Conversational Image Editing Loop
Supply an image, describe one change, inspect it, and iterate in plain language
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 94%
The loop treats an existing image as the starting state and natural language as the editing control. The user supplies one or more images, describes a concrete transformation, and inspects the generated result. Rather than packing every preference into one prompt, the user identifies the most important discrepancy and requests another targeted revision. The same mechanism can add or remove visual elements, colorize a black-and-white image, or fuse material from multiple images. This dramatically lowers the tool-learning barrier, but it does not remove the need to verify identities, text, factual details, ownership rights, and production-level precision before publishing the result.
Origin
Extracted from Marketing Against The Grain
Core principles
- 01Begin from the visual artifact rather than recreating it
- 02Express desired changes in ordinary language
- 03Make one inspectable transformation at a time
- 04Use multimodal composition when several images contribute to the result
How to run it
- 1
Select the source
Choose a clear image that contains the subject and composition worth preserving.
Pro tip Use the highest-quality lawful source available.
Watch out Confirm that you have permission to edit and publish the image.
- 2
Request one concrete change
Describe the subject, location, material, color, or style transformation in plain language.
Pro tip State what should remain unchanged as well as what should change.
Watch out Vague prompts invite unnecessary alterations.
- 3
Inspect the output
Compare the result against the original and the intended change.
Pro tip Check faces, hands, text, logos, edges, and repeated patterns closely.
Watch out A visually convincing result may contain fabricated details.
- 4
Refine the largest discrepancy
Ask for a focused correction rather than restarting with an overloaded prompt.
Pro tip Change one major variable per iteration when possible.
Watch out Repeated editing can gradually distort features meant to remain stable.
- 5
Combine sources when needed
Provide a second image when the desired composition requires elements from both.
Pro tip Explain which image supplies each subject or style element.
Watch out Image fusion can create misleading representations of real people or events.
- 6
Perform final checks
Verify visual accuracy, resolution, rights, disclosures, and suitability for the intended use.
Pro tip Use professional tools for any final precision corrections.
Watch out Do not present synthetic documentary imagery as authentic evidence.
In the wild
A marketer supplies a photograph of croissants and asks the model to add a strawberry drizzle while preserving the pastries, plate, and lighting. After inspecting the first output, the marketer asks for a thinner drizzle and fewer changes to the background.
→ A usable concept image is produced through a short conversational iteration.
A creator supplies a black-and-white family photograph and requests restrained, historically plausible colorization. The creator inspects faces and clothing carefully and retains the original alongside the generated version.
→ The image becomes easier to experience while its synthetic alteration remains clear.
Common mistakes
Overloading the first prompt
Requesting many simultaneous changes makes it difficult to diagnose why the result diverged from the intent.
Ignoring unintended edits
The model may alter faces, text, backgrounds, or composition even when the request targets one small element.
Skipping authenticity checks
Edited images can create false evidence or misrepresent people and events if their synthetic nature is hidden.
Is it for you?
Best for
Rapid concepting, simple corrections, colorization, compositing, and iterative creative exploration.
Not ideal for
Precision production work requiring pixel-perfect control, provable authenticity, or complex rights management.
From the transcript
“you can pass in an image and you can say, hey, update this image”
“You can take in two images and you can sort of fuse them together.”
“You can just do it with like a very, very simple text prompt.”
From the episode
Google's Secret AI Advantage (Why DeepMind Will Dominate)