Queen B Character Studio

The problem

Queen B: Makeover Merge is a dress-up merge game: every character has to look right across multiple ethnicities, multiple hairstyles, and a full set of expressions for each — before a single outfit is applied. Multiply those out and the roster runs into hundreds of illustrations before content even starts shipping, all needing the same face, proportions and identity. Illustrating that by hand doesn't scale to a live-ops schedule, and inconsistency here isn't a minor art bug — it's the thing players notice first.

The system

I built an AI-assisted character production pipeline that took a base identity through generation, correction, and slicing as one tracked flow, with a human approval gate before anything reached the game:

• Base generation: a neutral pose generated for each ethnicity and hairstyle combination, built from an LLM-driven brief so the variation stayed systematic rather than ad hoc.
• Expression spritesheet generation: the full expression set generated per combination from the locked neutral base.
• Face swap: an automated pass that reconciled each generated expression back against the locked identity, so five separate generations still read as one character.
• Background removal and slicing: sheets cut into individual, aligned sprite frames matching the animation rig.
• Review gate: every combination reviewed as Locked, Approved, or Rejected before promotion into the game.

The face-swap step was the piece that made the rest of the pipeline trustworthy. Generation alone produced expression sets that looked plausible individually but drifted from each other in subtle ways: a slightly different jaw, a slightly different eye shape. Reconciling every expression back against one locked reference is what let five independently generated images stand in as one consistent character.

What broke

• Face-swap seams: visible artifacts at the blend boundary on close inspection, worse on some hairstyles than others.
• Proportion drift between ethnicities: the same pose and framing reading slightly differently depending on skin tone and hair volume.
• Slicing misalignment: sheets that looked correct as a whole but broke the animation rig once cut into individual frames.

Each of those became a specific rejection reason at the review gate rather than a vague 'redo it', which is what made the pipeline improve over time instead of just producing a constant rate of rework.

What it proved

The pipeline shipped roughly 1,000 sprites for Queen B, a roster that would not have been art-directable by hand on a live-ops schedule. It's also the clearest evidence that the EverGen approach wasn't a one-off: the same principle, generate at volume, hold identity constant, gate everything through human review, applied to an entirely different production problem and held up.