Safety is critical to building AI applications, especially when dealing with rich media like images and video. That’s why today we’re releasing sieve/text-moderation and sieve/visual-moderation: two pipelines that enable quality content moderation for digital platforms dealing with text, image, audio, and video.

We’ve built these pipelines in partnership with some of the largest video streaming, online shopping, and creative AI platforms in the world — all of whom wanted a more modern, tunable, cost-effective solution for moderation.

Key Features

Our moderation pipelines are built using a combination fine-tuned multi-modal LLMs, embedding models, and traditional classification systems — to ensure high quality and low computational cost. Some key features this enables include:

• Broad Class Support: 20+ text classes and 40+ visual classes that cover a wide variety of moderation use cases.
• Developer-friendly: Single API call integration with pay-as-you-go, self-serve access.
• Highly Tunable Settings: Tweak processing FPS when processing video, enable additional filtering, add custom classes, and more.
• Cost-effective: $0.50 / 1M text characters and $0.95 / 1K images (~40% cheaper with batch processing enabled).

Performance

We evaluated our content moderation system against an expert-labeled dataset of 400,000 text characters (10,000 text samples) and 10,000 images equally balanced in classes across social media and live streaming content. We selecting a standardized set of classes to compare Sieve and other moderation solutions against our expert-labeled dataset.

Text Evaluation

Class	Sieve	Provider 1	Provider 2	Provider 3
Sexual Content	0.962	0.943	0.924	0.911
Violence	0.978	0.954	0.947	0.936
Bullying	0.965	0.932	0.919	0.908
Hate Speech	0.973	0.949	0.934	0.919
Spam	0.921	0.928	0.925	0.919
Drugs	0.967	0.938	0.917	0.905
Child Exploitation	0.982	0.957	0.946	0.933
Child Safety	0.981	0.954	0.942	0.931
Gibberish	0.916	0.923	0.920	0.914
Personal Information	0.968	0.946	0.932	0.918
Promotional Content	0.964	0.941	0.927	0.915
Weapons	0.911	0.918	0.916	0.909

Visual Evaluation

Class	Sieve	Provider 1	Provider 2	Provider 3
Nudity & Sexuality	0.947	0.932	0.878	0.901
Violence & Gore	0.963	0.943	0.961	0.959
Substance Use	0.928	0.853	0.894	0.919
Offensive Behavior	0.961	0.953	0.909	0.938
Child Safety	0.972	0.945	0.911	0.894
Copyright	0.885	0.912	0.851	0.924
Virtual Content	0.912	0.913	0.891	0.813

Cost

	Sieve	Provider 1	Provider 2	Provider 3
Text Moderation	$0.50 / 1M chars	Contact Sales	$0.50 / 1M chars	$10 / 1M chars
Visual Moderation	$0.95 / 1K images ($0.58 / 1K images batched)	Contact Sales	$1.50 / 1K images	$1.00 / 1K images

Conclusion

You can try sieve/text-moderation and sieve/visual-moderation by creating an account today. We’re excited to keep improving on this work as our customers continue deeply integrating AI into their products with safety in mind.