EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Fine-Grained Semantic Alignment Method for Text-to-Image Generation Based on Reordering-Guided Pseudo-Sample Distillation

Anyang Shen

ITM Web of Conferences · 2026 · 93 · 01002

DOI: 10.1051/itmconf/20269301002 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ITM Web of Conferences, online erschienen am 13.11.2026. Verfasst von Anyang Shen. Die Arbeit nennt 14 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Text-to-image diffusion models can generate images with high visual quality, but they remain unstable under fine-grained semantic constraints such as multiple objects, color binding, counting, and spatial relationships. This paper proposes a reordering-guided pseudo-sample distillation method for lightweight adaptation of Stable Diffusion 1.5. For each training prompt, multiple candidate images are first generated and then evaluated by detector-based rules. Semantically correct candidates are selected as pseudo image-text pairs and mixed with COCO general samples and fine-grained samples for UNet cross-attention LoRA training. Instead of simply increasing the amount of data, the method focuses on the quality and task distribution of pseudo samples. Experiments on GenEval show that the final variant V58c improves the average Overall score across three random seeds from 0.42181 to 0.42472. Although the improvement is limited, the method shows positive trends on two-object, colors, and counting subtasks. Ablation results further show that excessive pseudo samples, excessive general replay, and high-confidence filtering may weaken the compositional semantic training signal.

Zitieren

APA

Shen, A. (2026). A Fine-Grained Semantic Alignment Method for Text-to-Image Generation Based on Reordering-Guided Pseudo-Sample Distillation. ITM Web of Conferences, 93, 01002. https://doi.org/10.1051/itmconf/20269301002

Vancouver

Shen A. A Fine-Grained Semantic Alignment Method for Text-to-Image Generation Based on Reordering-Guided Pseudo-Sample Distillation. ITM Web of Conferences. 2026;93:01002. doi:10.1051/itmconf/20269301002

BibTeX
@article{Shen2026,
  author = {Shen, Anyang},
  title = {A Fine-Grained Semantic Alignment Method for Text-to-Image Generation Based on Reordering-Guided Pseudo-Sample Distillation},
  journal = {ITM Web of Conferences},
  year = {2026},
  volume = {93},
  pages = {01002},
  doi = {10.1051/itmconf/20269301002}
}
Kostenlos starten →