Legacy Concept Lab

Multimodal Foundations: Vision Encoders, Contrastive Learning & Cross-Attention Fusion

CLIP-style encoders are front-end for multimodal LLMs (how images become tokens) and conditioning backbone for text-to-image diffusion

Concept 32 of 100RepresentationsPhase 8
#32Multimodal VLPRepresentations
key equation
L=logesiijesijlogesiijesji\mathcal{L} = -\log\frac{e^{s_{ii}}}{\sum_j e^{s_{ij}}} - \log\frac{e^{s_{ii}}}{\sum_j e^{s_{ji}}}
Phase 8: Scaling, theory & multimodalConcept 32 of 100

Why It Matters for Modern Models

  • CLIP-style encoders are front-end for multimodal LLMs (how images become tokens) and conditioning backbone for text-to-image diffusion
  • SigLIP 2 (2025): unified training recipe extending image-text with captioning, self-supervised losses, online curation—better encoders for VLMs
  • CLOC (2024): adds region-level objectives for localization/dense features—contrastive ≠ dense understanding, need explicit losses
  • After diffusion (#9) and representations (#10), this completes the bridge: how text embeddings condition generative vision models
  • Opens multimodal arc: sets up VLM safety/robustness, grounding evals, multimodal RLHF as natural next concepts

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Contrastive ≠ dense understanding—CLIP-like training yields global semantics, localization/dense features require extra losses or architecture
  • Negatives and batch size matter—contrastive training shaped by similarity matrix, diagonal dominance emerges from large batches
  • Fusion choice is capability choice—dual-encoder retrieval is cheap, cross-attention fusion enables richer grounding but costs compute
  • Temperature τ controls sharpness—low τ makes softmax peaked (hard negatives), high τ smooths (easier learning but less discriminative)
  • Shared semantic space is learned geometry—images and text don't naturally align, contrastive loss pulls matched pairs together, pushes mismatches apart

Interactive Visualization

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and the visualization. Come back here for the full walkthrough.

Key Equation
L=logesiijesijlogesiijesji\mathcal{L} = -\log\frac{e^{s_{ii}}}{\sum_j e^{s_{ij}}} - \log\frac{e^{s_{ii}}}{\sum_j e^{s_{ji}}}

Vision-Language Pretraining (VLP) trains image encoder + text encoder so images and text land in a shared space, enabling multimodal LLMs and diffusion conditioning.

Dual-encoder similarity (normalized embeddings):

ui=fI(Ii)fI(Ii),vj=fT(Tj)fT(Tj),sij=uivjτu_i = \frac{f_I(I_i)}{|f_I(I_i)|}, \quad v_j = \frac{f_T(T_j)}{|f_T(T_j)|}, \quad s_{ij} = \frac{u_i^\top v_j}{\tau}

Temperature-scaled cosine similarity between image and text embeddings.

CLIP-style contrastive loss (InfoNCE / symmetric cross-entropy):

L=12(logesiijesijlogesiijesji)\mathcal{L} = \frac{1}{2}\left( -\log\frac{e^{s_{ii}}}{\sum_j e^{s_{ij}}} - \log\frac{e^{s_{ii}}}{\sum_j e^{s_{ji}}} \right)

Maximize diagonal (matched pairs), minimize off-diagonal (negatives)—contrastive learning.

Cross-attention fusion (text attends to vision tokens):

CrossAttn(HT,HI)=softmax ⁣((HTWQ)(HIWK)d)(HIWV)\text{CrossAttn}(H_T, H_I) = \text{softmax}\!\left(\frac{(H_TW_Q)(H_IW_K)^\top}{\sqrt{d}}\right)(H_IW_V)

Enables grounding—text tokens can attend to image patches for dense understanding.

Canonical Papers

Learning Transferable Visual Models From Natural Language Supervision

Radford et al.2021ICML
Read paper →

SigLIP 2: Multilingual Vision-Language Encoders

Tschannen et al.2025arXiv
Read paper →

Contrastive Localized Language-Image Pre-Training

Chen et al.2024arXiv
Read paper →

Connections

Next Moves

Explore this concept from different angles — like a mathematician would.