The follow-up to Why AI image edits lose the part you liked.
The move is simple. Once you’ve seen it, you stop gambling.
The move
-
Get one good generation. Roll the prompt until something in the image is right. The face. The pose. The lighting. The mood. Doesn’t have to be most of it. One thing.
-
Stop rerolling. This is the hard part, because the rest of the image is wrong and your reflex is to hit generate again. Don’t. The next roll will not give you the good part back.
-
Put it on the canvas. Drop the image into Hyperdraw as a layer. Now it’s yours. It can’t change unless you change it.
-
Mark the part that’s good. Mentally or with a quick selection. That part is locked in. It is no longer up for negotiation with the model.
-
Work the bad parts locally.
- Mask the region you want changed. Regenerate only inside the mask. The good part stays exactly as it was.
- Erase what’s wrong (a stray hand, a weird artifact). Let the model fill just the hole.
- Extend the canvas if the framing is off. Generate only the new strip. The original is untouched.
- Fix small mistakes by hand - a stray pixel, a wrong colour. You don’t need a model for everything.
-
Repeat per region. Each fix is local. Each fix can fail without destroying the parts that already work. If a regenerated mask comes back worse, undo it; the rest of the image doesn’t care.
Why this works
Every step keeps something. Prompting kept nothing - every roll was all-or-nothing. Canvas work is all-or-part: a bad regeneration costs you the masked region, not the whole image.
That changes the math. With prompts, the cost of trying is “lose everything I had.” With a canvas, the cost of trying is “lose this patch.” You will try more things. You will get further. You will stop hoarding the one good roll out of fear of losing it, because you can’t lose it - it’s pinned to a layer.
The line, again
Prompting rerolls the whole bet. Canvas work lets you keep the win.