You ask ChatGPT to fix the hand.
It returns the image with a different face. The hand is fine now. The character isn’t your character anymore. You ask it to put the original face back. The face comes back, sort of, but the hand is wrong again. The lighting drifted. The pose moved. You wanted one local fix; the model renegotiated the entire image.
Same story with Gemini. Same story with most chat-driven image tools. This isn’t a “you prompted it wrong” problem - it’s a contract problem. The chat surface only knows how to send the whole image back through the model. The pixels you liked are at risk every single message.
The fix is older than AI. It’s called layers.
Once it’s a layer, it’s yours
When a ChatGPT or Gemini result has the part you wanted, save it and Import it into Hyperdraw. It lands as its own layer. From this moment, those pixels cannot move unless you move them. No prompt, seed, model version, or context window can touch them.
That’s the difference between chat and a canvas. Chat sends the whole image back through the model on every turn. The canvas sends nothing unless you ask it to. The pixels you liked are off the table.
Erase a hole, don’t ask for a replacement
The reliable fix isn’t “mask the bad area and beg the model to inpaint it.” It’s:
- Import the chat result onto layer 1.
- Eraser the broken part. A bad hand, a glitched logo, a stray prop - gone. Now layer 1 has a transparent hole where the problem was.
- Add Layer above. On this new layer, paint the replacement - freely, at your speed, without the model in the loop. Even a rough shape works for many fixes.
- If you need the model’s help on the hole, now press Dream. It’s only generating into the transparent area you carved out; the rest of the image is below it on layer 1, untouched.
The whole image used to be at risk on every message. Now only the part you erased is in play, and the part you liked is safely below.
Compose, don’t renegotiate
Two chat results, each good at one thing? Import them onto separate layers, Eraser the parts you don’t want from the top one, and the bottom one shows through. That’s compositing - the same move every photo editor has had for thirty years. It works every time, no prompt required.
The trap of chat-driven image work is treating it as a single conversation about a single image. Hyperdraw lets you treat it as a pile of partial wins you composite by hand. The model contributes fragments; layers and erase decide which fragments live.
When you can’t phrase what’s wrong
Sometimes you can see the problem and can’t put it in words.
Hyperdraw has Text Tools and Vision / Describe modes for that bridge: describe the current image, rewrite a weak prompt into something the model handles better, or pull words out of a reference. Helpful when you hit the next wall - but the layer-and-erase loop above doesn’t need any of it.
A practical loop
- Generate or Import an image you mostly like → it’s on its own layer.
- Pick one thing to fix.
- Eraser that area on the import layer.
- Add Layer above. Paint the fix by hand, or Dream to fill the hole.
- Hate the result? Hide or delete the top layer. The original is still underneath, untouched.
- Repeat on the next area.
Every step preserves the pixels that already worked. The chat couldn’t do that. The canvas can.
The next time ChatGPT or Gemini gives you a result with one good thing, save it, drop it into hyperdraw.art via Import, erase the broken part, and paint the fix on a new layer above. Stop asking the chat to please not move everything else.