🚀 Introducing Science-T2I - Towards bridging the gap between AI imagination and scientific reality in image generation! [CVPR 2025] 📜 Paper: arxiv.org/abs/2504.13129 🌐 Project: jialuo-li.github.io/Science-… 💻 Code: github.com/Jialuo-Li/Science… 🤗 Dataset: huggingface.co/collections/J… 🔍 The Problem: Current text-to-image models (like Stable Diffusion/FLUX) often create scientifically implausible visuals (e.g., tomatoes floating unnaturally in water 🍅💧, copper that burns with yellow flame 🔥). These "science illusions" reveal a critical blind spot in generative AI! 🤔 Why It Matters: From education 🌍 to simulation 🔬, AI-generated content must obey real-world physics/chemistry. We' re towards making synthetic visuals not just pretty—but physically accurate! 💡 Our Solution: ✅ Science-T2I Dataset: Over 20K *expert-annotated* adversarial image pairs + 9K prompts across physics, chemistry & biology. ✅ SciScore: A novel reward model that evaluates images like a science teacher. ✅ Two-Stage Training: Fine-tunes models to align with scientific realism, boosting accuracy by 50%+. Check out more results and insights we conclude from our training in the thread below. 👇 #CVPR2025