Spaces:

xjodoin
/

moge3-webgpu

Running

Publish GPU autotuning and real-photo visual benchmarks

996fbaa verified 13 days ago

4.15 kB

<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>MoGe-3 WebGPU experiment</title><style>
*{box-sizing:border-box}body{font:16px system-ui;background:#f6f4ee;color:#24483e;max-width:1100px;padding:28px;margin:auto}h1{font-size:clamp(30px,5vw,50px);letter-spacing:-2px;margin:12px 0}p{line-height:1.55;color:#51665e}.badge{font-size:12px;text-transform:uppercase;letter-spacing:2px}.toolbar{display:flex;gap:12px;flex-wrap:wrap;align-items:center;margin:24px 0}button,select{padding:14px;border-radius:8px;font:inherit}button{background:#24483e;color:white;border:0;cursor:pointer}button:disabled{opacity:.5;cursor:wait}select{background:white;border:1px solid #b5c6bc}input{max-width:250px}.views{display:grid;grid-template-columns:1fr 2fr;gap:16px}.view{min-width:0;background:#dfe6df;border-radius:12px;overflow:hidden}.view h2{font-size:13px;padding:12px 16px;margin:0}img{width:100%;aspect-ratio:1.5;object-fit:contain}canvas{width:100%;height:360px;display:block}pre{background:#e9eee7;border-radius:12px;padding:16px;font-size:12px;max-height:340px;overflow:auto;white-space:pre-wrap}details{margin-top:24px}.note{font-size:13px}@media(max-width:680px){.views{grid-template-columns:1fr}img{max-height:260px}canvas{height:300px}}
#weight-mode-control{display:grid;gap:6px;max-width:100%}#weight-mode-control[hidden]{display:none}#weights{max-width:100%}</style>  <script type="module" crossorigin src="/assets/index-DW8o3Ijx.js"></script>
</head><body><span class="badge">Open-source browser research prototype</span><h1>A photo, reconstructed in your browser.</h1><p>MoGe-3 with its learned sparse 3D refinement, running on WebGPU. Voxel indexing and camera recovery run on your GPU. Your photo stays on this device.</p><div class="toolbar"><select id="profile" aria-label="Resolution"><option value="fast">512 · Fast</option><option value="detail">768 · Detail</option></select><select id="steps" aria-label="Refinement passes"><option value="1">1 pass · Fastest</option><option value="2">2 passes</option><option value="3">3 passes · Original default</option></select><select id="case" aria-label="Test photo"><option value="sample-backyard">Backyard</option><option value="brick-ranch">Brick ranch</option><option value="raised-deck">Raised deck</option></select><label id="weight-mode-control" hidden>Weights <select id="weights" aria-label="Weight download"><option value="compact">Smaller download · rounded weights</option><option value="exact">Original weights · exact upgrade</option></select></label><label>Your photo <input id="upload" type="file" accept="image/*"></label><button id="run">Run with refinement</button><button id="unload">Unload model</button><button id="clear-cache">Clear downloaded model</button></div><p class="note">Experimental landscape profiles · 1200–1733 tokens · 1–3 refinement passes · Downloaded model files saved on this device when storage is available. Higher resolutions require a capable GPU and several GB of available memory. Estimated measurements need real-world calibration.</p><div class="views"><div class="view"><h2>Input photo</h2><img id="photo" crossorigin="anonymous" src="/models/profiles/fast/cases/sample-backyard/photo.jpg" alt="Input to the reconstruction"></div><div class="view"><h2>Refined point cloud · drag to orbit, scroll to zoom</h2><canvas id="scene"></canvas></div></div><p class="note">Real-photo exact-weight validation: 34/36 cases pass; two Detail multipass cases differ from original CUDA. <a href="/benchmarks/real-visual/index.html">View depth comparisons</a>.</p><details open><summary>Timing and numerical validation</summary><p class="note">Sample accuracy is checked against original MoGe-3 on NVIDIA in strict FP32. MLX comparisons remain diagnostic. Uploaded photos have no reference validation.</p><pre id="status">Ready. The first run downloads shared model files and saves them on this device when storage is available. Later visits reuse saved downloads. Unload releases GPU memory; Clear downloaded model removes saved weights.</pre></details></body></html>

	<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>MoGe-3 WebGPU experiment</title><style>
	*{box-sizing:border-box}body{font:16px system-ui;background:#f6f4ee;color:#24483e;max-width:1100px;padding:28px;margin:auto}h1{font-size:clamp(30px,5vw,50px);letter-spacing:-2px;margin:12px 0}p{line-height:1.55;color:#51665e}.badge{font-size:12px;text-transform:uppercase;letter-spacing:2px}.toolbar{display:flex;gap:12px;flex-wrap:wrap;align-items:center;margin:24px 0}button,select{padding:14px;border-radius:8px;font:inherit}button{background:#24483e;color:white;border:0;cursor:pointer}button:disabled{opacity:.5;cursor:wait}select{background:white;border:1px solid #b5c6bc}input{max-width:250px}.views{display:grid;grid-template-columns:1fr 2fr;gap:16px}.view{min-width:0;background:#dfe6df;border-radius:12px;overflow:hidden}.view h2{font-size:13px;padding:12px 16px;margin:0}img{width:100%;aspect-ratio:1.5;object-fit:contain}canvas{width:100%;height:360px;display:block}pre{background:#e9eee7;border-radius:12px;padding:16px;font-size:12px;max-height:340px;overflow:auto;white-space:pre-wrap}details{margin-top:24px}.note{font-size:13px}@media(max-width:680px){.views{grid-template-columns:1fr}img{max-height:260px}canvas{height:300px}}
	#weight-mode-control{display:grid;gap:6px;max-width:100%}#weight-mode-control[hidden]{display:none}#weights{max-width:100%}</style> <script type="module" crossorigin src="/assets/index-DW8o3Ijx.js"></script>
	</head><body><span class="badge">Open-source browser research prototype</span><h1>A photo, reconstructed in your browser.</h1><p>MoGe-3 with its learned sparse 3D refinement, running on WebGPU. Voxel indexing and camera recovery run on your GPU. Your photo stays on this device.</p><div class="toolbar"><select id="profile" aria-label="Resolution"><option value="fast">512 · Fast</option><option value="detail">768 · Detail</option></select><select id="steps" aria-label="Refinement passes"><option value="1">1 pass · Fastest</option><option value="2">2 passes</option><option value="3">3 passes · Original default</option></select><select id="case" aria-label="Test photo"><option value="sample-backyard">Backyard</option><option value="brick-ranch">Brick ranch</option><option value="raised-deck">Raised deck</option></select><label id="weight-mode-control" hidden>Weights <select id="weights" aria-label="Weight download"><option value="compact">Smaller download · rounded weights</option><option value="exact">Original weights · exact upgrade</option></select></label><label>Your photo <input id="upload" type="file" accept="image/*"></label><button id="run">Run with refinement</button><button id="unload">Unload model</button><button id="clear-cache">Clear downloaded model</button></div><p class="note">Experimental landscape profiles · 1200–1733 tokens · 1–3 refinement passes · Downloaded model files saved on this device when storage is available. Higher resolutions require a capable GPU and several GB of available memory. Estimated measurements need real-world calibration.</p><div class="views"><div class="view"><h2>Input photo</h2><img id="photo" crossorigin="anonymous" src="/models/profiles/fast/cases/sample-backyard/photo.jpg" alt="Input to the reconstruction"></div><div class="view"><h2>Refined point cloud · drag to orbit, scroll to zoom</h2><canvas id="scene"></canvas></div></div><p class="note">Real-photo exact-weight validation: 34/36 cases pass; two Detail multipass cases differ from original CUDA. <a href="/benchmarks/real-visual/index.html">View depth comparisons</a>.</p><details open><summary>Timing and numerical validation</summary><p class="note">Sample accuracy is checked against original MoGe-3 on NVIDIA in strict FP32. MLX comparisons remain diagnostic. Uploaded photos have no reference validation.</p><pre id="status">Ready. The first run downloads shared model files and saves them on this device when storage is available. Later visits reuse saved downloads. Unload releases GPU memory; Clear downloaded model removes saved weights.</pre></details></body></html>