In　this　paper,　we　propose　a　novel　technique　for　zero-shot　generation　of　3　-dimension　models　using　only　target　text　prompt.　This　paper　builds　upon　existing　research　results　by　utilizing　a　pre-trained　CLIP　model　as　the　core,　which　bridges　the　gap　between　text　and　visual　entities　by　comparing　the　semantic　similarity　between　the　input　text　prompt　and　the　3D　model　renderings.　Furthermore,　we　integrate　a　diffusion　model　that　generates　corresponding　CLIP　image　embeddings　based　on　the　text　embedding　information.　This　integration　enables　the　generation　of　more　faithful　results　to　the　text　prompt　in　the　stage　of　predicting　the　color　and　local　geometric　details　that　match　the　target　text　prompt.　©　2024　IEEE.