
Cloudinary Generative Remove does not treat every nearby mark as part of a named object. Its documented defaults leave that object's shadows and reflections, and prompt text controls whether removal targets one instance, multiple objects or lettering on a particular object. Specify the intended scope, then review the whole result for changes you did not request.
Key Takeaways
- Object removal leaves shadows and reflections by default.
- Multiple prompts can broaden removal beyond one instance.
- Object-specific text removal differs from asking to remove all text.
Write what must disappear and what must remain
Keep the original image and identify the authorized change. A product distraction, a cast shadow and a neighboring printed label are different image regions. State which belong to the task before choosing the effect's options.
The Cloudinary Generative Remove guide documents natural-language prompts and region-based targeting. Its examples show that a more specific object description can select a different target from a broad category.
The controlled-selection guide explains the general need to protect useful background evidence. Its painted-mask controls are not the same as Cloudinary's prompt grammar.
Check the scope of instances and text
Cloudinary documents multiple_true for removing multiple occurrences of an item. When more than one prompt is supplied, it says all instances of each specified item are removed regardless of the multiple setting.
That distinction matters if a scene contains several similar products or signs. Do not add a list of prompts as though it merely makes one target description more detailed.
For lettering, prompt_text requests text removal generally. The documented text:<object> form restricts the request to text on the specified object, such as a mobile screen. Preserve other signs and labels as explicit acceptance checks rather than assuming natural-language scope is perfectly reliable.
Decide whether shadows and reflections belong to removal
The guide says shadows and reflections cast by the prompted object remain by default. It provides remove-shadow set to true when those should also disappear.
For a proposed product cleanup, compare the original cast shadow, contact area and any reflection before choosing that option. The double-shadow guide explains why dark shapes need a lighting interpretation, not automatic deletion. The character-replacement shadow guide concerns a different moving-image task and does not supply this still-image API's controls.
Reject collateral or invented evidence
Inspect the removed region, background lines, retained labels and nearby subjects. Cloudinary warns that generated results may be inaccurate and that the effect works best on simple, visible objects. It also advises against attempting faces or hands.
Generated fill is a proposed reconstruction, not recovered factual content hidden behind the original object. Accept only an output whose removal scope and remaining scene meet the brief; retain the untouched source for comparison. No generation or object-detection result is claimed here.
Related tasks
- Remove objects from a photo with a controlled AI selection
- Why Does My Product Photo Have Two Conflicting Shadows?
- Why Is the Old Dancer’s Shadow Still Visible After Character Replacement?
Sources and scope
- Cloudinary: Generative remove: Prompt scope, multiple instances, regions, remove-shadow and text-based object targeting control requested removal. The effect accepts non-transparent images, excludes animated/fetched inputs and has documented regional and processing limits; outputs are generated rather than recovered factual pixels.
Documentation-based guidance from the cited official pages captured on October 25, 2026. Scenarios and acceptance procedures are proposed, not executed tests. No account operation, measured demand, AI-citation frequency, processing benchmark or search-indexing outcome is claimed.
Prepared by Imagild Editorial, the company editorial team responsible for these guides. AI assisted research and drafting. This is documentation-based guidance with proposed examples, not a claim of hands-on testing or an independent product review. Report a correction.
Cover: AI-generated conceptual editorial illustration; not a real application screenshot, observed output or before-and-after test.