About 9 minutes to read.
What is up people!
Imagine using AI to make an ad for a ceramic mug you sell. The original photo is ordinary: a green mug on a kitchen counter, with a squared handle and a small maker’s mark near the base. The generated version puts it beside a sunlit window. The shadows are beautiful, the composition looks expensive, and the handle has become round.
You now have a better picture of a different mug.
That is an easy mistake to overlook when the lighting is the first thing that catches your attention. But someone choosing a product from the image is also choosing its shape, finish, and details. A convincing scene can make an invented feature look like something you can buy.
Google’s Nano Banana 2.1 update, released last week, promises improvements in visual design, editing selected areas, and keeping subjects consistent. Those are useful directions for image generation. They also make this a good moment to get more precise about what counts as a successful image.
For work involving a real product, start by deciding which details must survive. Then judge whether AI has made that product look good. This gives you room to experiment with the setting without accidentally approving a redesign.
📌 TL;DR
Define what must stay true. A product reference establishes the object; a style reference suggests how to present it. Tell AI which role each image serves.
Check identity before aesthetics. Shape, markings, parts, and written claims need their own inspection. A favorable overall impression cannot settle those questions.
Choose a repair that fits the error. A targeted edit may help, but an exact product photo may be the better starting point when repeated generations keep changing the item.
Decide what AI is allowed to invent
The mug example is hypothetical, but the decision it creates is practical. If you are developing a new mug design, a rounded handle might be an interesting suggestion. If you are advertising stock already on the shelf, it is a mistake. The same image can be useful in one job and unsuitable for another.
Decide which job you are doing before asking for a more polished result. Words such as “premium,” “cinematic,” or “beautiful” describe a treatment. They do not tell the model whether it can reinterpret the object itself. A prompt can ask for attractive styling while leaving the most important restriction unstated.
For an existing product, make a short distinction in the brief. The mug’s silhouette, squared handle, glaze, and maker’s mark belong to the product. The table, background, and arrangement of props are choices you might allow the system to explore. If a change would make a customer expect a feature the real mug lacks, put that detail in the fixed group.
This does not mean every pixel has to remain identical. A different light changes reflections; a different camera angle changes the visible outline. The useful question is whether the image remains consistent with the physical object. A bright reflection on the glaze is different from turning a matte surface into a mirror finish.
Reference images need equally clear roles. Suppose you supply your mug photo alongside an inspiration image of a luxury perfume campaign. Without further direction, you are asking the system to reconcile two very different objects. Explain that the mug photo supplies the product and the campaign supplies only the lighting or mood. The perfume bottle is not a second candidate for what belongs on the table.
Use a reference that actually shows the details you want preserved. If the maker’s mark is unreadable, include a clear close-up. If the handle is hidden, take another photo. Asking for precision from an ambiguous source creates a problem before generation even begins.
There is plenty of creative space left after those decisions. The same mug could sit beside a rainy window, on a breakfast table, or against a simple colored backdrop. You can explore different moods and compositions while holding the product constant. That is a useful creative brief because it tells you which changes to welcome and which to reject.
Inspect the product before choosing the picture
Google’s October 6 model card separates overall image preference from categories such as product consistency and editing. It also lists remaining problems, including small text, partial instruction following during selected-area edits, and occasional confusion about spatial relationships. These are Google’s evaluations and limitations, not results from a test conducted for this newsletter.
That separation is worth carrying into your own review. When choosing between images, first check the details that could disqualify one. Only then compare which of the acceptable images has the stronger composition. Otherwise, the picture you most want to use can become the picture whose errors you keep excusing.
An older research project offers a concrete way to think about this. In the 2023 GenEval study, researchers evaluated generated scenes through individual properties such as object count, color, and position. Their human study collected 6,000 annotations across 1,200 images. The method could identify a specific mismatch instead of reducing the whole picture to one impression.
GenEval studied text-to-image composition, not whether a modern editor preserves an exact product photograph, and it did not test Nano Banana 2.1. The transferable idea is to ask separate, answerable questions about the image. The mug’s handle is squared or it isn’t; that question deserves an answer before the lighting wins your vote.
Keep the original reference open beside the generated image. Compare the silhouette, the handle’s attachment points, the mark, and any visible surface pattern. Zoom in enough to read actual lettering rather than deciding that the lettering looks plausible. If the image contains a price, date, or quantity, compare it with the approved text character by character.
Then inspect the whole scene. A picture of one mug beside a matching saucer can imply that the saucer comes with it. A hand holding a miniature-looking mug can create a misleading sense of size. The object may be recognizable while the surrounding composition changes what the viewer thinks is included or how it will feel to use.
Some discrepancies will be hard to judge. A handle can look narrower because of perspective, and a green glaze can look different under warm light. When you cannot tell whether the image is faithful, use another reference or request the original viewing angle. Uncertainty is a reason to get a clearer comparison, rather than declare the image accurate because no obvious defect jumps out.
After that factual pass, give the design its own attention. Does the mug stand out? Is the background helping the mood or competing with the product? Can someone understand the image when it appears small on a phone? An accurate image can be dull, cluttered, or hard to read. Passing the first check gives you a set of honest options to make more appealing.
Repair the detail without losing the version that worked
Suppose the generated scene is strong except for the rounded handle. “Make it more accurate” leaves too much open. A more useful correction names the mismatch and points back to the reference: restore the squared handle, retain its original attachment points, and preserve the accepted background and framing.
When your image tool supports selecting an area to edit, use that control to indicate the handle. Google calls this kind of capability mask-based editing: the selected region communicates where the change should happen. Its improvement claims make it worth exploring, but selection should be treated as an instruction rather than a guarantee that everything outside it stays untouched.
Keep the accepted version before making the edit, and compare the new image with both that version and the original product. The first comparison catches unwanted changes to the scene; the second checks whether the repair represents the mug correctly. Fixing the handle is not progress if the maker’s mark disappears in the process.
It can be tempting to keep requesting one more correction because so much of the picture already works. The cost is that you may spend the afternoon protecting details that were present in the photograph you started with. Decide whether another generation is likely to solve the remaining problem, or whether you need a different production method.
For an exact listing image, an ordinary product photograph with careful lighting and conventional retouching may be the sensible choice. You can also use AI to explore a scene concept, then reproduce that direction with the real object. The concept earns its value by helping you choose the shot; it does not have to become the published shot itself.
Be especially cautious about requesting a view your references do not show. If the model turns the mug around, it must produce a view of its unseen side. That view may be plausible without being correct. Supply the missing view or keep the camera position within what you can verify.
Creative illustration has a different standard. An obviously imaginary mug floating through space can be exactly what an editorial image needs. A new handle design can be useful during product development. Label concepts according to their purpose, and give imagination more room when the image is not claiming to document an existing item.
The decision is about the promise the finished image makes. Once that promise is clear, you can choose how much freedom to give the model without treating every deviation as either a disaster or an artistic improvement.
Also this week
Claude Haiku 5.5 arrived for smaller, repeated jobs. Anthropic’s October 7 announcement positions it for tasks such as summaries, classification, and quick supporting work alongside larger models. The company also says its larger models remain better choices for complex coding. If you run repeated tasks, a narrow job with an answer you can check is a sensible place to compare it with your current option. Lower API prices do not by themselves mean your chat subscription costs less.
Claude can work inside Google Docs, Sheets, and Slides. The new Google Workspace add-on is in public beta on paid Claude plans. It can work with an open file and propose or apply changes; the default “Ask before edits” mode presents approval cards. For a first try, use a copy of a document and inspect a small requested change in place. Availability and organizational controls apply, so check your setup before planning around it.
Try this on one product image
The following is a suggested exercise, not a workflow tested for this issue. Use a product you own or have permission to depict, with reference photos you are allowed to upload. Choose an image you can inspect yourself before it goes anywhere public.
Start with a clear source. Pick a photo where the identifying details are visible. Keep it unchanged, and add a close-up only if it resolves a detail the main photo cannot show. A pile of near-duplicate references is less useful than one that answers the missing question.
Write the fixed details in ordinary language. Name the specific shape, parts, markings, and finish that define the object. Then say what may change. For the imaginary mug, the setting can change while the squared handle and maker’s mark remain fixed.
Ask for one new scene from the same view. Keeping the angle familiar makes the first result easier to inspect. Try a different setting before asking for dramatic camera movement, elaborate props, and several products together. Add complexity after the simpler version works.
Review the result against the reference. Check the product details, all visible words, and what the surrounding scene implies. If you ask AI to help spot discrepancies, use its observations as leads to examine yourself. An automated assurance that everything matches is not a substitute for seeing the match.
Choose the next action from the error. A single clear mismatch may justify a targeted edit. An unseen detail needs another reference. Repeated changes to the object may justify returning to real photography. Once an image passes, inspect the final exported file at the size and crop people will actually see.
Here is a starting request you can adapt to your own object:
Create a staged promotional image using the attached product photo as the identity reference. Keep the product’s shape, proportions, visible parts, surface finish, and markings faithful to that photo. Preserve the original viewing angle. Change only the surrounding setting to a simple breakfast table beside a window. Do not add lettering, accessories, or product features. If a required detail is unclear in the reference, ask for a clearer photo before generating. The result should look composed and appealing while showing the same product.
Before sending it, replace the generic preservation language with the details you actually care about. For the mug, explicitly name the squared handle and small mark near the base. The prompt is a brief for the model, and your original photo remains the source for checking whether it followed that brief.
🧰 Tools to try
Nano Banana 2.1, where you already have access. Google’s official model documentation describes image generation and conversational editing with image references. Check which model your chosen product actually uses; a familiar app name alone does not identify the underlying model. There is no need to enable API billing to begin inspecting images made through a tool you already use.
Your existing photo editor. Keep it available for tasks where the real object needs to remain intact. Cropping, exposure adjustments, and conventional retouching may be sufficient for a usable photograph. Choose the method around the image’s purpose, rather than making generation a requirement for every visual.
For the next product image you make, leave the reference beside the final export until the review is finished. Imagine the person ordering the mug and then lifting the real one out of the box. Check the handle they would touch, the finish they would see, and the mark they would recognize. Once those details agree, choose the image that makes them want to pick it up.
P.S. An AI wrote these sentences using this week’s primary sources and earlier image-evaluation research. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.