Change Clothes Without Changing the Face
If you've tried outfit swaps with a full-image editor, you know the pattern. You ask for a new jacket. You get the jacket — plus a slightly different face. Different cheekbones, different expression, the uncanny feeling of meeting a stranger wearing your character's clothes.
For hobby art that's annoying. If it's a character with a following, a VTuber persona, a game concept, or a brand mascot, it's disqualifying.
Why the face keeps drifting
The face changes because every pixel goes back through the model. Full-image editing has no concept of "the face is correct, leave it alone." Your character's identity is re-synthesized from scratch on every run, and "close enough" is the best guarantee available.
The only real protection is structural: keep the face's pixels out of the edit entirely.
The layer setup that works
Upload your character image to the editor and decompose. On a typical character render you'll get layers in the neighborhood of subject, garment, hair, background. Names vary with the image; what matters is what's inside them.
Before editing anything, click through the stack and find where the face landed. Usually it's inside the main subject layer. If decomposition split it oddly — occasionally the face region gets its own layer, which is a gift — just note what you have.
Now the important part. The outfit lives on its own layer. Editing that layer is a scoped operation: new clothes come in through the model, and the face never enters the pipeline. It doesn't drift because it isn't sampled.
Two habits that make this reliable:
Lock the layers you're not touching. Locked layers are passed to the edit as preserve rules — an explicit instruction that those regions are fixed. It's belt and suspenders, but with a face on the line, wear both.
One outfit per edit. Wardrobe sheets come from repetition, not from one ambitious prompt. Same face layer, same pose, swap the garment layer: spring jacket, summer tee, winter coat. Export each combination. The character stays coherent across the set because identity pixels are the same pixels every time.
For anime and illustration
Cel-shaded art separates cleanly — flat regions decompose into flat layers, and edits land with crisp boundaries. It's the easiest category for this workflow.
Loose brushwork and heavy motion blur are harder. When detail is smeared across region boundaries, decomposition sometimes fuses them. The workaround is a higher layer count: more, smaller layers means fragile details end up inside layers you weren't going to touch anyway.
Checking your result
Trust toggles, not vibes. Hide the edited layer and confirm the character underneath is untouched. Then compare against your reference at full zoom — silhouette first, then landmarks. If anything outside the outfit moved, the edit scope was wrong somewhere, and the fix is a narrower instruction, not another roll of the dice.
We keep a longer writeup on editing characters without recasting them, including the verification habit in more detail.
The one-line version: your character's identity is pixels, and those pixels deserve to be treated as an asset you protect, not a description you re-roll.
