We corrected how your product photo is presented to the image model. The picture you select was being passed as supporting material rather than as the subject to work from, which put it at odds with the instruction the model received. The result was inconsistent: whatever the model could copy straight from the visible pixels came out fine, while anything it had to infer — the angle, the proportions, the parts not directly in view — drifted away from your actual product.
Packshots and photo shoots now present your photo as the subject of the generation. In our measurements the same request, with the same photo and the same instruction, went from almost never matching the product to matching it in the large majority of attempts.
Two related changes come with it. When you supply additional reference images, a job that cannot load one of them now reports an error instead of quietly generating without it — you see that something went wrong rather than receiving a picture that silently ignored part of your input. And a limit that rejected perfectly valid jobs, such as a packshot accompanied by four references, has been raised to what the model actually supports.