Product mockup prompt with set lighting and camera angle

Ask for "a nice photo of a black mug" and the model picks the handle side, the finish and the light direction again for every image. Four results then differ in ways you never chose, so you can't compare them as drafts. The example product below is fictional: a matte black ceramic mug, 350 ml, handle on the right.

[1] Write the product spec before the prompt. Keep only what the photo can show. A value the camera can't see, like 350 ml, becomes a size cue: about 9 cm tall, fits in one hand.

[2] The product mockup template in Google's Gemini docs (Image generation) has five slots: the product, the surface it sits on, the lighting setup and its purpose, the camera angle and what it shows, and the detail in sharp focus. Fill them from step 1.

A high-resolution, studio-lit product photograph of a {matte black ceramic mug, cylindrical, about 9 cm tall, unglazed fine sand-like texture, handle on the right side of the frame}on a {light natural oak tabletop}.The lighting is a {single large softbox from the left at 45 degrees, with a soft white bounce card on the right} to {show the texture without any glossy reflections}.The camera angle is a {slightly elevated front view, about 15 degrees above the rim} to showcase {the full silhouette and the handle}.Ultra-realistic, with sharp focus on {the surface texture of the mug body}.Square 1:1 image. Plain, uncluttered background.

Braces mark the slots; remove them before sending. A few words were added to the documented template for this product.

[3] For a black matte product, spend your words on lighting. Black absorbs light and the faces go flat; matte means no highlights to show the shape. The bounce card opposite the softbox keeps the right side from merging into the background. Don't stop at "studio lighting" — state how many lights, from where, and how large. The Vertex AI best practices page for Gemini images also says to describe composition with photography and film terms such as angle, lens and perspective.

[4] Turn the prompt into a checklist and mark each result O or X. Judge the light by the shadow: with the softbox at 45 degrees on the left, the shadow falls behind the mug to the right.

Drop any image with an X. If all four fail the same row, that sentence is too weak: rewrite that one line and generate again. Change several lines at once and you won't know which change worked.

[5] Check the ratio in pixels, not by eye. The files below are gray test images made with convert to show the check — they are not generated results. Anything more than 1% away from 1:1 is flagged.

$ identify -format "%f %w %h\n" mug-*.png | awk '{r=$2/$3; printf "%s %sx%s %s\n",$1,$2,$3,(r>0.99&&r<1.01)?"OK":"NOT 1:1"}'mug-1.png 1024x1024 OKmug-2.png 1024x1024 OKmug-3.png 1024x1024 OKmug-4.png 1200x896 NOT 1:1

[6] To learn which line drives the result, build a comparison set: drop one line from the base prompt, generate the same number of images, and fill the step 4 table again.

The right column is what gets left to the model, not an observed result. Which removal piles up X marks is something only your own runs will show.