How to get a generated character to use the hand you asked for
Tell Vizard Agent which hand does what in each scene. A model often takes "left" to mean the left of the image, so it rewrites every instruction in screen terms for that pose, such as "his left hand, on the right of the picture", chooses a view where the hands can be told apart, and crops into the hands to check before moving on.
What do you type into Vizard Agent?
Write the hands into the scene the way you would brief an illustrator, one action per hand. In a story the hand matters: a bundle carried on the left shoulder in one scene should not jump to the right in the next, and a character who points with one hand and waves with the other only reads if both are where you said.
The stick with the cloth bundle lies on his left shoulder, and his left hand holds the end of it. His right hand knocks on the gate.
What does Vizard Agent actually do?
It treats left and right as a translation problem. A character facing the camera has their left hand on the right of the picture; seen from behind, the two agree; seen from the side, one hand may not be visible at all. In the session behind this article it:
- Added an explicit hand assignment to every prompt, converting the character's own left and right into screen sides for that pose.
- Cropped the shoulders and hands of a rendered scene, enlarged them three times, and checked which arm actually held the pole after the first version put it on the wrong side.
- Found the side view hid the left arm, and redid the scene as an almost straight rear view, where the character's left is also the screen's left.
- Kept a rear three-quarter view in the following scenes so the carry stayed readable, and restated which hand stayed on the staff when the other took a new object.
- Re-checked the hands before every upload, including after a style pass.
When does this not work well?
When the pose hides one hand. No wording makes a model draw a hand the camera cannot see, so the fix is a different angle, not a stronger prompt. Very small figures are the other weak point: at a distance the hands are a few pixels and the check cannot settle which is which.
Common questions
Why does the character use the wrong hand? "Left" got read as screen-left. Vizard Agent translates it per pose.
Can I keep the same hand across scenes? Yes. Vizard Agent restates the assignment in every scene.
How does it check? Vizard Agent crops and enlarges the hands before moving on.