Two new research papers address the challenge of diversity in text-to-image generation models, particularly when using long and complex prompts. The first paper introduces TPSO (Token-Prompt Embedding Space Optimization), a training-free module that enhances diversity by exploring underrepresented regions of the token embedding space while maintaining image quality through semantic constraints. The second paper proposes PromptMoG, a method that samples prompt embeddings from a Mixture-of-Gaussians distribution to restore generative flexibility for long prompts, supported by a new benchmark called LPD-Bench for evaluating diversity and fidelity. Both approaches aim to prevent repetitive outputs and improve creative exploration in diffusion models. AI
IMPACT These methods could lead to more creative and varied outputs from text-to-image models, benefiting artists and designers.
RANK_REASON Two research papers published on arXiv introducing new methods for improving text-to-image generation diversity.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
