Ethanfel
3953d97163
Includes algorithm comparison, node wiring, and parameter guide diagrams. SVGs use <picture> tags for GitHub compatibility. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
88 lines
6.4 KiB
XML
88 lines
6.4 KiB
XML
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 880 520">
|
|
<defs>
|
|
<linearGradient id="abg" x1="0%" y1="0%" x2="100%" y2="100%">
|
|
<stop offset="0%" style="stop-color:#0d1117"/>
|
|
<stop offset="100%" style="stop-color:#161b22"/>
|
|
</linearGradient>
|
|
</defs>
|
|
<rect width="880" height="520" rx="12" fill="url(#abg)"/>
|
|
|
|
<!-- Title -->
|
|
<text x="440" y="38" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="20" font-weight="700" fill="#e6edf3" text-anchor="middle">Algorithm: Standard CFG vs SMC-CFG</text>
|
|
|
|
<!-- Left: Standard CFG -->
|
|
<rect x="30" y="60" width="390" height="430" rx="10" fill="#1c2333" stroke="#30363d" stroke-width="1.5"/>
|
|
<text x="225" y="90" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="16" font-weight="600" fill="#f0883e" text-anchor="middle">Standard CFG (P-Control)</text>
|
|
|
|
<!-- Step boxes - Standard -->
|
|
<rect x="55" y="110" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="75" y="127" font-family="monospace" font-size="11" fill="#7ee787">1.</text>
|
|
<text x="95" y="127" font-family="monospace" font-size="11" fill="#c9d1d9">Compute conditional prediction</text>
|
|
<text x="95" y="145" font-family="monospace" font-size="12" font-weight="600" fill="#79c0ff">v_cond = model(x_t, t, prompt)</text>
|
|
|
|
<rect x="55" y="164" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="75" y="181" font-family="monospace" font-size="11" fill="#7ee787">2.</text>
|
|
<text x="95" y="181" font-family="monospace" font-size="11" fill="#c9d1d9">Compute unconditional prediction</text>
|
|
<text x="95" y="199" font-family="monospace" font-size="12" font-weight="600" fill="#79c0ff">v_uncond = model(x_t, t, "")</text>
|
|
|
|
<rect x="55" y="218" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="75" y="235" font-family="monospace" font-size="11" fill="#7ee787">3.</text>
|
|
<text x="95" y="235" font-family="monospace" font-size="11" fill="#c9d1d9">Guidance error (fixed gain)</text>
|
|
<text x="95" y="253" font-family="monospace" font-size="12" font-weight="600" fill="#79c0ff">e_t = v_cond - v_uncond</text>
|
|
|
|
<rect x="55" y="272" width="340" height="44" rx="6" fill="#21262d" stroke="#f85149" stroke-width="1.5"/>
|
|
<text x="75" y="289" font-family="monospace" font-size="11" fill="#7ee787">4.</text>
|
|
<text x="95" y="289" font-family="monospace" font-size="11" fill="#f85149">Linear combination (can diverge!)</text>
|
|
<text x="95" y="307" font-family="monospace" font-size="12" font-weight="600" fill="#ff7b72">v_out = v_uncond + w * e_t</text>
|
|
|
|
<!-- Problem callout -->
|
|
<rect x="55" y="340" width="340" height="70" rx="6" fill="#2d1117" stroke="#f85149" stroke-width="1"/>
|
|
<text x="75" y="362" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="12" font-weight="600" fill="#f85149">Problem at high CFG scales:</text>
|
|
<text x="75" y="382" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="11" fill="#f0883e">Unbounded linear gain causes oversaturation,</text>
|
|
<text x="75" y="398" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="11" fill="#f0883e">artifacts, and semantic drift</text>
|
|
|
|
<!-- Right: SMC-CFG -->
|
|
<rect x="460" y="60" width="390" height="430" rx="10" fill="#1c2333" stroke="#30363d" stroke-width="1.5"/>
|
|
<text x="655" y="90" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="16" font-weight="600" fill="#7ee787" text-anchor="middle">SMC-CFG (Sliding Mode Control)</text>
|
|
|
|
<!-- Step boxes - SMC -->
|
|
<rect x="485" y="110" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="505" y="127" font-family="monospace" font-size="11" fill="#7ee787">1.</text>
|
|
<text x="525" y="127" font-family="monospace" font-size="11" fill="#c9d1d9">Same: compute v_cond, v_uncond</text>
|
|
<text x="525" y="145" font-family="monospace" font-size="12" font-weight="600" fill="#79c0ff">e_t = v_cond - v_uncond</text>
|
|
|
|
<rect x="485" y="164" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="505" y="181" font-family="monospace" font-size="11" fill="#7ee787">2.</text>
|
|
<text x="525" y="181" font-family="monospace" font-size="11" fill="#c9d1d9">Compute sliding surface</text>
|
|
<text x="525" y="199" font-family="monospace" font-size="12" font-weight="600" fill="#d2a8ff">s_t = (e_t - e_{t-1}) + lambda * e_{t-1}</text>
|
|
|
|
<rect x="485" y="218" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="505" y="235" font-family="monospace" font-size="11" fill="#7ee787">3.</text>
|
|
<text x="525" y="235" font-family="monospace" font-size="11" fill="#c9d1d9">Nonlinear switching control</text>
|
|
<text x="525" y="253" font-family="monospace" font-size="12" font-weight="600" fill="#d2a8ff">u_sw = -K * sign(s_t)</text>
|
|
|
|
<rect x="485" y="272" width="340" height="44" rx="6" fill="#21262d" stroke="#30363d"/>
|
|
<text x="505" y="289" font-family="monospace" font-size="11" fill="#7ee787">4.</text>
|
|
<text x="525" y="289" font-family="monospace" font-size="11" fill="#c9d1d9">Bounded correction</text>
|
|
<text x="525" y="307" font-family="monospace" font-size="12" font-weight="600" fill="#d2a8ff">e_corrected = e_t + u_sw</text>
|
|
|
|
<rect x="485" y="326" width="340" height="44" rx="6" fill="#21262d" stroke="#7ee787" stroke-width="1.5"/>
|
|
<text x="505" y="343" font-family="monospace" font-size="11" fill="#7ee787">5.</text>
|
|
<text x="525" y="343" font-family="monospace" font-size="11" fill="#7ee787">Stable guided output</text>
|
|
<text x="525" y="361" font-family="monospace" font-size="12" font-weight="600" fill="#7ee787">v_out = v_uncond + w * e_corrected</text>
|
|
|
|
<!-- Solution callout -->
|
|
<rect x="485" y="394" width="340" height="70" rx="6" fill="#0d2818" stroke="#238636" stroke-width="1"/>
|
|
<text x="505" y="416" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="12" font-weight="600" fill="#7ee787">Correction bounded to [-K, +K]:</text>
|
|
<text x="505" y="436" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="11" fill="#56d364">Prevents divergence at any CFG scale.</text>
|
|
<text x="505" y="452" font-family="Segoe UI, Helvetica, Arial, sans-serif" font-size="11" fill="#56d364">Lyapunov-stable finite-time convergence.</text>
|
|
|
|
<!-- Arrow between panels -->
|
|
<path d="M 425 275 L 455 275" stroke="#484f58" stroke-width="2" fill="none" marker-end="url(#arrowhead)"/>
|
|
<defs>
|
|
<marker id="arrowhead" markerWidth="8" markerHeight="6" refX="8" refY="3" orient="auto">
|
|
<polygon points="0 0, 8 3, 0 6" fill="#484f58"/>
|
|
</marker>
|
|
</defs>
|
|
</svg>
|