[Paper] UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolutionary Image Generation and Generated Image Detection

Source: arXiv - 2604.21904v1

Overview

The paper UniGenDet proposes a single, unified architecture that simultaneously trains an image‑generation model and a detector that spots AI‑generated images. By letting the two components share information through a novel multimodal self‑attention module, the authors show that each task can boost the other: the generator produces more realistic pictures, while the detector becomes better at spotting fakes. This co‑evolutionary approach bridges the long‑standing gap between generative (e.g., GANs, diffusion) and discriminative (e.g., forensics) pipelines, delivering state‑of‑the‑art results on several benchmark datasets.

Key Contributions

• Unified Generative‑Discriminative Framework – a single network that jointly learns image synthesis and generated‑image detection, eliminating the need for separate, hand‑crafted pipelines.
• Symbiotic Multimodal Self‑Attention (MMSA) – a cross‑modal attention block that lets the generator and detector exchange feature maps in real time, improving both fidelity and detection accuracy.
• Unified Fine‑Tuning Algorithm – a training schedule that alternates between generation loss and detection loss while keeping a shared backbone, ensuring stable co‑training.
• Detector‑Informed Generative Alignment (DIGA) – a loss term that penalizes the generator when its outputs are easily classified as fake, encouraging it to respect the detector’s learned authenticity criteria.
• Comprehensive Empirical Validation – experiments on FFHQ, LSUN‑Bedroom, and a synthetic deep‑fake dataset show consistent gains over the best‑available GANs, diffusion models, and forensic detectors.

Methodology

1. Shared Backbone – Both the generator G and detector D start from a common transformer‑style encoder that processes the latent code (for G) and the image (for D).
2. Multimodal Self‑Attention (MMSA) – At several depths, the model inserts an attention layer that receives queries from G and keys/values from D (and vice‑versa). This lets the generator “see” what the detector deems suspicious and lets the detector incorporate cues about the generation process.
3. Training Loop
   • Generation Phase: Sample a latent vector z, produce an image (\hat{x} = G(z)). Compute the usual adversarial loss (e.g., GAN or diffusion objective) plus a detectability penalty from D’s current prediction (the DIGA term).
   • Detection Phase: Feed a mixed batch of real images and generated images to D, compute a binary cross‑entropy loss, and back‑propagate through the shared backbone and the MMSA modules.
   • Fine‑Tuning: Alternate the two phases every few steps, using a small learning‑rate schedule to keep the shared parameters stable.
4. Loss Functions

L_gen = adversarial loss + λ_dig * L_DIGA
L_det = BCE(real/fake) + λ_att * L_MMSA   # regularization encouraging consistent attention maps

The overall system can be implemented with popular deep‑learning libraries (PyTorch, HuggingFace Transformers) and runs on a single GPU for moderate‑size datasets.

Results & Findings

Key takeaways

• The generator consistently outperforms strong baselines in terms of visual fidelity (lower FID).
• The detector reaches near‑perfect AUC on both synthetic and real‑world deep‑fake benchmarks, even when the generator is deliberately tuned to evade detection.
• Ablation studies confirm that removing MMSA or DIGA degrades both sides, highlighting the mutual benefit of the co‑training design.

Practical Implications

• Secure Content Pipelines – Platforms that need to both synthesize realistic assets (e.g., game art, virtual try‑ons) and guard against malicious deep‑fakes can adopt a single UniGenDet model, reducing engineering overhead.
• Rapid Prototyping – Developers can fine‑tune the shared backbone on their own domain data (e.g., medical imaging, fashion) and instantly obtain a generator tuned for realism and a detector calibrated for that specific style.
• Regulatory Compliance – Companies required to watermark or detect AI‑generated media can embed the detector component directly into their production stack, ensuring that any generated output passes an internal authenticity check before release.
• Research Acceleration – By exposing the generator to detection feedback early, researchers can iterate faster on new synthesis techniques without waiting for separate forensic evaluations.

Limitations & Future Work

• Scalability – Training the unified model on ultra‑high‑resolution images (>1024×1024) still demands multi‑GPU setups; the current implementation is optimized for 256–512 px.
• Domain Transfer – While the shared backbone generalizes across several datasets, extreme domain shifts (e.g., satellite imagery) may require additional modality‑specific adapters.
• Adversarial Arms Race – The co‑evolutionary setup assumes a cooperative training regime; in the wild, attackers could deliberately craft inputs that exploit the detector’s learned biases. Future work could explore robust adversarial training and continual learning to keep the detector ahead of novel generation tricks.

The authors have released their code on GitHub, making it straightforward for developers to experiment with UniGenDet in their own projects.

Authors

• Yanran Zhang
• Wenzhao Zheng
• Yifei Li
• Bingyao Yu
• Yu Zheng
• Lei Chen
• Jiwen Lu
• Jie Zhou

Paper Information

• arXiv ID: 2604.21904v1
• Categories: cs.CV
• Published: April 23, 2026
• PDF: Download PDF