DALL·E 3 revision mode guide
Editing a DALL·E 3 image is really re-prompting: every follow-up message regenerates the picture from scratch, conditioned on what came before. That’s why a request to “make the car red” can also swap the background and change the lighting. The trick is phrasing that anchors the parts you like and isolates the one thing you’re changing. This guide builds those instructions for you from five common edit types.
How it works
You pick a change type — recolour, add, remove, reposition or restyle — and the tool drops your original prompt and the target element into a template that:
- Restates the original scene so the model re-anchors on it.
- Names the single change in unambiguous, visual language.
- Locks the rest with an explicit “keep everything else identical” clause.
That structure is what separates a clean one-element edit from a complete regeneration.
The five edit types and how to phrase them
Recolour — Specify the exact colour and the exact element. “Change only the car to cobalt blue; keep the background, lighting, and all other elements identical to the previous image.”
Add an element — Name where it goes and how large. “Add a red umbrella in the upper-left background; everything else stays exactly the same.”
Remove an element — Describe what fills the space, not just what goes. “Remove the lamppost and replace it with the same cobblestone street texture that surrounds it; leave the rest of the image unchanged.”
Reposition — Describe the new location precisely. “Move the cat from the left window ledge to the right window ledge; keep the cat’s appearance, the room, and all other elements identical.”
Restyle — Name the artistic style clearly. “Render the same scene in the style of a Japanese woodblock print; keep the composition, subjects, and layout identical.”
Why DALL·E 3 keeps drifting
DALL·E 3 regenerates the full image each time using a diffusion process that samples from its training distribution. Minor changes to the prompt can shift which region of that distribution the model samples from, changing elements you did not mention. The anchor-and-lock structure in this guide counteracts that drift, but it is not perfect — spatially precise edits (fixing a hand, changing a word in text) genuinely require the drag-to-select inpaint editor in ChatGPT.
Tips
- Always reply in the same conversation. Starting a fresh chat loses the image context and forces a full restart.
- Change one thing per turn. Bundling multiple changes produces unpredictable results — iterate one edit at a time and review before the next.
- Use the inpaint selector for spatial precision. When the change is location-specific (e.g. fix one hand), the drag-to-select editor beats any text instruction.
- Describe, don’t negate. “Remove the text” is weaker than “a clean, textless version of the same scene” — describe what you want to see, not what you want gone.
- Save good results immediately. The image URL in the API response expires; download the image before closing the conversation.