Hugging Face's logo

Spaces:
yrtf89
/
manifoldcache
Running

App Files Community
manifoldcache
File size: 5,097 Bytes
3e10512
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bdfcb07
 
 
 
 
 
 
 
 
 
 
3e10512
 
42538b3
 
3e10512
 
 
 
aaa71fb
3e10512
 
 
 
 
 
58dde9b
3e10512
 
 
 
 
42538b3
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

    <!DOCTYPE html>
    <html lang="en">
    <head>
        <meta charset="UTF-8">
        <meta name="viewport" content="width=device-width, initial-scale=1.0">
        <title>ManifoldCache: Training-Free Diffusion Acceleration</title>
        <style>
            body {
                font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
                margin: 0;
                padding: 20px;
                background-color: #f9fafb;
                color: #111827;
            }
            .container {
                max-width: 1100px;
                margin: 0 auto;
            }
            h1 { font-size: 1.85rem; margin-bottom: 8px; color: #047857; }
            p { margin-top: 0; color: #4b5563; font-size: 0.95rem; line-height: 1.5; }
            .links-bar {
                margin: 15px 0 25px 0;
                display: flex;
                gap: 15px;
            }
            .btn {
                display: inline-block;
                padding: 8px 16px;
                border-radius: 6px;
                font-weight: 500;
                text-decoration: none;
                font-size: 0.9rem;
            }
            .btn-primary { background-color: #059669; color: white; }
            .btn-primary:hover { background-color: #047857; }
            .btn-secondary { background-color: #e5e7eb; color: #374151; }
            .btn-secondary:hover { background-color: #d1d5db; }
            .visual-preview {
                background: white;
                padding: 15px;
                border-radius: 8px;
                border: 1px solid #e5e7eb;
                margin-bottom: 25px;
                text-align: center;
            }
            .visual-preview img {
                max-width: 100%;
                height: auto;
                border-radius: 4px;
            }
            .iframe-container {
                position: relative;
                width: 100%;
                height: 900px;
                border-radius: 8px;
                overflow: hidden;
                border: 1px solid #e5e7eb;
                box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.05);
            }
            iframe {
                width: 100%;
                height: 100%;
                border: none;
            }
        </style>
    </head>
    <body>
    
    <div class="container">
        <h1>ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching</h1>
        <!-- <p>
        </p> -->
      <p class="project-description">
  <strong>ManifoldCache</strong> delivers acceleration for scientific diffusion models through:
  <strong>(a)</strong> a unified CMDM abstraction covering several models across five domains, including brain MRI, molecules, proteins, crystals, and 4D scenes spanning radically different ambient spaces from voxel grids to motif token graphs to crystal lattices to 4D video latents under one closed-form schedule;
  <strong>(b)</strong> true backbone-agnosticism across fundamentally different families, including volumetric 3D ConvNets, SE(3)/E(3)-equivariant networks, graph diffusion transformers, and multi-view video DiTs including discrete diffusion;
  <strong>(c)</strong> systematic failure analysis showing quantization causes OOM, pruning breaks constraints, fast ODE solvers drift off the manifold, and standard caching induces mode confusion;
  <strong>(d)</strong> a provably sharp safe-caching threshold T* with both bounded-error and guaranteed-confusion regimes confirmed by ablations;
  <strong>(e)</strong> depth-adaptive caching from Jacobian decomposition, where deeper blocks get larger strides with bounded and competitive VRAM overhead; and
  <strong>(f)</strong> completely training-free and data-free with zero calibration, zero retraining, working out-of-the-box on existing pretrained checkpoints while delivering consistent joint dominance on both speed and quality across all models.
       </p>
    
        <div class="links-bar">
            <a href="https://openreview.net/forum?id=4P0EjrDCr1" target="_blank" class="btn btn-primary">📄 Read Paper (OpenReview)</a>
            <a href="https://yrtf89.github.io/manifoldcache.github.io/index.html" target="_blank" class="btn btn-secondary">🌐 Open Demo Website</a>
        </div>
    
        <!-- GIF Banner Section -->
        <div class="visual-preview">
            <!-- <p style="font-weight: 600; margin-bottom: 10px; color: #374151;">Acceleration Output Preview (Constraint-Manifold Preservation)</p> -->
            <img src="chrome-capture-2026-06-19%20(3).gif" alt="ManifoldCache 3D Trajectory Optimization Pipeline">
        </div>
    
        <!-- Embedded Live Interactive Portfolio -->
        <h2 style="font-size: 1.3rem; margin-bottom: 12px; color: #374151;">Interactive 3D Visualizations & Benchmarks</h2>
        <div class="iframe-container">
            <iframe src="https://yrtf89.github.io/manifoldcache.github.io/index.html" title="ManifoldCache Project Web Demo"></iframe>
        </div>
    </div>
    
    </body>
    </html>