Indie Studio Trains AI Model on Its Own Artists for Mobile Game

Studio builds game and AI engine in parallel

Studio Atelico is developing Bobium Brawlers, a mobile card-and-dice battler where players create creatures through text descriptions, alongside an on-device AI engine for game developers. The image generation model powering the game was trained exclusively on artwork commissioned from the studio's own participating artists, according to CEO and Co-Founder Piero Molino.

The studio, formed by veterans from Uber, Meta, SEGA, and Creative Assembly, treats generative AI as a player-facing mechanic rather than a production tool. Players describe creatures in text, and the game generates both a portrait matching the established visual style and a playable deck following designer-authored rules.

Why it matters

As AI adoption in game development faces scrutiny over training data sources and artist displacement, Studio Atelico's approach demonstrates an alternative model: consent-based training, on-device inference to eliminate per-use costs, and AI constrained by human-authored game design. The studio's publicly available Artist Rights Contract proposes explicit consent, project-specific usage rights, and revenue participation—addressing concerns that have made AI controversial among game developers and artists.

Testing revealed need for constraints

After building more than ten prototypes, the team discovered that maximizing AI control did not improve gameplay. "Too little AI didn't create experiences that felt genuinely new, while too much AI often made the game feel arbitrary and made it harder for players to develop mastery," Molino explained in an interview with 80.lv.

The breakthrough came from pairing AI with strong authored mechanics. Designers define rules, creative direction, and constraints, while AI provides personalization within those boundaries. The creature generation system exemplifies this: players invent original creatures, but the output must fit the game's visual identity and gameplay framework.

Iterative training with artist feedback

Artist José Luna initially drew approximately 20 creatures to train the model. When integrated into development builds, the model revealed gaps—generating all dragons as purple because the training set contained only one purple dragon, or failing to consistently add eyes to inanimate objects described as creatures.

Luna created additional artwork targeting these deficiencies, and the model was retrained. After beta release, playtester feedback uncovered new edge cases. Luna has now created around 100 creatures, with more planned post-launch. "The more people experiment with the system, the more unusual cases we discover," Molino said.

On-device inference removes economic constraints

The studio prioritizes running AI models directly on player hardware rather than cloud servers. Beyond technical benefits like lower latency and privacy, local inference removes per-interaction costs that can limit how developers design AI-driven mechanics. "If every AI interaction has a cost attached to it, developers inevitably begin designing around that limitation," Molino noted.

Statistical testing replaces deterministic QA

Testing AI systems requires evaluating whether outputs stay within creative and gameplay constraints rather than producing identical results. The team uses statistical evaluation methods similar to clinical trials, aiming for confidence rather than certainty. Checklists define desired and prohibited behaviors, while red teaming stress-tests the system for inappropriate content.

Content safety involves balancing false negatives (inappropriate content passing through) against false positives (legitimate requests blocked). The team tracks these metrics and adjusts boundaries as more players interact with the system.

Studio Atelico is also working with other studios using the engine. These details were first reported by 80.lv in an interview with Molino.