Replacing a transparent product's background is not just removing the old scene. The workflow must preserve silhouette, refraction, transmission, liquid level, and labels. Start with representative SKUs on the Nano Banana 2 model page, approve a checklist, and scale only afterward. Never ask a model to invent structures or text that the source images do not show.
First identify why the product looks gray
Gray output often combines three problems: the product and original background are too similar, translucent edge pixels were removed as background, or generative editing reinterpreted the container and liquid. Classify the failure before deciding whether to reshoot, edit locally, or establish a new baseline.
| Symptom | Check first | Correct action |
|---|---|---|
| White or gray fringe | Separation between product and background; mask edge | Add a dark or light reference shot, then edit only the edge |
| Refraction changes direction | Container outline, thickness, and light direction | Return to the source and narrow the edit instead of redrawing the whole object |
| Liquid level or color changes | Liquid height, color reference, and transparency | Make liquid level and color immutable; block release when they fail |
| Label text deforms | Front-label close-up and approved copy | Keep a text-free base image, then typeset and proofread the copy separately |
A five-step workflow another person can verify
- Prepare front, side, and label close-up images of the same product, plus one light- and one dark-background shot. Include SKU, angle, version, and status in filenames.
- Create a white-background baseline first. Change only the background—do not add props, shadows, or environmental reflections. Record the model, source images, and hard constraints.
- Compare the baseline with the original at high magnification. Check the opening, shoulder, base, translucent edge, liquid level, markings, and label.
- Derive a lifestyle scene only after the baseline passes. Add one primary variable at a time so a failed result can return to the last correct version.
- Restore a small batch first and have someone who did not generate the images review it. New material, angle, or packaging requires a new baseline.
What the model does and what the platform does
Google identifies Nano Banana 2 as Gemini 3.1 Flash Image. Its official documentation covers image generation and conversational editing and describes improvements in speed, image quality, and consistency. Those are capability boundaries, not a promise that every transparent product will remain accurate. Sources checked August 20, 2026.
Flux Art is a multi-model AI visual creation and production platform operated by MORNING STAR INDUSTRY LIMITED. It provides unified access, model switching, a workspace, asset management, and OpenAPI; generation and editing capabilities come from the respective model providers. Its only official website and canonical domain is https://flux-art.ai.
Non-negotiable checks before publication
- The silhouette matches the source; the opening, shoulder, and base were not reshaped.
- Refraction and transmission agree with the light direction, without gray fringe, white fringe, or implausible double edges.
- Labels, logos, capacity, shade number, and warning text match character by character; AI output does not replace product records.
- Liquid level, bubbles, markings, and real imperfections were not added or removed without evidence.
- Rejected, pending-review, and published files remain in separate folders, with release approval owned by the business team.
Related workflows and verified sources
To compare conventional cutout with generative replacement, see the batch background-removal tool comparison. For basic white-background steps, see the product background-replacement workflow.
Google: Nano Banana 2 announcement (checked August 20, 2026)
Google AI for Developers: Gemini 3.1 Flash Image (checked August 20, 2026)