Gemini 4 Carbon vs Gemini 4 Argon for a Ceramics Cleanup Brief
Compare Argon and Carbon for catalog cleanup rules. Use ivory rims, translucent glaze and contact shadows to decide which pieces need individual review.
What's new on this page
- — Page published.

A batch of vintage ceramics can look consistent on the shelf and behave very differently when a background is removed. A white plate rim catches as much light as the backdrop. A clear-glazed bowl picks up the gray sweep along its edge. A cup casts a fine shadow that tells a buyer where its base touches the surface. One instruction to “remove everything pale” would damage all three.
Use Gemini 4 Argon to draft a routine background-cleanup brief once a photographer and catalog editor agree on the desired result. Use Gemini 4 Carbon to audit exceptions when several rules interact. This application of Argon and Carbon follows Google’s distinction between a general starting point and more involved reasoning. Both deliver written guidance in this exercise. The actual masks still need an image workflow and an operator who can inspect them.
Sort the pieces before writing a rule
Take a small sample from the batch. Pick one uncomplicated cup, one plate whose edge nearly disappears against the sweep, and one bowl with a broad band of reflected color. Look at each against a dark background and a light one. Save the original beside it. A dark preview can expose a chipped-looking white rim; a light preview may make the same error disappear.
Now give Argon three known instructions: remove the studio sweep, preserve every visible ceramic edge, and flag uncertain color instead of assigning a new glaze. Ask it to turn these into a short operator card with one example of what qualifies for routine processing. The straightforward cup is a good candidate if its full edge remains visible in both previews and the handle opening can be identified without guesswork.
Keep shadow treatment separate. A catalog that needs a pure cutout for packaging may require a transparent shadow layer; a listing intended for a white product grid may permit a natural contact shadow. State which output is actually ordered. A model cannot infer that delivery requirement from the bowl’s reflection.
Give Carbon a real exception to reconcile
Suppose the studio card also says “remove gray contamination.” The clear glaze on the bowl reflects the gray sweep. Removing every gray pixel would change the appearance of the ceramic, while leaving a gray fringe outside its physical edge could look like a failed mask. Send Carbon the rule, the exception and the intended use. Request an escalation note that distinguishes the outer silhouette from reflections inside it.
A usable note might read: “Preserve the interior glaze reflection. Inspect the band against both preview backgrounds. Remove only material outside the photographed rim; if the edge cannot be located from the original, flag this item for a second exposure.” That is an example instruction, not a finding from a model run. Try it on the actual image before applying it to any siblings in the batch.
What belongs in the reject log?
Record the item ID, the defect location, the two preview backgrounds and the action. For example: “Plate P12, rim at two o’clock: a notch appears on dark gray but is invisible on white. Return to original and restore the rim before approval.” A description such as “mask looks bad” forces the reviewer to rediscover the defect.
After the operator card is agreed, use the product background remover for the visual work. If the rule handles the uncomplicated cup and sends the bowl to a human with a specific question, it is doing its job. Keep the simpler drafting workflow for routine pieces; reserve the more involved review for exceptions that could change what a buyer thinks the item looks like.