Describe what should change, name the details that should survive, and decide how you will check both. An attractive image can miss a requested change or quietly replace something you meant to keep. A useful editing brief gives you a way to notice that before the next revision.

Start with a picture you want to edit and a short list of visible requirements. Keep the starting version available throughout the process: it is the reference for preservation.

A published edit with an unresolved instruction

In his 13 November 2025 experiments, Max Woolf starts with a generated skull-shaped pancake, topped with blueberries and maple syrup. His selected editing experiment asks for five changes:

  • Assign the strawberry to the socket labeled left.
  • Assign the blackberry to the socket labeled right.
  • Garnish the top with mint.
  • Use a chocolate-chip cookie as the support instead of the plate.
  • Populate the background with cheerful people.

Woolf reports that all five were implemented. He also describes blueberries being removed to accommodate mint and syrup pooling differently on the new support. Then an update qualifies the result: “left eye socket” and “right eye socket” depend on whose perspective you use. Read his original prompt, images and update.

That leaves a precise conclusion: the edits were reported as present, but correct fruit placement cannot be established against an undefined reference frame. The brief needs a direction convention before that criterion can receive a clear pass or fail. A revision could specify the viewer’s left edge of the picture, rather than the skull’s left. This diagnoses an ambiguity in the request before blaming the model.

Write what changes and what stays

An invariant is a detail you want an edit to preserve. For a product picture, it might be the logo, shape, camera angle or surrounding objects. Name the important details rather than relying entirely on “keep everything else the same.”

Suppose you want to recolor a mug in a product photo. A brief could read:

Change the red ceramic body of the mug to dark green. Keep its shape, handle, printed logo, position, camera angle and background. Preserve the logo’s lettering and colors. Allow the reflections on the recolored body to adjust to its new color. Add no objects or text.

You can turn the brief into checks: the body changed color; the handle retained its shape; the logo still has the same letters; the background retained its objects and arrangement.

Decide how much preservation the job needs. “The same subject and composition” allows more variation than “retain the exact lettering and outline.” A prompt expresses those requirements; compare the output to discover whether it met them. For details that must match exactly, a similar-looking substitute is still a mismatch.

Also resolve competing requirements. Replacing a background while insisting that every existing shadow remain identical may leave you asking for an incoherent scene. State which local effects may change to support the edit and which features are protected.

Give each reference a job

The starting image tells the editor what already exists. An additional reference can specify a different color, the appearance of a subject or a desired composition. Say which of those jobs it performs. “Use this picture” leaves open whether you want its object, lighting, background or style.

For the mug brief, a green color sample should supply color, while the product photo supplies shape and lettering. If you introduce a room photograph only to set the background, explicitly keep the product’s appearance tied to its original picture. These are instructions to check, not locks on the output.

Robin Rombach is among the authors of the FLUX.1 Kontext report. The team separates local editing, character reference, style reference and other tasks in its evaluation. Its discussion also documents missed instructions, identity loss and artifacts across repeated edits. That makes reference preservation a useful question to inspect on each revision, even when earlier examples look convincing. Read the original report’s evaluation and discussion.

Read a control as a tradeoff

Tim Brooks, Aleksander Holynski and Alexei Efros’s InstructPix2Pix project pairs an image with an editing instruction. Its published failure cases include unwanted excessive changes and difficulty isolating an object. Those failures show why checking the requested change alone is incomplete. The original project includes the examples and limitations.

Its implementation offers separate image and text guidance weights. The authors explain that stronger image guidance favors similarity to the input, while stronger text guidance favors the instruction. Their README’s tips discuss both insufficient change and too much change. Those are controls in that particular implementation, not settings every editor shares.

Guidance is a way of steering generation toward a condition. In diffusion models, an image emerges through repeated steps from noise; guidance changes the direction of those steps. Sander Dieleman explains this with diagrams in his geometry essay. Small changes to the steps can produce different final outcomes. If your workflow exposes a guidance setting, read what that implementation says it controls before treating a larger number as better quality. Open the original explanation.

Compare before asking for more

Check the result in two passes. First, examine each requested change. Then compare the protected details with the starting picture. In the mug example, a convincing green surface does not settle whether the logo was redrawn or the handle changed shape.

Record a specific outcome:

  • Met: the requested detail changed and its protected neighbors remained acceptable.
  • Mismatch: an instruction was missed or a protected detail changed.
  • Unresolved: the brief does not yet define what counts as correct.

Carry one brief through the edit

From a written image brief to one inspected revision.

  1. Step 1

    State the brief

    Name the requested change, the details to preserve and what each reference supplies.

    Condition Resolve directions and competing instructions before judging the output.

  2. Step 2

    Compare the output

    Check each requested change, then compare the protected details with the starting picture.

    Condition An attractive result can still miss an instruction or change something you wanted to keep.

  3. Step 3

    Make one revision

    Describe a specific remaining mismatch and keep the last acceptable version for comparison.

    Condition Recheck the whole brief after a revision; earlier successes may not survive another edit.

Accept an edit only against a clear brief. A precise prompt makes the comparison possible; it does not guarantee preservation.

Check both the changed and protected details after each revision.

Use that observation to write the next revision. “Restore the original handle shape; retain the current green body” gives you a narrower comparison than another request to improve the image. Afterward, check the whole brief again, including details that passed before. Keep the last acceptable version so a later repair does not erase your comparison point.

The same habit matters when a visual project has several parts. In his 1 August 2026 video-editor account, Pieter Levels describes adjusting a prompt and regenerating an individual clip in a timeline. This is his account of a commercial product workflow; the agent sidebar discussed in that post was a planned addition. The useful reading question is how an individual part can be revised while the surrounding edit remains available. Read his original account.

Choose the next reading task

For more published visual examples, use the creative AI reading selection. Choose a continuation according to the part of the process you want to understand:

  • Repeat the comparison across many requests: evaluate an AI application on your task, with examples and explicit failure criteria.
  • Inspect a local implementation: read what to check before downloading a local model, then the local AI selection. That article concerns language models; its model-versus-runtime distinction is the relevant continuation, rather than an image-editor setup guide.
  • Build the surrounding workflow: use coding-agent workflows and review for the software work around saving briefs, versions and comparisons.
  • Read the underlying model research: use LLM research for language-model foundations, or return to the image papers linked above for visual generation. Different model families require different explanations.