Raw image models pull every character toward an average body and quietly grow back a missing arm. Mirva holds the design: a one-armed hero stays one-armed in every pose.


A turnaround of the design, and the one thing that must never change: he has only a left arm.
Mirva builds the expression sheet and action poses from the design.
Each result is checked against the character’s rule; any pose where the model grew the arm back is regenerated before it joins the set.
Climbing, fighting, sprinting, riding on his brother’s shoulders: the rule holds in every scene.
Raw image models learn from millions of bodies, nearly all with two arms. Ask one for a dynamic pose and it drifts toward that average: the missing arm quietly grows back, most often in poses that use both hands. A longer prompt can’t fix that, because every run starts over. Mirva works differently: the rule stays with the character, and Mirva reviews each pose against it, keeping a set of poses that pass and regenerating the ones that don’t. Below: raw model output next to the poses Mirva kept, from the same session.




Mirva’s own check, from the same session:
Both fixed ✅ Sprint: one arm pumping, right side is just cloth flowing off the stump. Punch: single left fist extended, right shoulder stump clearly visible with no arm below it.
This one nailed it ✅ One arm reaching up to grip the bark, right shoulder clearly just a stump with cloth, bare feet pressing into the tree for support.






