Image Moderation 2.0 detects risky image content across multiple scenarios and supports pay-as-you-go and resource plan billing.
1. Introduction to Image Moderation 2.0
Features
The Image Moderation 2.0 API detects image content that violates online regulations, disrupts platform governance, or degrades user experience. It supports 40+ content risk tags and 40+ risk control items. Based on the risk tags and confidence scores returned by the API, you can define custom review or governance actions tailored to your industry standards or platform content policies.
Image Moderation 2.0 also supports identifying whether images contain elements involving minors. Note that Content Moderation only detects the presence of minor-related elements and does not automatically determine whether to block such content based on your specific business requirements. If you need to make content-blocking decisions for minor-related content in specific business scenarios, you must perform a secondary review and implement the corresponding handling logic on your own after receiving the detection results.
Version comparison
Compared with Image Moderation 1.0, Image Moderation 2.0 offers broader risk coverage, more granular risk labels, and more flexible console configuration options.
Comparison item | Image Moderation2.0 | Image Moderation 1.0 |
Default QPS | Non-large-model version: 100 Large-model version: 50 | 50 |
Risk detection scope |
| Pornography, politically sensitive and terrorist content, text-image violations, inappropriate scenes, special logos, QR codes |
Risk detection labels | 100+ Note For a full list of labels, see the Risk label interpretation table. | 40+ Note For a full list of labels, see label. |
Console |
Note For more information, see the Console operation guide. |
|
Supported image formats | PNG, JPG, JPEG, BMP, WEBP, TIFF, SVG, GIF, ICO, HEIC | PNG, JPG, JPEG, BMP, GIF, WEBP |
Billing | Billed per service. Each service detects multiple risks simultaneously. Cost = Number of images × Number of services Billed separately for the following services:
| Billing is based on threat scenarios (scene). Billed per risk scenario (scene). Cost = Number of images × Number of risk scenarios Billed separately for the following risk scenarios:
|
Services
Image Moderation V2.0 supports the following services:
Scenario | Service | Detected content | Applicable scenarios |
LLM moderation | Image Moderation for Large and Small Model Integration (postImageCheckByVL_global) | Combines large-model and expert-model capabilities to detect pornography, sexually explicit content, politically sensitive content, terrorist content, prohibited content, religious content, flags, advertising traffic diversion, inappropriate content, abuse, and other violations. Returns detailed labels. For a full list of detectable items, see the AI Guardrails Console. | Use this service when high accuracy is required. Recommended as the preferred choice. For more information, see LLM-based image moderation enhanced service. |
LLM moderation (entering the Chinese mainland) | Image Moderation for Large and Small Model Integration_China Version (postImageCheckByVL_ec) | Designed for entering-the-Chinese-mainland scenarios. Combines large-model and expert-model capabilities to detect pornography, sexually explicit content, politically sensitive content, terrorist content, prohibited content, religious content, advertising traffic diversion, and inappropriate content in images. Returns detailed labels. | Use this service when high accuracy is required for content entering the Chinese mainland. Recommended as the preferred choice. |
General-purpose scenarios | Baseline Check (baselineCheck_global) | Detects pornography, sexually explicit content, terrorist content, prohibited content, flags, inappropriate content, abuse, and special elements in images. Includes both visual content and embedded text in 18 languages: Chinese, English, French, German, Indonesian, Malay, Portuguese, Spanish, Thai, Vietnamese, Japanese, Arabic, Filipino, Hindi, Türkiye, Russian, Italian, and Dutch. For a full list of detectable items, see the AI Guardrails Console. | Use this service to detect whether images contain noncompliant or unsuitable content. Apply it to all images exposed to the public internet. |
AIGC scenarios | AIGC image detection (aigcDetector_global) | Detection target: Images. Detection content: Whether images are suspected to be AI-generated or synthesized. | Use this service to determine whether images are AI-generated. Apply it to tag image sources. For more information, see Image Moderation Enhanced 2.0 AIGC scenario detection service. |
AIGC scenarios | AI-Generated Image Detection_Ultimate (aigcDetector_ultra_global) | Detection target: Images. Detection content: Whether images are suspected to be AI-generated or synthesized, whether they are suspected to be composites (including face synthesis), and whether they show signs of AI local editing. | Use this service to determine whether images are forged. Recommended for image forensics, especially when determining whether images have been locally edited by AI. For more information, see Image Moderation Enhanced 2.0 AIGC scenario detection service. |
Use postImageCheckByVL_global for businesses operating outside mainland China. Use postImageCheckByVL_ec for businesses serving users within mainland China.
2. Billing
Pay-as-you-go
Image Moderation 2.0 uses pay-as-you-go billing by default. You are charged only when you call a service. The following table lists the API pricing.
Moderation type | Supported business scenarios (services) | Unit price |
Standard image moderation (image_standard) |
| 0.6 USD per 1,000 calls Note One call to Baseline Check counts as one billing unit. For example, 100 calls cost 0.06 USD. |
Advanced image moderation (image_advanced) |
| 1.2 USD per 1,000 calls Note You are charged for each invocation. For example, 100 invocations cost 0.12 USD. |
Large-model image moderation (China) (image_vl_standard_cn) |
| 0.7 USD per 1,000 calls Note One call to the China mainland LLM image moderation service counts as one billing unit. For example, 100 calls cost 0.07 USD. |
Content Moderation 2.0 pay-as-you-go usage is billed every 24 hours. In the billing detail report, the moderationType field corresponds to the moderation type field described earlier. You can view Billing Details.
Basic pornography and politically sensitive content detection for images is provided by the Baseline Check service (baselineCheck_global), which belongs to the Standard image moderation (image_standard) moderation type in the preceding table. Each successful call counts as one billing unit. Quota consumption follows two rules:
Resource plan deduction: When you use a resource plan, the deduction coefficient of
image_standardis 2. Each image that is checked once deducts 2 units from the resource plan quota. The coefficients of all moderation types are listed in the Resource plan deduction section.Large-model moderation enabled: If Enable Large Model Moderation is selected on the Rule configuration > Moderation scope configuration page of the Content Moderation console, small-model services such as Baseline Check also generate usage for Basic large-model image moderation (
image_vl_basic). The same image then produces usage under two moderation types, and each counts as one billing unit.
Resource plan deduction
If your moderation volume is large or predictable, purchase a resource plan. Larger plans offer lower per-unit costs than pay-as-you-go pricing. For more information, see Purchase Content Moderation 2.0 resource plans.
Content Moderation 2.0 resource plans offset usage of Content Moderation 2.0 and cannot be shared with the Content Moderation 1.0 data transfer plan. The specific offset factors are as follows:
Moderation type | Offset factor |
Standard image moderation (image_standard) | 2: Each successful API call deducts 2 units from the resource plan quota. Note For example, if your resource plan quota is 10 units and you make 1 successful API call, 2 units are deducted, leaving 8 units. |
Advanced image moderation (image_advanced) | 4: Each successful API call deducts 4 units from the resource plan quota. Note For example, if your resource plan quota is 10 units and you make 1 successful API call, 4 units are deducted, leaving 6 units. |
Large-model image moderation for entering the Chinese mainland (image_vl_standard_cn) | 2.34: Each successful API call deducts 2.34 units from the resource plan quota. Note For example, if your resource plan quota is 100 units and you make 1 successful API call, 2.34 units are deducted, leaving 97.66 units. |
After purchase, Image Moderation 2.0 API usage is deducted from your resource plan quota first. When the quota runs out, pay-as-you-go billing applies automatically. Monitor your remaining quota and charges. You can set low-balance alerts in the Resource Plan System in the Alibaba Cloud User Center.
3. Usage
Integration
Image Moderation 2.0 supports SDK integration and native HTTPS integration:
For SDK integration, see Image Moderation 2.0 SDK and integration guide.
For native API integration:
For large-model image moderation, see Image moderation with LLMs.
For general image detection, see Image moderation V2.0 API.
For AIGC and image authentication scenarios, see AIGC and forgery detection service of image moderation 2.0.
Console operations
Before your first API call, configure your image moderation policy in the AI Guardrails Console.
In the console, you can adjust the detection scope, configure per-business policies, manage custom image libraries and dictionaries, and view call results and usage.
After adjusting your moderation policy, run effect testing in the console to verify the results.
For step-by-step instructions, see the Console
4. FAQ
Q: What are the differences between Image Moderation Enhanced Edition and Image Moderation 1.0?
A: The main differences between Image Moderation 2.0 and Image Moderation 1.0 are as follows:
Risk label count: Image Moderation 2.0 supports 100+ risk labels. Image Moderation 1.0 supports 40+ risk labels.
Billing method: Image Moderation 2.0 is billed per service. Each service call simultaneously detects multiple risk types, so the cost formula is: number of images × number of services. Image Moderation 1.0 is billed per risk scenario (scene). Each scene is counted and billed separately.
Scenario support: Image Moderation 2.0 provides preset policies for dedicated business scenarios, including avatar, post and comment, and marketing material moderation. Image Moderation 1.0 supports general detection scenarios only.
Q: Why does an image that violates regulations return a compliant result, or why cannot non-compliant text embedded in an image be detected?
A: This situation falls within the scope of model detection capabilities (false negative or false positive). If this occurs, we recommend re-testing the image to verify the result. For adversarial scenarios such as images with embedded non-compliant text, the Content Moderation team continuously optimizes detection strategies. A common reason for an advertising violation verdict is that the image contains traffic diversion elements such as QR codes, disguised contact details, or inappropriate guidance that lures users to another destination. When such content is matched, advertising violation labels are returned. These labels are returned by services whose detection scope covers advertising traffic diversion, such as Large- and small-model fusion image moderation (postImageCheckByVL_global). If an advertising violation verdict looks incorrect, or if you encounter advertising-related false negatives or false positives, the behavior may result from ongoing backend adjustments to detection strategies. Re-submit the image to verify the current result. If specific false negative cases or advertising violation misjudgements persist, provide the corresponding RequestID and submit a ticket with detailed information so that the backend team can investigate and adjust further.
Q: How do I estimate the number of images scanned based on my billing charges?
A: You can estimate the number of scans based on the unit price of the corresponding service. Refer to the billing table in the 2. Billing section of this document for the unit price of each service. For example, Standard image moderation is priced at USD 0.6 per 1,000 calls. If your bill shows USD 6, this corresponds to approximately 10,000 scans.
For exact scan counts and detailed usage information, log in to the Content Moderation console and view Usage Statistics, or access Bill Details in Alibaba Cloud User Center.
Q: Does the temporary OSS storage used by Image Moderation Enhanced Edition affect my main Alibaba Cloud account?
A: No. The temporary OSS storage is used solely for the detection process and does not affect the data security of your main Alibaba Cloud account. After detection is complete, the system automatically deletes the temporary data on a regular basis.
Q: What should I do if the recognition accuracy for images photographed from a screen (including moiré patterns) is low?
A: Moiré patterns produced when photographing a screen can reduce model recognition accuracy. We recommend adjusting the photo clarity first and then retrying the detection. If false negatives persist after the adjustment, provide the specific RequestID via a support ticket. The Content Moderation backend team can then further investigate and optimize the detection results.
Q: Why do baselineCheck and baselineCheckByVL return different results for the same image?
A: The two services are independent and are built on different model capabilities, so different results for the same image are expected behavior rather than a defect.
Baseline Check (
baselineCheck) is a small-model service. It focuses on explicit high-risk violations such as pornography, sexually explicit content, politically sensitive content, terrorist content, and prohibited content, and it covers both the visual content and the text embedded in the image. This service belongs to the Standard image moderation (image_standard) moderation type.Large-model image moderation (
baselineCheckByVL) is based on a moderation large model. It understands image semantics and implicit intent more deeply, detects risks such as pornography, politically sensitive content, terrorist content, prohibited content, inappropriate content, abuse, and advertising, and returns top-level labels. It can therefore cover risk dimensions that Baseline Check does not identify. This service belongs to the Large-model image moderation (image_vl_standard) moderation type.
To choose between the two services, compare their unit prices and resource plan deduction coefficients in the 2. Billing section, and decide based on the detection depth and cost that your business requires. To compare results directly, call both services on the same image on the effect testing page of the Content Moderation console, and then review the difference in the returned labels.
Q: How often is data about sensitive public figures, such as officials removed from office, updated?
A: Data about sensitive public figures is not updated in fixed batch cycles. The backend updates the recognition data once the information is verified, and the update takes effect from then on, so there is no fixed periodic update delay. If a specific sensitive public figure is not recognized or is recognized incorrectly, provide the corresponding RequestID and submit a ticket to technical support so that the backend team can verify the information and update the recognition data.