Kakao Presents HG-DPO Paper for Person Image Generation at CVPR 2025

The Kakao Kanana team unveiled the HG-DPO study, which was selected as a highlight paper at CVPR 2025. This technology integrates DPO (Direct Preference Optimization), an existing LLM alignment technique, into diffusion-based person image generation models, helping to generate high-quality, photorealistic images by fine-tuning the Base model.

Real-World-Based Learning Strategy Existing Diffusion-DPO methods created preference pairs from generated images for training, but had the limitation of not reaching real-world quality. To overcome this, HG-DPO introduced a new training approach that uses real-world images as Winning images and generated images as Losing images. This guides the model to learn towards 'photorealistic quality' rather than just relative superiority.

3-Stage Curriculum Learning To address the instability of single-stage training, a Curriculum Learning approach was applied to gradually increase difficulty.

• Easy Stage: Forms basic preferences by training on Winning/Losing pairs with large preference differences evaluated by PickScore among generated images.
• Normal Stage: Learns the realism of composition and pose by using Intermediate images, created by adding Noise to real-world images and restoring them, as Winning images.
• Hard Stage: Performs final fine-tuning to match real-world details by using actual real-world images as Winning images.

Performance and Application HG-DPO demonstrated quantitative and qualitative superiority by ranking highly on major benchmark metrics such as P-Score, HPS, and FID. It is also effectively applied to Personalized text-to-image generation, enabling the creation of high-quality profile photos. This technology is being applied to the advancement of Kakao's proprietary image generation model, 'Kollage by Kanana'.