Home▸AI & Machine Learning▸Cross-Modal Attention

3D Multimodal Embedding Space Simulator

Interactive 3D shared-embedding-space simulator: watch image-token and text-token points drift through a vector space and lock onto matching concept pairs as cross-modal alignment strength increases, with alignment strength, embedding noise and pair count tunable in real time.

AI & Machine Learning3DEasy60 FPS📱 Mobile-adapted⇄ 2D version

Beyond attending across modalities token-by-token, multimodal LLMs are trained so that images and text sharing a concept land close together in one shared vector space — the geometry that makes cross-modal retrieval and grounding possible. This simulation renders that space in 3D: image-token points and text-token points drift and jitter under embedding noise while a spring proportional to alignment strength pulls each true concept pair together, visualizing convergence the way a contrastive training objective would over time. Rotate and zoom to inspect how pairs cluster as alignment strength rises and noise falls.

⚙ Under the hood

In the 3D space, image-token and text-token points drift and are pulled towards each other in a shared embedding volume, illustrating contrastive-training convergence.

attentionmultimodalembedding

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install