Models
Truebees detection is powered by a set of specialised machine learning models. Understanding which model produced a verdict helps you interpret results, tune your thresholds, and reason about edge cases.
Pipelines vs. models
A pipeline is a named detection configuration you reference in the API (e.g., pipeline=cantaloupe). Internally, a pipeline may invoke one or more models in sequence. The model name returned in the verdict identifies which specific model scored the submitted image.
For example, the cantaloupe pipeline runs both the Cantaloupe deepfake detector and the Azalea social-media classifier. Each model contributes its own score to the verdict.
See Verification API Deep Dive for how to specify a pipeline in a request, and Platform Concepts for how to interpret the score and status fields in a verdict.
Available models
| Model | Type | Brief description |
|---|---|---|
| Cantaloupe | Deepfake image detector | Residual-network detector that scores the squared distance of image features from learned reference centres |
| Azalea | Social-media classifier | Hierarchical forensic classifier that detects social-media sharing and attributes supported source platforms |
| Honeydew | Source attribution & localization | Foundational-vision-transformer model that classifies image origin (real, GAN, diffusion, or partially tampered) and produces a tampering-localization heatmap |
| Pitaya | Video deepfake detector | Foundational-vision-transformer model that detects fully AI-generated videos by analyzing spatio-temporal artifacts across temporally distributed clips |
| Kumquat | Screenshot detector | Multi-stream detector that combines file-encoding, pixel-forensic, and semantic signals to distinguish screenshots from original images |
| Ginkgo | Deepfake image detector | Foundational-vision-transformer detector with a lightly adapted frozen backbone, robust to compression and real-world degradations |
Model cards
Each model page provides a high-level technical explanation suitable for integration and debugging. Detailed model cards — covering training datasets, evaluation benchmarks, and specific weight versions — will be published as sub-pages under each model as they become available.