Industry News
Using GPT Image 2.6 to generate product images for an entire season results in bottle caps appearing more frequently than the background in these images.
Those who create product images, upon reading the release announcement for GPT Image 2.6, would likely skip the GIF of the dancing cat and go straight to the next point: During multiple rounds of editing, the characters, product design elements, and brand visuals remain more consistent and stable.
If this statement is true, one person would be able to complete the production of all the promotional images for an entire season. However, if it isn’t true, the savings made on shooting costs would be completely offset by the resulting returns and customer complaints.
On November 4th, OpenAI launched Images 2.6, which divided its API into two separate models: Flare, designed for speed; and Sunburst, aimed at providing detailed results and ensuring stability during editing. ChatGPT also added a new feature that allows users to save reference images. Once characters, pets, or other objects are saved, there is no need to re-upload them repeatedly.
The handle was moved to a different position, and no one reported any issues.
On the day of release, some developers used versions 2.5 and 2.6 to create the same set of promotional materials for an e-commerce campaign: one thumbnail with dimensions of 1536×864, one product image measuring 1024×1024, as well as images that had undergone adjustments in color and price. Out of these six static images, the text in all 24 designated areas was correctly placed, and the dimensions of all images were accurate.
The problem lies with the product itself.
In version 2.5, the thumbnail shows a side handle, while the product image displays a top handle; in version 2.6, the design is exactly reversed. The color of the letters also varies between different versions, and there is an additional white mark on the badge that no one requested. These errors do not trigger any warning messages and cannot be detected by text checks either; they can only be identified by comparing the images with the original design.
Since each item is only tested once, it is impossible to determine which generation is superior. However, this experiment does highlight the key aspects that need attention during inspection: whether the background design looks good can be immediately apparent, whereas the fact that the bottle cap is 2 millimeters taller can only be noticed upon closer examination. Once customers receive their products, they will be able to recognize these differences at a glance.
Which tasks can be safely entrusted to it?
Based on the actual test results reported by various companies, I would categorize them in the following way:
- Rest assured: We can change the background, create different seasonal atmospheres, recreate real-life scenarios, and produce concept drafts for event posters.
- When examining these images—those showing combinations of multiple items or the front side with packaging information—it is essential to verify each one individually.
- Don’t click on those areas: the specification comparison charts on the details page, the close-ups of product labels, and any sections labeled “photographed in real life”.
OpenAI’s own recommendations tend to be conservative: the edited images must remain completely identical to the original ones before being merged back together. SVG exports are also of little use when it comes to creating packaging materials, as they can only generate new graphics based on textual descriptions and cannot convert existing logos into vector format.
A five-round test that can be copied verbatim
Before putting the new model into production, test it using your own products for about half an hour – that’s sufficient.
Use a product image that has been approved by the customer as a reference. Open Sunburst and select the medium or high quality setting. Make five consecutive changes, each involving only one modification: change the desktop background, the lighting effects, the seasonal props, the background color, or add an accessory. In each round, compare the resulting design with the reference image and note at which round the bottle cap, label, or proportions began to deviate from the original design.
If the fifth item still matches the first one, then this product category can be considered successful.
The phrase that definitely deserves the most emphasis in these instructions is “Do not touch anything.” For example, in this sentence: “Place this table lamp on the walnut-colored side table in a room with a warm and minimalist design; keep its original shape as well as its sage-green glaze intact.” The first half describes the scene, while the second half explicitly warns against making any changes. Many people only write the first half, resulting in the model automatically optimizing even the lampshade without anyone realizing it.
How can a team that doesn’t write code get started with this approach?
When using the above process via APIs, one must handle reference images, masks, and usage statistics independently; operational staff generally have no involvement in this process. I would recommend using PixPix’s GPT Image 2.6 platform directly. Upon logging in, Sunburst will be selected by default. You can then upload the reference image and clearly specify which aspects of the product’s design—such as its shape, color, and composition—should be retained, as well as the areas that need to be modified.
Its tutorial also involves making adjustments one by one, and only moving on to the next step after observing the results. This aligns perfectly with the rhythm of the five rounds of testing. There is an example on the page that you can refer to: Take the same amber-colored perfume and place it in three different settings – a dark photography studio, a scene with plant projections, and a surface made of stone. You can try doing the same with your own product to see if there are any changes to the appearance of the bottle.
For those categories that have passed the testing, you can start generating scenario images in bulk. For the PixPix section, as well as the E-commerce Tools and A+ Content categories, you can continue to create materials on their respective detail pages. The specific amount of credits required for these tasks will be indicated on the workbench.
The final step still requires manual intervention. Before releasing the product, it is necessary to compare the final image with the actual object – this step cannot be skipped at present.