AI Character Design on Astorie

Character design is a multi-deliverable phase, not a one-image generation. An indie game studio building a 2D platformer protagonist needs a master portrait, a turnaround sheet (front, back, three-quarter), an expression sheet (happy, angry, surprised, sad, sleeping, determined), and outfit variations (default, alternate costume, holiday skin) — all anchored to the same character. Tab-based AI image tools generate one portrait per session and have no way to lock the design across the dozen-plus deliverables that constitute a real character bible.

The deeper problem is consistency at the design stage itself. A novelist writing a Substack series with recurring protagonists, an animator building a 2D-rigged short with one returning character, a tabletop publisher producing a campaign book with a fixed cast — every long-form creative project depends on the character holding identity across hundreds of generations. Without an upstream anchor that locks silhouette, color palette, and outfit signatures, the character drifts within the same project and the world breaks immersion.

And there is the legal honesty layer. Character design in the style of Pixar, Studio Ghibli, Marvel, Disney, or a living artist (Loish, Karla Ortiz, Greg Rutkowski) is exactly where AI image work runs into trademark and right-of-publicity risk. Style mimicry of a registered IP or a recently active artist is litigation-prone for commercial release. A workflow that frames AI character design as concept exploration — build an original style, anchor it, and frame the output as reference for the real character artist to clean — is honest about where AI fits in the production pipeline.

Astorie treats character design as a multi-anchor canvas problem. The protagonist description lives as a text node — 'wiry teen runner, deep navy hood, copper-rimmed goggles, signature scar on the left cheek.' The style reference (line weight, color script, world tone) lives as another labeled image node. Wire both into Midjourney for the master portrait. Once the master lands, every downstream deliverable — turnaround, expression sheet, outfit variations — wires into the master portrait as the new upstream anchor. The character holds identity because the chain pulls from one locked source, not from a re-typed prompt.

Multi-model fanout for the master portrait, single-model for the deliverables. For the master, fan across Midjourney (editorial composition), Flux (high-fidelity rendering), Nano Banana 2 (reference fidelity for sequel work), Flux Kontext (edit-aware refinement), and GPT Image 2 (cleanup pass). Pick the strongest direction. For the turnaround, expression sheet, and outfit variations — the catalog of character views that make up the bible — lock the winning model and run them all from the same master portrait anchor. The bible scales on the chain rather than on per-view re-prompting.

Downstream chaining hands off to the production pipeline. Once the bible lands, chain into the image-upscale tool for high-resolution masters that the rig artist can clean and rig in Spine, Live2D, or DragonBones. Save the canvas as a character template; future sequels, season-two storylines, or alternate-universe variations inherit the locked anchors and only swap the per-view prompts. Frame the deliverable honestly — concept-stage character design for an artist team to clean and finalize, not engine-ready production sprites.