NSFW Filter API for Image Moderation

The NSFW Filter is built around an asynchronous inference workflow that works naturally in production applications. Submit one publicly accessible image_url to create a task, receive the task ID, and then monitor the task with GET /v1/inferences/{task_id}. You can also provide a callback URL to receive an HTTPS notification after the task reaches a terminal state. This design gives your product time to manage processing, store task information, update dashboards, and apply moderation actions only when a completed result is available. It is a strong fit for AI image generators, content platforms, creative asset systems, marketplaces, and any service that needs structured image moderation without interrupting the broader user flow.

A completed NSFW Filter task returns clear, machine-readable moderation data. The result includes a flagged boolean that states whether the model flagged the image using your selected settings. It also returns category labels in two groups: nsfw and special_care. This makes it easier to create consistent workflow logic, such as sending flagged content to a review queue, delaying publication, requiring a second approval, adding an age gate, or applying your platform’s own policy rules. An empty category array means no labels were reported for that category group during the task; it does not guarantee that an image is universally safe, appropriate in every context, or fully compliant with every policy.

Every product has a different definition of acceptable visual content. The NSFW Filter gives you controls for adapting model behavior to your own moderation posture. threshold_offset is used to tune normal NSFW matching behavior. It defaults to 0.02, accepts values from 0 to 0.1 in 0.01 increments, and becomes less strict at higher values. A less strict setting may reduce the number of flags, but it can also increase the chance that potentially unsafe images are not flagged. strict_special_care is enabled by default. When enabled, special-care matches strengthen normal NSFW concept matching; when disabled, special-care labels can still be returned, but they do not strengthen normal NSFW matching or guarantee that content will pass.

The most dependable NSFW Filter workflow pairs automated detection with thoughtful human oversight. Use a flagged result as a signal to pause publication, create a moderation ticket, notify an operator, request a replacement, or send an image into a human-review queue. Use a non-flagged result as one useful input, rather than a final legal, safety, or contextual decision. Model outputs may vary with image quality, visual style, context, threshold settings, and the content standards your organization applies. Build clear escalation rules for uncertain images, retain audit information where appropriate, verify callback authenticity before processing webhook events, and ensure that your human moderators can make the final call when content is ambiguous or high impact.