DETECT-World • REsemble Research

DETECT-World: a world model for deepfake detection

DETECT-World combines DETECT-3B Omni's artifact-based detection with a new layer that checks physical consistency across lighting and motion. This gives it a faster, more confident path to day-zero coverage as new generators launch.

Get started
Trusted by
THE PROBLEM

Detectors need to keep pace with new generative models

New AI models and variants are released every day. Detection systems that rely solely on recognizing patterns from known generators can see false negative rates rise in the gap before they're updated to cover a new one. DETECT-World reviews both known generation patterns and physical consistency together, so it can flag a generator it hasn't been trained on.
‍

STEP 1 • CONNECT

Send audio, image, or video through the API, using real-time streaming or batch processing.

STEP 2 • EVALUATE

DETECT-World checks the content for known generation artifacts and for inconsistencies with physical reality, such as lighting and motion.

STEP 3 • SCORE

Returns a confidence score per modality, in real time. Act on the score in-workflow or incorporate with Resemble Intelligence for auditable results.

HOW THE MODEL WORKS

Two detection methods, combined

DETECT-World builds on the foundation of DETECT-3B Omni, which analyzes raw pixel and audio data for statistical patterns generation models leave behind. DETECT-World adds a second layer on top of that same model: checking whether content is consistent with physical reality. Every piece of content is evaluated on both dimensions, so a deepfake that evades one check can still be caught by the other.
Lighting and visual consistency
DETECT-World is trained to evaluate whether visual details in a scene are physically consistent, for example, whether lighting behaves the way it would in a real environment.
Continuity over time
DETECT-World is trained to evaluate whether a scene remains physically consistent over time, for example, whether objects and materials behave the way they would in reality.
Combined detection
Together, the two techniques — precision on recognizable threats, and broader coverage on the new ones — makes DETECT-World built to cover zero-day attacks from new generators.
DETECT-WORLD IN ACTION

Benchmarked against 250+ generators

A sample of what DETECT-World returns on real submissions, across modalities achieveing between 99.80% and 100% accuracy on generators like Midjourney, Kling 3, Seedance, Open AI, and more.

Original image generated by Midjourney v7 2025
‍
Detected with a confidence score of 100% with DETECT-World
‍
Original video generated by Kling 3 2026
‍
Detected with a confidence score of 98.80% with DETECT-World
‍
BENCHMARKS

Reliable detection across every modality

All figures below are internal benchmarks, pending third-party validation.

Detection accuracy by modality
Audio
99.47%
Video
98.2%
Image
95.77%
0%
Accuracy % (higher = better)
100%
Audio has been publicly benchmarked via Podonos. Video and image are internal results, pending external validation.
BUILT ON RESEMBLE AI

DETECT-World: one layer of a broader safety stack

Detection answers whether content is synthetic. Watermarking detects any AI records, establishes provenance, and satisfies compliance. Intelligence sheds light on whether the presence of AI indicates fraud or malicious intent. Together they provide the most comprehensive approach to stopping AI misuse.
product
Multimodal Watermarking
Ship the watermark as part of your production pipeline. Watermark generated content and verify it later via API.
EXPLORE RESEMBLE WATERMARKER
product
Deepfake Detection
Real-time, multimodal detection across audio, video, and image. The complement to watermarking for content you didn’t generate.
Explore resemble Detect
PRODUCT
Human-readable Reports
Explains what triggered a flag, and why, in plain language. The context layer that pairs with Resemble Detect on every result..
EXPLORE RESEMBLE INTELLIGENCE
Frequently asked questions
What is DETECT-World, and how is it different from DETECT-3B Omni?
DETECT-World expands on DETECT-3B Omni's foundation, reviewing both known generation patterns and physical consistency, including lighting and continuity. This gives it a faster, more confident path to catching content from generators it hasn't seen before.
What does "World" mean in DETECT-World?
The name reflects DETECT-World's approach of reviewing physical consistency in a scene, such as lighting and continuity, in addition to detecting known generation patterns.
Does DETECT-World need to be retrained when a new generator is released?
DETECT-World's architecture is designed to train faster against new generators than the prior approach, and its physical-consistency checks give it a sizable head start on day-zero coverage even before retraining happens.
What languages does DETECT-World support?
DETECT-World supports 54 languages, validated against MLAADv10, up from 51 in DETECT-3B Omni. Because the model identifies generation patterns in audio structure rather than relying on specific words, accuracy holds up on languages not included in training.
How do I integrate DETECT-World?
DETECT-World uses the same API as Resemble Detect. See the API documentation for endpoint details.
Get complete generative AI security
Join thousands of developers and enterprises securing with Resemble AI
Contact us