File size: 5,265 Bytes
ce35403 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 | import { Arrow, FigBox, FILL, LoopFigure, type LoopStop } from "./LoopFigure";
const STOPS: LoopStop[] = [
{
x: 160,
y: 110,
ids: ["vision"],
inspect: "vision",
title: "Image patches",
note: "The image is cut into patches. This is the new front-end; the language model still has not seen pixels.",
},
{
x: 160,
y: 170,
ids: ["vision"],
inspect: "vision",
title: "ViT encoder",
note: "A vision Transformer (often CLIP-ViT) turns patches into a sequence of embeddings.",
},
{
x: 160,
y: 270,
ids: ["projector"],
inspect: "projector",
title: "Projector W",
note: "A linear layer or small MLP maps vision width onto the LLM hidden size.",
},
{
x: 500,
y: 170,
ids: ["decoder"],
inspect: "decoder",
title: "Text tokens",
note: "The prompt is embedded as usual. Vision tokens will sit in the same residual stream.",
},
{
x: 330,
y: 360,
ids: ["decoder", "projector"],
inspect: "decoder",
title: "Concat [vision ; text]",
note: "z = LLM([W · ViT(image); text]). Cross-attention variants keep the ViT as a separate memory.",
},
{
x: 330,
y: 450,
ids: ["decoder"],
inspect: "decoder",
title: "LLM decoder → logits",
note: "The top layer is still next-token prediction. Once vision tokens are in context, the same stack can caption or follow instructions.",
},
];
export function VlmLoop({ onSelect }: { onSelect: (id: string) => void }) {
return (
<LoopFigure
stops={STOPS}
loopMs={15000}
viewBox="0 0 680 520"
ariaLabel="Vision-language model architecture"
onSelect={onSelect}
>
{({ on, markerId }) => (
<g>
<rect width="680" height="520" fill="#f7f4ec" />
<text x="340" y="28" textAnchor="middle" fontSize="16" fontWeight="700" fill="#222">
Vision-Language Model
</text>
<text x="340" y="48" textAnchor="middle" fontSize="12" fill="#555">
z = LLM([W · ViT(image); text])
</text>
<text x="160" y="78" textAnchor="middle" fontSize="12" fill="#555">
Vision
</text>
<g>
{[0, 1, 2].map((r) =>
[0, 1, 2].map((c) => (
<rect
key={`${r}-${c}`}
x={118 + c * 28}
y={86 + r * 16}
width="24"
height="13"
rx="2"
fill={on("vision") ? FILL.green : "#cfe8d4"}
stroke="#3a7a4a"
/>
)),
)}
</g>
<Arrow x1={160} y1={140} x2={160} y2={148} active={on("vision")} markerId={markerId} />
<FigBox
x={70}
y={148}
w={180}
h={48}
fill={FILL.cyan}
lit={on("vision")}
label="ViT encoder"
onClick={() => onSelect("vision")}
/>
<Arrow x1={160} y1={196} x2={160} y2={242} active={on("vision") || on("projector")} markerId={markerId} />
<FigBox
x={70}
y={248}
w={180}
h={44}
fill={FILL.yellow}
lit={on("projector")}
label="Projector W"
onClick={() => onSelect("projector")}
/>
<text x="500" y="88" textAnchor="middle" fontSize="12" fill="#555">
Language
</text>
<FigBox
x={410}
y={148}
w={180}
h={48}
fill={FILL.pink}
lit={on("decoder")}
label="Text tokens"
onClick={() => onSelect("decoder")}
/>
<Arrow x1={500} y1={196} x2={500} y2={248} active={on("decoder")} markerId={markerId} />
<FigBox
x={410}
y={248}
w={180}
h={44}
fill={FILL.orange}
lit={on("decoder")}
label="Text embed"
onClick={() => onSelect("decoder")}
/>
<path
d="M160 292 V 330 H 330"
fill="none"
stroke={on("projector") || on("decoder") ? "#111" : "#888"}
strokeWidth={on("projector") ? 2.2 : 1.2}
markerEnd={`url(#${markerId})`}
/>
<path
d="M500 292 V 330 H 330"
fill="none"
stroke={on("decoder") ? "#111" : "#888"}
strokeWidth={on("decoder") ? 2.2 : 1.2}
markerEnd={`url(#${markerId})`}
/>
<FigBox
x={220}
y={336}
w={220}
h={48}
fill={FILL.purple}
lit={on("decoder")}
label="[vision tokens ; text]"
onClick={() => onSelect("decoder")}
/>
<Arrow x1={330} y1={384} x2={330} y2={418} active={on("decoder")} markerId={markerId} />
<FigBox
x={200}
y={426}
w={260}
h={52}
fill={FILL.orange}
lit={on("decoder")}
label="LLM decoder → logits"
onClick={() => onSelect("decoder")}
/>
</g>
)}
</LoopFigure>
);
}
|