T2I Generation

Text-to-image generation studies how models turn natural-language instructions into images, primarily using diffusion and flow-matching models, autoregressive visual-token models, and unified multimodal systems. Researchers improve prompt following through better text and visual representations, spatial planning, reference-image conditioning, and post-training with preference or rubric-based rewards. Benchmarks test compositional alignment, counting, text rendering, world knowledge, and visual quality, while audits examine bias, privacy, and harmful outputs. A central challenge is making generations reliably match complex instructions and real-world constraints without sacrificing diversity, image quality, efficiency, or safety.