RekognitionGuard
Amazon Rekognition for Blueprints and C++

Overview

RekognitionGuard connects Unreal Engine 4.27 and 5.8 to Amazon Rekognition. Its core is UGC moderation: hand it a texture, a render target, a screenshot, a file, bytes or an S3 object and it fires On Allowed, On Review or On Blocked, with the reasons. Requests and image encoding run on background worker threads; every result arrives on the game thread.

• Is Image Safe, Moderate Texture, Moderate Render Target, Moderate Screenshot: one node from player content to a verdict.
• Moderation policies with category rules and thresholds, presets (All Ages, Teen, Mature), lenient mode for drawings, video verdicts with timestamps.
• Image input from bytes, files, UTexture2D (also compressed and streamed textures), UTextureRenderTarget2D, the game viewport and S3; JPEG encoding with the engine ImageWrapper; automatic downscaling under the 5 MB limit.
• Vision: labels with boxes, dominant colours and image quality, text, faces with attributes, face comparison, protective equipment, celebrities.
• Face collections and users: index, search, list, delete, associate — for duplicate-account checks, ban evasion detection or face sign-in.
• Stored video (S3): content moderation and label detection jobs, with Wait For ... nodes that poll and collect every page.

Requirements

Installation

1. Install RekognitionGuard from Fab into your engine, or copy the RekognitionGuard folder into your project's Plugins folder.
2. Enable RekognitionGuard in Edit > Plugins and restart the editor.
3. Open Project Settings > Plugins > RekognitionGuard and set Region and Identity Pool Id.
4. C++ only: add PublicDependencyModuleNames.Add("RekognitionGuard"); to your Build.cs.

Setting up AWS

1. In Amazon Cognito create an identity pool in your Rekognition region. Enable guest access if players may submit images before signing in.
2. Give the role the pool hands to players only the image analysis calls the game needs. Collection writes (Index Faces, Create User, Delete ...) and video jobs belong on a server or a tool with its own role.

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "PlayerImageChecks",
      "Effect": "Allow",
      "Action": [
        "rekognition:DetectModerationLabels",
        "rekognition:DetectLabels",
        "rekognition:DetectText",
        "rekognition:DetectFaces",
        "rekognition:CompareFaces",
        "rekognition:DetectProtectiveEquipment"
      ],
      "Resource": "*"
    },
    {
      "Sid": "SearchOneCollection",
      "Effect": "Allow",
      "Action": ["rekognition:SearchFacesByImage", "rekognition:SearchUsersByImage"],
      "Resource": "arn:aws:rekognition:us-east-1:123456789012:collection/mygame-players"
    }
  ]
}

Server / tool role, in addition: rekognition:CreateCollection, DeleteCollection, DescribeCollection, ListCollections, IndexFaces, ListFaces, DeleteFaces, SearchFaces, CreateUser, DeleteUser, ListUsers, AssociateFaces, DisassociateFaces, SearchUsers, RecognizeCelebrities, GetCelebrityInfo, StartContentModeration, GetContentModeration, StartLabelDetection, GetLabelDetection. Images and videos in S3 need s3:GetObject on the objects for the caller; video jobs with an SNS topic need iam:PassRole for the topic role.

Quick start

Moderate an avatar

1. Get the picture: a texture the player picked (Moderate Texture), a file (Load Image File then Is Image Safe) or bytes from an HTTP upload (Make Image From Bytes then Is Image Safe).
2. Leave Policy unconnected to use the settings preset (Teen by default), or connect Make Moderation Policy.
3. On Allowed: upload Image > Bytes (already JPEG and under 5 MB). On Review: hide it until a moderator looked at it; log Verdict To Text. On Blocked: tell the player; Verdict > Reasons lists the labels.

Moderate a spray canvas

Players paint into a render target. Before replicating it, call Moderate Render Target with the canvas. Set Lenient For Illustrations in the policy so drawings are reviewed rather than blocked at the same confidence. Add Detect Text + Get Text Lines to run written words through your own word filter.

Profile photo check

Detect Faces with Attributes = All, then Check Profile Photo: one sharp, bright, frontal face without sunglasses; Problems holds player-friendly messages.

Images

Captured images are downscaled so the longer side is at most Max Image Dimension (2048 px) and encoded at Jpeg Quality (85); if the JPEG is still above 5 MB the quality and then the size are reduced until it fits. Every node returns or works on exactly the bytes that were sent.

Moderation policies

Rekognition returns moderation labels in a three-level taxonomy (version 7). The 10 top-level categories are: Explicit, Non-Explicit Nudity of Intimate parts and Kissing, Swimwear or Underwear, Violence, Visually Disturbing, Drugs & Tobacco, Alcohol, Rude Gestures, Gambling, Hate Symbols. Get Moderation Categories lists them and their level-2 labels.

A policy is a list of rules Category, Action (Ignore / Review / Block), Min Confidence plus an action for unlisted labels. Evaluation: only the most specific detected labels are judged (a parent is skipped when one of its children was detected); for each, the closest category in its chain (label, parent, top-level) that has rules decides; among those rules the strongest one whose confidence is reached wins. So Weapons: Review 60 overrides Violence: Block 80 for a weapon, while graphic violence still follows the Violence rule.

Node reference (75)

Latent nodes run in the background, can be stopped with Cancel on their return value and fire exactly one final pin on the game thread. Collection Id empty = Default Collection Id; thresholds 0 = Default Face Match Threshold.

Config and identity

Images

Moderation

Vision

Face collections and users

Stored video (S3)

Types

Settings

Project Settings > Plugins > RekognitionGuard (saved to DefaultGame.ini). Changes apply to the next request.

C++

// Build.cs: PublicDependencyModuleNames.Add("RekognitionGuard");
#include "RekognitionGuardAsyncActions.h"
#include "RekognitionGuardLibrary.h"

URekognitionGuardModerateAction* Node = URekognitionGuardModerateAction::IsImageSafe(this,
    URekognitionGuardLibrary::MakeImageFromBytes(UploadBytes), FRekognitionGuardModerationPolicy());
Node->OnAllowed.AddDynamic(this, &UMyUploader::HandleAllowed);   // (Verdict, Image, Error)
Node->OnReview.AddDynamic(this, &UMyUploader::HandleReview);
Node->OnBlocked.AddDynamic(this, &UMyUploader::HandleBlocked);
Node->OnFailure.AddDynamic(this, &UMyUploader::HandleFailure);
Node->Activate();

URekognitionGuardModerationLibrary::EvaluateWithPolicy evaluates labels with exactly the given policy on any thread.

Troubleshooting

Credits

© 2026 Alpha XP — alphaxp.net · Support: support@alphaxp.net. Includes the AWS SDK for C++ and the AWS Common Runtime (Apache License 2.0). Built for Unreal Engine 4.27 and 5.8.