LLM image moderation - AI Guardrails - Alibaba Cloud Documentation Center

How it works

Image Moderation 2.0 uses a custom-trained Qwen LLM based on image content risk characteristics, providing the following image moderation services:

Image Moderation for Large and Small Model Integration:

Combines large and expert models to detect a wide range of non-compliant content in images, such as pornography, suggestive material, politically sensitive content, violence, terrorism, prohibited items, religious content, spam, and other undesirable content. (Note: All large model inference is processed in the Singapore region.)For details on the detection categories, see Rules.
China Mainland Edition of the Large and Small Model Fusion Image Moderation Service (postImageCheckByVL_ec):

Uses large models and expert models to comprehensively detect non-compliant content in images, including pornography, suggestive material, political content, violence and terrorism, prohibited items, religious content, spam, and other harmful content.Image moderation service tailored for the Chinese mainland region, providing fine-grained risk labels and localized content understanding. For details on the detection categories, see Rules.

For the US (Virginia), Germany (Frankfurt), and China (Hong Kong) regions, LLM inference runs in Singapore. Data and logs are stored locally in the respective regions (Frankfurt and Hong Kong).

Select your service

Service	Description	Supported regions	Use cases
Image Moderation for Large and Small Model Integration(`postImageCheckByVL_global`)	Combines an LLM and expert models to provide more granular labels, such as pornography subcategories, specific behaviors, and specific objects. Offers a wider detection range and richer labels. Provides low false positive and false negative rates.	Singapore, China (Hong Kong), US (Virginia), Germany (Frankfurt)	Social media, live streaming, gaming, e-commerce, and education businesses that require strict risk control and fine-grained policies. Businesses that need detailed risk labels. Highly recommended for new users with high performance requirements.
postImageCheckByVL_ec	Designed for cross-border businesses entering the Chinese mainland. Extends the Image Moderation for Large and Small Model Integration service with optimizations for Chinese-language context and domestic regulatory compliance requirements.	China (Shanghai), China (Hangzhou), China (Beijing), China (Shenzhen), China (Chengdu)	Cross-border platforms and overseas applications that need to moderate Chinese-language content. Overseas businesses that must meet Chinese mainland compliance requirements.

The detection capabilities of postImageCheckByVL_cb cover all detection capabilities of baselineCheck_cb, and additionally provide 12 detection labels (such as pornographic_underage, QRCode, and pt_logo). Some risk labels use different names between the two services, for example: pornographic_art (baselineCheck_cb) vs. pornographic_artwork (postImageCheckByVL_cb), sexual_shoulder (baselineCheck_cb) vs. sexual_femaleShoulder (postImageCheckByVL_cb), and violent_blood (baselineCheck_cb) vs. violent_bloody (postImageCheckByVL_cb). To obtain the original label names, call the two services separately.

Prerequisites

Before you begin, ensure that you have:

An active Image Moderation 2.0 subscription (pay-as-you-go)
An Alibaba Cloud account or a RAM user with the AliyunYundunGreenWebFullAccess policy
An AccessKey pair for authentication

Get started

Step 1: Activate the service

Visit the service activation page and activate Image Moderation 2.0. After activation, the default billing method is pay-as-you-go. Fees are settled daily based on actual usage. You are not charged if you do not call the service. For details on the billing methods, see Billing.

Step 2: Grant permissions to a RAM user

Create an AccessKey pair for your Alibaba Cloud account or a RAM user. The RAM user must have the AliyunYundunGreenWebFullAccess policy to call Content Moderation APIs.

Log on to the RAM console using your Alibaba Cloud account or as a RAM administrator.
Create a RAM user. See Create a RAM user.
Attach the AliyunYundunGreenWebFullAccess policy to the RAM user. See Manage permissions for RAM users.

Step 3: Install the SDK

Follow the Image Moderation SDK and integration guide to install the SDK and configure your endpoint.

The following table lists all supported regions and endpoints:

Region	Public endpoint	VPC endpoint	Supported service
China (Shanghai)	green-cip.cn-shanghai.aliyuncs.com	green-cip-vpc.cn-shanghai.aliyuncs.com	postImageCheckByVL_ec
China (Hangzhou)	green-cip.cn-hangzhou.aliyuncs.com	green-cip-vpc.cn-hangzhou.aliyuncs.com	postImageCheckByVL_ec
China (Beijing)	green-cip.cn-beijing.aliyuncs.com	green-cip-vpc.cn-beijing.aliyuncs.com	postImageCheckByVL_ec
China (Shenzhen)	green-cip.cn-shenzhen.aliyuncs.com	green-cip-vpc.cn-shenzhen.aliyuncs.com	postImageCheckByVL_ec
China (Chengdu)	green-cip.cn-chengdu.aliyuncs.com	N/A	postImageCheckByVL_ec
Singapore	green-cip.ap-southeast-1.aliyuncs.com	green-cip-vpc.ap-southeast-1.aliyuncs.com	postImageCheckByVL_global
Malaysia (Kuala Lumpur)	green-cip.ap-southeast-3.aliyuncs.com	N/A
Indonesia (Jakarta)	green-cip.ap-southeast-5.aliyuncs.com	N/A
China (Hong Kong)	green-cip.cn-hongkong.aliyuncs.com	green-cip-vpc.cn-hongkong.aliyuncs.com
US (Virginia)	green-cip.us-east-1.aliyuncs.com	green-cip-vpc.us-east-1.aliyuncs.com
Germany (Frankfurt)	green-cip.eu-central-1.aliyuncs.com	green-cip-vpc.eu-central-1.aliyuncs.com

Important

In the Germany (Frankfurt) and China (Hong Kong) regions, LLM inference runs in the Singapore region. Data and logs are stored locally in the respective regions (Frankfurt and Hong Kong).

Note

The UK (London), Malaysia (Kuala Lumpur), and Indonesia (Jakarta) regions reuse the console configuration of the Singapore region. The US (Silicon Valley) region reuses the console configuration of the US (Virginia) region.

Step 4: Configure detection rules (optional)

In the Content Moderation console, configure detection rules: enable or disable detection categories, copy a service, configure a custom image library, configure a custom glossary, query detection records, and review usage data. See Console guide.

API reference

API overview

API: ImageModeration
Service code: postImageCheckByVL_global, postImageCheckByVL_ec
QPS limit: 50 calls/second per user. Exceeding this limit throttles requests. Contact your business manager for a quota increase.
Billing: Billed per successful request (HTTP 200), settled daily. Pricing varies by service. For details, see the Billing section.

Debug the API

Before integration, you can use Alibaba Cloud OpenAPI online debugging to try the and Image Moderation 2.0 APIs. You can also view sample code and SDK dependency information to obtain an overview of how to use these APIs and their parameters.

Important

API calls through the online debugger are billed.

Image requirements

Constraint	Limit
Supported formats	PNG, JPG, JPEG, BMP, WEBP, TIFF, SVG, HEIC (longest edge < 8,192 px), GIF (first frame), ICO (last image)
Max file size	20 MB
Max dimensions	16,384 px (height or width); 250 million total pixels
Optimal resolution	At least 200 x 200 px (lower resolutions reduce accuracy)
Download timeout	3 seconds
URL restrictions	Publicly accessible; max 2,048 characters; no Chinese characters; one URL per request

Submit an image

Submit an image using one of the following methods per request:

Method	Required parameters	Notes
URL	`imageUrl`	URL must be publicly accessible
OSS authorization	`ossBucketName`, `ossObjectName`, `ossRegionId`	Grant `AliyunCIPScanOSSRole` on the Cloud Resource Access Authorization page
Local upload	Upload via SDK	File is deleted 30 minutes after upload; does not consume OSS storage. See the Image Moderation SDK guide for code examples.

Request parameters

The request body is a JSON object. For required common request parameters, see the Integration guide.

Top-level parameters

Parameter

Type

Required

Example

Description

Service

String

Yes

postImageCheckByVL_global

The detection service. Valid values:

postImageCheckByVL_global: Image Moderation for Large and Small Model Integration
postImageCheckByVL_ec: Hybrid Model Image Moderation Service

ServiceParameters

JSONString

Yes

—

A JSON string containing the content detection parameters.

ServiceParameters fields

Parameter	Type	Required	Example	Description
`imageUrl`	String	Conditional	`https://img.alicdn.com/tfs/TB1U4r9AeH2gK0jSZJnXXaT1FXa-2880-480.png`	The URL of the image to moderate. Required when submitting by URL.
`ossBucketName`	String	Conditional	`bucket_01`	The name of the authorized Object Storage Service (OSS) bucket. Required when submitting by OSS.
`ossObjectName`	String	Conditional	`2022023/04/24/test.jpg`	The object key of the image in the OSS bucket. Required when submitting by OSS.
`ossRegionId`	String	Conditional	`cn-beijing`	The region where the OSS bucket is located. Required when submitting by OSS.
`dataId`	String	No	`img123****`	A unique identifier to associate the result with your business data. Alphanumeric characters, underscores, hyphens, and periods. Max 64 characters.
`infoType`	String	No	`customImage`	Specifies additional information to return. Valid values: `customImage` (returns custom image library match details). Separate multiple values with commas.
`referer`	String	No	`www.aliyun.com`	The `Referer` request header, used for hotlink protection. Max 256 characters.

Request example

{
    "Service": "postImageCheckByVL_global",
    "ServiceParameters": {
        "imageUrl": "https://img.alicdn.com/tfs/TB1U4r9AeH2gK0jSZJnXXaT1FXa-2880-480.png",
        "dataId": "img0307****"
    }
}

Response parameters

Top-level response fields

Parameter	Type	Example	Description
`RequestId`	String	`70ED13B0-BC22-576D-9CCF-1CC12FEAC477`	The unique request ID. Used for troubleshooting.
`Code`	Integer	`200`	The status code.
`Msg`	String	`OK`	The response message.
`Data`	Object	—	Detection results.

Data fields

Parameter	Type	Example	Description
`RiskLevel`	String	`high`	Overall risk level, based on the highest-risk label. Valid values: `high`, `medium`, `low`, `none`.
`DataId`	String	`img123******`	Data ID from the request. Returned only if `dataId` was specified in the request.
`Result`	Array	—	Array of detected risk labels. Each entry contains `Label`, `Confidence`, `Description`, and `RiskLevel`.
`Ext`	Object	—	Supplementary information, including custom library matches.

Result fields

Parameter	Type	Example	Description
`Label`	String	`violent_explosion`	Risk label. A single image can match multiple labels.
`Confidence`	Float	`81.22`	Confidence score (0–100, two decimal places). Higher score means higher confidence.
`Description`	String	`Fireworks content`	Human-readable label description. Use `Label` — not `Description` — to determine actions, as this field may change.
`RiskLevel`	String	`high`	Risk level for this label, based on configured score thresholds. Valid values: `high`, `medium`, `low`, `none`.

Ext fields

Parameter	Type	Description
`CustomImage`	JSONArray	Custom library match details. Returned when a submitted image matches a library entry.

CustomImage fields

Parameter	Type	Example	Description
`LibId`	String	`lib0001`	ID of the matched custom library.
`LibName`	String	`Custom Image Library A`	Name of the matched custom library.
`ImageId`	String	`20240307`	ID of the matched image in the library.

Response example

{
    "RequestId": "70ED13B0-BC22-576D-9CCF-1CC12FEAC477",
    "Code": 200,
    "Msg": "OK",
    "Data": {
        "RiskLevel": "high",
        "DataId": "img0307****",
        "Result": [
            {
                "Label": "violent_explosion",
                "Confidence": 92.40,
                "Description": "Fireworks content",
                "RiskLevel": "high"
            },
            {
                "Label": "violent_burning",
                "Confidence": 67.15,
                "Description": "Burning scenes",
                "RiskLevel": "medium"
            }
        ],
        "Ext": {}
    }
}

Request and response examples are formatted for readability. Actual API responses do not include line breaks or indentation.

Risk labels

The service returns risk labels grouped by category. Each label has a confidence score (0–100). Higher score means higher confidence. Enable or disable individual labels in the Content Moderation console.

Understanding risk levels and handling results

Each label has a RiskLevel and Confidence. The RiskLevel in the Data object reflects the highest risk across all labels.

Use RiskLevel to guide your moderation workflow:

Risk level	Recommended action
`high`	Block or remove content immediately
`medium`	Route to manual review
`low`	Process only if your use case requires high recall; otherwise treat as no risk
`none`	No risk detected

Tuning confidence thresholds: Default risk score thresholds determine when a label is assigned high, medium, or low. Lowering a threshold increases recall (fewer missed violations) but also increases false positives. Raising a threshold improves precision but may miss some violations. Adjust thresholds per label in the Content Moderation console to match your platform's tolerance for false positives versus false negatives.

Store returned risk labels and confidence scores. Use them to prioritize manual review, build annotation datasets, and apply tiered governance policies.

Label reference

Labels are grouped into the categories below. The _tii suffix indicates text detected within the image (text-in-image), not visual content.

Pornographic content (pornographic_*)

Label	Description
`pornographic_adultContent`	Image suspected to contain pornographic content.
`pornographic_adultToys`	Image suspected to contain adult toy content.
`pornographic_artwork`	Image suspected to contain erotic artwork content.
`pornographic_underage`	Image suspected to contain underage pornographic content.
`pornographic_adultContent_tii`	Image text suspected to contain pornographic content.

Sexually suggestive content (sexual_*)

Label	Description
`sexual_suggestiveContent`	Image suspected to contain vulgar or sexual hint content.
`sexual_breastBump`	Image suspected to contain nipple outline content.
`sexual_cleavage`	Image suspected to contain female cleavage features.
`sexual_femaleUnderwear`	Image suspected to contain underwear or swimsuit content.
`sexual_maleTopless`	Image suspected to contain male shirtless content.
`sexual_femaleShoulder`	Image suspected to contain shoulder sexy content.
`sexual_femaleLeg`	Image suspected to contain leg sexy content.
`sexual_pregnancy`	Image suspected to contain pregnancy or breastfeeding content.
`sexual_cartoon`	Image suspected to contain cartoon sexy content.
`sexual_underage`	Image suspected to contain underage sexy content.
`sexual_feet`	Image suspected to contain foot sexy content.

Politically sensitive content (political_*)

Label	Description
`political_historicalNihility`	Image suspected to involve historical nihilism or inappropriate historical events.
`political_historicalNihility_tii`	Image text suspected of historical nihilism.
`political_politicalFigure_metaphor_tii`	Image text suspected to contain metaphor or alias for key leaders.
`political_politicalFigure_name_tii`	Image text contains leader names.
`political_prohibitedPerson_tii`	Image text suspected to contain names of fallen officials.
`political_taintedCelebrity_tii`	Image text suspected to contain names of tainted celebrities.
`political_taintedCelebrity`	Image suspected to contain tainted or heavily negative public figures.
`political_Chinaflag`	Image suspected to contain Chinese national flag.
`political_otherflag`	Image suspected to contain other national flags.
`political_Chinamap`	Image suspected to contain Chinese map.
`political_logo`	Image suspected to contain banned media logos.
`political_outfit`	Image suspected to contain military or police uniforms.
`political_badge`	Image suspected to contain national or party emblems.
`political_politicalFigure_1`	Image suspected to contain current or former leaders.
`political_politicalFigure_2`	Image suspected to involve leader family members.
`political_politicalFigure_3`	Image suspected to involve provincial or municipal government personnel.
`political_politicalFigure_4`	Image suspected to involve foreign leaders and family members.
`political_prohibitedPerson_1`	Image suspected to contain national-level fallen officials.
`political_prohibitedPerson_2`	Image suspected to contain provincial or municipal-level fallen officials.
`political_racism_tii`	Image text suspected to contain racist expressions.
`political_medicalOutfit`	Image suspected to contain medical attire.

Violence and terrorism (violent_*)

Label	Description
`violent_crowding`	Image suspected to contain crowd gathering.
`violent_explosion`	Image suspected to contain fireworks or explosion elements.
`violent_armedForces`	Image suspected to contain armed forces elements.
`violent_gun`	Image suspected to contain firearms.
`violent_knives`	Image suspected to contain knives.
`violent_gunKnives_tii`	Image text contains firearm or knife descriptions.
`violent_blood`	Image suspected to contain bloody content.
`violent_horrific`	Image suspected to contain horrific content.
`violent_horrific_tii`	Image text suspected to describe violence or terror content.
`violent_burning`	Image suspected to contain burning content.
`violent_ACU`	Image suspected to contain combat uniforms.

Contraband (contraband_*)

Label	Description
`contraband_drug`	Image suspected to involve drugs or medicines.
`contraband_drug_tii`	Image text suspected to describe drugs, contraband, or prohibited items.
`contraband_gamble`	Image suspected to contain gambling items.
`contraband_gamble_tii`	Image text suspected to describe gambling behavior.
`contraband_certificate_tii`	Image text suspected to contain certificate or cash-out adtraffic diversion.
`contraband_illegal`	Image suspected to contain contraband items.
`contraband_illegal_tii`	Image text suspected to contain contraband descriptions.

Religious content (religion_*)

Label	Description
`religion_funeral`	Image suspected to contain funeral or memorial hall content.
`religion_buddhism`	Image suspected to contain Buddhist or Taoist attire or symbols.
`religion_christianity`	Image suspected to contain Christian attire or symbols.
`religion_muslim`	Image suspected to contain Muslim attire or symbols.
`religion_tii`	Image text contains religious elements or information.

Flags

Label	Description
`political_Chinaflag`	Image suspected to contain Chinese national flag.
`political_otherflag`	Image suspected to contain other national flags.

Spam and promotional content (pt_*)

Label	Description
`pt_logotoSocialNetwork`	Image suspected to contain common social network platform watermarks.
`pt_qrCode`	Image suspected to contain QR codes.
`pt_programCode`	Image suspected to contain mini program codes.
`pt_toDirectContact_tii`	Image text contains contact information for traffic diversion.
`pt_toSocialNetwork_tii`	Image text contains social platform traffic diversion information.
`pt_toShortVideos_tii`	Image text contains short video platform traffic diversion information.
`pt_investment_tii`	Image text contains investment or financial advertising information.
`pt_recruitment_tii`	Image text contains part-time recruitment advertising information.

Inappropriate behavior (inappropriate_*)

Label	Description
`inappropriate_smoking`	Smoking.
`inappropriate_drinking`	Drinking.
`inappropriate_tattoo`	Tattoo.
`inappropriate_middleFinger`	Middle finger gesture.
`inappropriate_foodWasting`	Food wasting content.
`inappropriate_underage`	Underage content.

Profanity (profanity_*)

Label	Description
`profanity_oral_tii`	Image text contains vulgar or offensive language.
`profanity_offensive_tii`	Image text contains severe abusive content.

Custom image library labels

Configure a custom image library in the console. When a submitted image matches a library image, the system returns the label with a _lib suffix (for example, violent_explosion_lib). The Confidence score reflects similarity.

No-risk labels

Label	Confidence score	Description
`nonLabel`	Not returned	No threats detected, or all detection categories are disabled.
`nonLabel_lib`	0–100	The image is highly similar to an exempted image in your custom library.

Status codes

Requests are billed only for status code 200.

Code	Description
`200`	Request succeeded.
`400`	A required parameter is empty.
`401`	A parameter value is invalid.
`402`	A parameter exceeds the maximum length. Correct and retry.
`403`	QPS limit exceeded. Reduce concurrency and retry.
`404`	Image download failed. Check the URL or retry.
`405`	Image download timed out. Verify accessibility and retry.
`406`	Image file too large. Resize and retry.
`407`	Unsupported image format. Use a supported format and retry.
`408`	Insufficient permissions. Verify service activation, no overdue payments, and RAM user has the required policy.
`500`	Internal system error. Retry later.

Billing

Image Moderation 2.0 LLM image moderation service supports pay-as-you-go billing. Requests returning non-200 status codes are not charged.

Pay-as-you-go

After activating Image Moderation 2.0, the default billing method is pay-as-you-go. Fees are settled daily based on actual usage. You are not charged if you do not call the service.

Billing category	Included service	Unit price
Image Moderation advanced (`image_advanced`)	Image Moderation for Large and Small Model Integration: `postImageCheckByVL_global`	USD 1.20 per 1,000 calls Note Each call to any of the services listed on the left counts as one billing unit. Fees are based on actual usage. For example, 100 calls to the Image Moderation for Large and Small Model Integration service cost USD 0.12.
Image Moderation China Mainland Edition (`image_vl_standard_cn`)	Hybrid Model Image Moderation Service: `postImageCheckByVL_ec`	USD 0.70 per 1,000 calls

Note

Content Moderation 2.0 pay-as-you-go billing is settled once daily (within 24 hours). In bill details, the moderationType field identifies the moderation type. View your bill details.

Resource plan deduction

For large-volume or consistent moderation needs, purchase a resource plan in advance. Larger plans offer greater discounts. You can purchase and use multiple plans at the same time. For more information, see Purchase a resource plan for Content Moderation Enhanced Edition.

This resource plan applies to Content Moderation Enhanced Edition. It cannot be shared with the resource plans for Content Moderation 1.0. The deduction ratios are as follows:

Type

Deduction ratio

Image Moderation advanced (image_advanced)

Deduction ratio: 4. Each API call deducts 4 from your resource plan quota.

Example: a 100-call quota minus one API call leaves a balance of 96 calls.

Image Moderation China Mainland Edition (image_vl_standard_cn)

Deduction ratio: 2.34. Each API call deducts 2.34 from your resource plan quota.

Example: a 100-call quota minus one API call leaves a balance of 97.66 calls.

Resource plan quota is applied before pay-as-you-go billing. When the quota is exhausted, usage is billed on a pay-as-you-go basis. Monitor your resource plan balance and pay-as-you-go bills. You can set up low-balance alerts in the Resource Plan system of the Alibaba Cloud User Center.

AI Guardrails:Image Moderation with LLMs

How it works

Select your service

Prerequisites

Get started

Step 1: Activate the service

Step 2: Grant permissions to a RAM user

Step 3: Install the SDK

Step 4: Configure detection rules (optional)

API reference

API overview

Debug the API

Image requirements

Submit an image

Request parameters

Request example

Response parameters

Response example

Risk labels

Understanding risk levels and handling results

Label reference

Custom image library labels

No-risk labels

Status codes

Billing

Pay-as-you-go

Resource plan deduction

What's next