You hit the roll. The image is gorgeous. The mood, the lighting, the character, the composition - somehow, all four landed at once.
The hand is broken.
You know what happens next. You’ll re-prompt to fix the hand. The next roll will have a great hand and a different character. You’ll re-prompt again. The mood is gone. You’ll spend forty more rolls trying to get back to what you had on roll one, and you won’t.
This is a known wound - covered in Why AI image edits lose the part you liked and the fix in Fix part of an AI image without regenerating it. This article is the same move applied specifically to Midjourney.
Pin the win on a layer
The moment a Midjourney roll has the mood, composition, character, or lighting you wanted - stop. Save it. Open Hyperdraw. Import.
Now it’s on a layer. That’s the whole game. The dice are no longer involved. Style tokens, seeds, aspect-ratio quirks, whatever Midjourney thought “cinematic” meant today - none of it can change those pixels anymore. The win is pinned, in the literal sense: it’s a layer in a file you control.
Pick the one thing
Don’t try to fix everything. Pick one:
- a broken hand
- nonsense lettering on a sign
- a too-busy background
- a costume shape that’s almost right
- a subject crowded against the edge
- a color accent in the wrong place
One region, one problem, one pass.
The broken hand: layers first, model second
Take the most common case. The reliable move doesn’t start with another generation - it starts with composition.
- Import the Midjourney roll → layer 1. This is your win.
- Eraser the broken fingers on layer 1. You’ve cut a transparent hole exactly where the problem was. Everything else - face, mood, lighting, every pixel that worked - is intact.
- Add Layer above. On this layer, draw the hand silhouette you actually wanted. Even a rough closed fist or open palm in the right color is often enough to read correctly.
- If hand-painting isn’t enough, now press Dream. The model only fills the transparent hole on the layer below; layer 2’s silhouette acts as a hint. The good Midjourney pixels around the hole are not part of the conversation.
- Bad result? Hide or delete the top layer. The win is still there, on layer 1, exactly the way Midjourney rolled it.
Compare this to re-prompting Midjourney: the cost of trying a fix used to be “lose the whole image.” Now it’s “hide a layer.” You’ll try more things. You’ll get further.
Use Image Strength to keep the model honest
When you do reach for Dream over a Midjourney import, some engines expose an Image Strength control. Push it higher than you’d push it for an original sketch - you want the model to respect the colors, shapes, and lighting that already work, not reinterpret them.
Lower strength = more freedom to invent. Higher strength = more loyalty to the imported image. For “fix the hand on this Midjourney roll,” loyal is what you want.
Liquify before regenerating
Some Midjourney problems aren’t generation problems. The shoulder line is a little off. The eye spacing is uncanny. The prop silhouette wants to be a hair longer.
Try Liquify first, on a duplicate of the layer if you want to be extra safe. Push the shape by hand. Don’t regenerate unless the image actually needs new pixels. A successful Liquify costs you zero generations and zero risk to the win.
Compose two good rolls instead of chasing one perfect one
You have two Midjourney rolls. One has the perfect face. One has the perfect background. Import both onto separate layers. Eraser the bad parts of the top one. The good parts of the bottom one show through.
That’s a finished image, made from two partial wins, in about a minute, with no further prompting. Midjourney rarely gives you that on a single roll. A canvas with layers gives it to you reliably.
Next time a Midjourney roll lands almost perfectly, save it, open Hyperdraw, Import, erase the bad part, and paint or Dream into the hole on a new layer above. Don’t burn the win to fix one finger.