathcal{I}$5, implemented as three identical sub-discriminators at different image scales to classify paired versus unpaired $\mathcal{I}$6, with a real/fake discriminator $\mathcal{I}$7, implemented as three patch discriminators producing logits at patch sizes $\mathcal{I}$8, $\mathcal{I}$9, and $\mathcal{P}$0. Spectral normalization is applied to every layer. Auxiliary terms include feature-matching losses on $\mathcal{P}$1 and $\mathcal{P}$2, an input-consistency loss on $\mathcal{P}$3 features between $\mathcal{P}$4 and $\mathcal{P}$5, and an $\mathcal{P}$6 reconstruction loss with $\mathcal{P}$7. The full objective is

$\mathcal{P}$8

Using FID and LPIPS, both lower-is-better metrics, the paper reports the following comparison on the same test set:

Method	FID	LPIPS
Pixel-level Domain Transfer	248.47	0.8220
pix2pix	55.44	0.5820
pix2pixHD	197.59	0.6857
Proposed method	30.38	0.5786

The full model achieves the best FID of 30.38 and a competitive LPIPS of 0.5786. Qualitatively, the authors attribute fine textures such as lace and metallic buckles, together with coherent global silhouettes, to the multi-patch $\mathcal{P}$9 and the input-consistency term; reported failure cases include outlier inputs, toy-like proportions, and very novel styles unseen during training (Tango et al., 2020).

A separate but adjacent line of work concerns layered 3D human generation. HumanCoser aims to generate physically-layered 3D humans from text prompts through a layer-wise dressed human representation, a dual-representation decoupling framework, multi-layer fusion volume rendering, and an SMPL-driven implicit field deformation network that enables the free transfer and reuse of clothing. The provided technical exposition applies this framework to layered 3D cosplay costumes by representing the minimally clothed body and successive clothing or armor pieces as independent NeRF-style volumes $\mathcal{S}$0, using dual SDS losses and density-separation regularization during training, and then adapting the learned layers to arbitrary body shapes and poses via SID-Net. A plausible implication is that physically layered representations are especially relevant to cosplay because reusable garments, armor plates, and part-level editing are central to virtual try-on and layered human animation (Wang et al., 2024).

5. Internet-connected cosplay and audience-driven interaction

“IoT Skullfort: Exploring the Impact of Internet Connected Cosplay” examines cosplay as a wearable interactive system rather than only as a costume artifact. The implemented costume embeds an Adafruit Feather nRF52 Bluefruit LE microcontroller based on the nRF52832, a Neopixel-style LED strip with 60 individually-addressable RGB LEDs per meter, a single-cell 3.7 V Lithium-Polymer battery, and a PowerBoost 1000 that produces a regulated 5.2 V rail for the LEDs. The helmet body is fabricated from EVA foam panels cut and assembled via Pepakura 2D flattening of a 3D model created in Blender, while a 3D-printed mold is used to vacuum form a PETG visor. Communication uses Bluetooth Low Energy in a star topology with a single central device and one peripheral, with typical BLE throughput of approximately 200 kbps raw payload and 20–100 ms round-trip latency (Beckett et al., 2019).

The paper’s distinctive design choice is to shift control from the wearer to the audience. A companion mobile app is described as exposing controls for selecting a static RGB color, launching pre-defined animations such as breathing, chasing, fade, strobe, or rainbow cycle, and sending short text messages if scrolling text mode is enabled. The firmware can be modeled as an event-driven state machine with Idle/Default Animation, Command Received, Transition, and Hold/Play states. For fade transitions, the mapping algorithm is given by

$\mathcal{S}$1

applied per LED to produce a smooth color fade.

Evaluation was qualitative rather than metric-driven. A single one-hour focus group with six cosplay community experts, each with decades of participation, alternated performer and audience roles after a system demonstration. No formal rating scales were used. Thematic findings covered learning curve for audience, attention management in crowds, group interaction and fairness, health and safety including photosensitivity and disability accommodations, and content moderation for hate speech or abusive messages. All experts agreed that audience-driven control significantly enhances engagement but raises new UX, safety, and moderation challenges. The paper also identifies power and reliability constraints: BLE updates must be less than 100 ms to feel live, crowded RF environments threaten connectivity, and a typical 1 Ah LiPo pack powers 60 LEDs for about 1.5–2 hours at medium brightness (Beckett et al., 2019).

The broader significance lies in reframing cosplay as performer–audience co-creation. Here the costume is not a static representation of character identity but a networked interface whose visible state can be manipulated by third parties.

6. Acronymic uses of “COSPLAY” in technical literature

Within arXiv and adjacent technical literature, “COSPLAY,” “COS-PLAY,” and “CoSPlay” also appear as acronymic names for machine-learning systems unrelated to costume play. This creates a terminological ambiguity that is especially relevant in search and citation practice.

In dialogue generation, COSPLAY expands to “COncept Set guided PersonaLized dialogue generation Across both partY personas.” The method represents self-persona, partner persona, and dialogue in concept sets over a ConceptNet-derived vocabulary of size approximately 2600, uses set algebra, set expansion, and set distance, and adds reinforcement signals for Mutual Benefit and Common Ground. On Persona-Chat, with train/dev/test splits of 8,939/1,000/1,000 dialogues, the model reports $\mathcal{S}$2 Hits@1 and $\mathcal{S}$3 F1 on the original benchmark, and human-evaluation scores of 4.52 fluency, 4.35 engagement, and 4.37 consistency, outperforming strong baselines on engagement and overall average (Xu et al., 2022).

In long-horizon agent research, COS-PLAY denotes “Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks.” The framework couples a decision agent $\mathcal{S}$4 and a skill bank agent $\mathcal{S}$5, both based on Qwen3-8B with LoRA adapters, and updates them through Group Relative Policy Optimization. The decision agent retrieves skills from a dynamic bank, while the skill bank agent segments rollouts, learns contracts, and curates reusable skills. Across six game environments, the paper reports that COS-PLAY with an 8B base model achieves over 25.1 percent average reward improvement against four frontier LLM baselines on single-player game benchmarks, while remaining competitive on multiplayer social reasoning games (Wu et al., 22 Apr 2026).

In code generation, CoSPlay stands for “Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test.” It maintains a code pool $\mathcal{S}$6, a test pool $\mathcal{S}$7, and an execution matrix $\mathcal{S}$8 where $\mathcal{S}$9 if code $$\frac{dU_i}{dt} = \alpha\sum_{j\in\mathrm{Peers}_i}\mathrm{Interaction}_{ij}(t) + \beta\,\mathrm{ResourceAvailability}_i(t),$$0 passes test $$\frac{dU_i}{dt} = \alpha\sum_{j\in\mathrm{Peers}_i}\mathrm{Interaction}_{ij}(t) + \beta\,\mathrm{ResourceAvailability}_i(t),$$1. Through exploration–attack idea generation, execution-matrix-driven iterative self-play, and output-consensus-based cluster selection, it co-evolves code and unit tests without ground-truth tests or model updates. On four competitive-programming-style benchmarks, it improves average BoN from 22.1% to 33.2% and UT accuracy from 14.6% to 78.3% on Qwen2.5-7B-Instruct, and when applied to CURE-7B it further improves BoN by 5.7% (Hu et al., 22 May 2026).

These acronymic usages are semantically independent of cosplay as costume practice. Their inclusion in arXiv indexes nevertheless makes “cosplay” a notable example of cross-domain naming collision between fan-culture terminology and ML system nomenclature.