Improving text-to-image generation without retraining
Abstract
Text-to-image (T2I) models have achieved strong visual realism, but they still struggle when prompts require world knowledge. In this work, we study how to enhance T2I generation without retraining large-scale diffusion models. We investigate three strategies: encoder routing, prompt rewriting using a large language model, and best prompt selection through a feature fusion scoring scheme, based on Stable Diffusion 3.5-Large as our base model. The evaluation is conducted on the WISE benchmark, which is designed to measure the semantic understanding of T2I models via three categories Consistency, Realism, and Aesthetic quality. The results show that our proposed strategies can significantly improve the performance of Stable Diffusion 3.5-Large on the WISE benchmark, which enhances the semantic capability of T2I generation without retraining the model.
Read the paper
Related studies
- Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2)
- Commentary on the clinical and preclinical dosage limits of interstitially administered magnetic fluids for therapeutic hyperthermia based on current practice and efficacy models
- Prosthesis: New Technological Opportunities and Innovative Biomedical Devices
- The history and future of neural modeling for cochlear implants
- Ryanodine receptors: waking up from refractoriness
- Novel Transgenic Rabbit Model Sheds Light on the Puzzling Role of Matrix Metalloproteinase-12 in Atherosclerosis
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.