Imager-1-Pro

Imager-1-Pro is the high-capacity model of the Imager family: an AI-generated image detector with 89M parameters, built on ConvNeXt-Base and trained on Defactify (DALL·E 3, Midjourney v6, SDXL, Stable Diffusion 3).

Results

Accuracy	F1-IA	AUC	TPR@1%FPR	Parameters
99.33%	99.60%	99.78%	99.00%	89M

Validation Accuracy: 99.33% — evaluated on a stratified group split (no data leakage between train/val/test).

F1-IA: 99.60% — F1 score for the AI-generated class, the most relevant metric for detection.

AUC: 99.78% — area under the ROC curve, showing strong separability between classes.

TPR@1%FPR: 99.00% — detects 99% of AI images while keeping false positives below 1%.

Parameters: 89M — optimized for high-accuracy deployment on servers with ONNX support.

Architecture

Field	Value
Base	ConvNeXt-Base
Parameters	89M
Pretrained weights	ImageNet-1K
Input size	224 × 224 RGB
Framework	PyTorch
Output	2 classes

Training

Field	Value
Dataset	Defactify (DALL·E 3, Midjourney v6, SDXL, SD3)
Optimizer	AdamW
Learning rate	1e-4 (backbone) / 1e-3 (head)
Weight decay	0.05
Scheduler	Cosine with warmup + re-warmup on unfreeze
Loss	CrossEntropyLoss (label_smoothing=0.05)
Batch size	32 (effective 64 with grad accumulation)
Precision	AMP FP16
Gradient clipping	1.0
Augmentation	JPEG, down/up sampling, blur, crop, flip
EMA decay	0.999
Training time	11.5 hours

About this release

This repository contains the base model of Imager-1-Pro, released for developers and researchers who want direct access to the weights and training pipeline.

Imager always publishes a separately verified and production-ready version of each model, tested against real-world generators before release. This base model is provided so the community can build on top of it, fine-tune it, or integrate it into their own systems.

Access to the most powerful Imager models is always granted through the official channels.

Task

Binary image classification.

Label	Meaning
0	Real
1	AI-generated

License

Apache 2.0