Why it matters

VLM design shapes capability. Understanding shapes selection.

Advertisement

The architecture

LLaVA: MLP adapter, vision tokens prefix.

Flamingo: gated cross-attention layers.

VLM architecturesLLaVAadapter + prefixFlamingocross-attentionIDEFICS / Molmohybrid variantsLLaVA simpler; Flamingo more expressive; hybrids common now
VLM types.
Advertisement

How it works end to end

LLaVA: simple + effective.

Flamingo: gated cross-attn, more expressive.

Molmo: pixel-level control.

Open source: LLaVA, IDEFICS, Molmo.