Pareto-Enhanced Portrait Generation: Vision-Aligned Text Supervision for Alignment, Realism, and Aesthetics
The article discusses a new method for improving text-to-image diffusion models in human portrait generation. This method addresses the challenges of text-image alignment, photorealism, and aesthetics by introducing a feature supervision paradigm. The proposed approach enhances image generation without degrading the model's original capabilities.
- ▪Text-to-image diffusion models face a trilemma involving alignment, realism, and aesthetics.
- ▪The proposed method uses a lightweight cross-modal alignment mechanism to improve image generation.
- ▪Extensive experiments demonstrate that the method achieves improvements across all three challenging aspects.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20640 (cs) [Submitted on 20 May 2026] Title:Pareto-Enhanced Portrait Generation: Vision-Aligned Text Supervision for Alignment, Realism, and Aesthetics Authors:Yunlong Wang, Jinjin Shi, Wenbin Gao, Xuran Xu, Runyu Shi, Ying Huang View a PDF of the paper titled Pareto-Enhanced Portrait Generation: Vision-Aligned Text Supervision for Alignment, Realism, and Aesthetics, by Yunlong Wang and 5 other authors View PDF Abstract:Text-to-image diffusion models often face a severe trilemma in human portrait generation: text-image alignment, photorealism, and human-perceived aesthetics inherently inhibit one another. Supervised Fine-Tuning (SFT) is an effective method for enhancing the photorealism of image generation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
