mjsushanth.github.io/text-to-pose.html at main · mjsushanth/mjsushanth.github.io · GitHub

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
<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>Text-to-Pose Diffusion - Joel Markapudi</title>
    <style>
        * {
            margin: 0;
            padding: 0;
            box-sizing: border-box;
        }

        body {
            font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, sans-serif;
            line-height: 1.6;
            color: #333;
            background: #f5f5f0;
        }

        .container {
            max-width: 920px;
            margin: 0 auto;
            padding: 50px 30px 40px 30px;
            background: #fafaf5;
            min-height: 100vh;
        }

        .back-link {
            display: inline-flex;
            align-items: center;
            color: #4a90e2;
            text-decoration: none;
            margin-bottom: 30px;
            font-size: 0.95rem;
            transition: color 0.3s;
        }

        .back-link:hover {
            color: #357abd;
        }

        h1 {
            font-size: 2.2rem;
            font-weight: 400;
            margin-bottom: 18px;
            color: #2c2c2c;
            line-height: 1.3;
        }

        .project-overview {
            color: #555;
            font-size: 1.05rem;
            margin-bottom: 25px;
            padding-bottom: 20px;
            border-bottom: 2px solid #e0e0d8;
            line-height: 1.7;
        }

        .highlights {
            margin-top: 25px;
        }

        .highlights ul {
            list-style: none;
            margin: 0;
            padding: 0;
        }

        .highlights li {
            margin-bottom: 16px;
            padding-left: 20px;
            position: relative;
            color: #555;
            line-height: 1.7;
        }

        .highlights li:before {
            content: "•";
            position: absolute;
            left: 0;
            color: #4a90e2;
            font-weight: bold;
            font-size: 1.2rem;
        }

        strong {
            color: #2c2c2c;
            font-weight: 500;
        }

        .external-links {
            margin-top: 35px;
            padding-top: 25px;
            border-top: 1px solid #e0e0d8;
        }

        .external-links h2 {
            font-size: 1.3rem;
            font-weight: 500;
            margin-bottom: 15px;
            color: #2c2c2c;
        }

        .external-links a {
            display: inline-block;
            margin-right: 20px;
            margin-bottom: 10px;
            color: #4a90e2;
            text-decoration: none;
            font-size: 1rem;
            transition: color 0.3s;
        }

        .external-links a:hover {
            color: #357abd;
            text-decoration: underline;
        }

        @media (max-width: 768px) {
            .container {
                padding: 35px 20px;
            }

            h1 {
                font-size: 1.8rem;
            }
        }
    </style>
</head>
<body>
    <div class="container">
        <a href="index.html" class="back-link">← Back to Portfolio</a>

        <h1>Text-to-Pose Diffusion: CLIP-Conditioned 3D Pose Synthesis</h1>

        <div class="project-overview">
            Cross-modal diffusion model that maps free-form text into 3D human poses, integrating CLIP semantics with a UNet-based diffusion backbone. Focuses on anatomically consistent skeletons, kinematic chain reasoning, and pose–text alignment rather than pretty images, built in PyTorch with CLIP as the semantic encoder.
        </div>

        <div class="highlights">
            <ul>
                <li><strong>Cross-Modal Architecture:</strong> Hybrid CNN–Transformer UNet where CLIP text embeddings are projected into pose-space and injected via cross-attention at multiple resolution levels, enabling the model to condition on high-level actions (e.g., “running”, “reaching”) and finer body-part cues simultaneously.</li>

                <li><strong>Specialized Cross-Attention:</strong> Hierarchical semantic conditioning that routes different aspects of the text into different spatial scales: coarse layers focus on global posture and balance, while mid/fine layers emphasize limb orientation, hand/foot placement, and spatial relationships between joints.</li>

                <li><strong>Biomechanics & Physical Plausibility:</strong> Enforces bone-length consistency, joint-limit style constraints, and N-joint kinematic chain validation (pelvis–spine–extremities) with auxiliary anatomical losses and center-of-mass sanity checks to down-weight physically implausible skeletons.</li>

                <li><strong>Guidance & Training Dynamics:</strong> Uses dual-pass classifier-free guidance (CFG) to trade off diversity vs. adherence to the text prompt, with ablations over guidance strength, noise schedules, and conditioning dropout to study stability, mode-collapse behavior, and semantic fidelity.</li>

                <li><strong>Tooling & Evaluation:</strong> Includes dataset EDA and joint-mapping utilities, pose visualizers for qualitative inspection, and experiment scaffolding to compare baselines (unconditioned / weak conditioning) against the full CLIP-conditioned model under identical sampling budgets.</li>
            </ul>
        </div>

        <div class="external-links">
            <h2>Project Resources</h2>
            <a href="https://github.com/mjsushanth/CLIP-Conditioned-Diffusion-T2Pose-Generation/blob/main/DESIGN_README.md" target="_blank">→ Design Documentation</a>
            <a href="https://github.com/mjsushanth/CLIP-Conditioned-Diffusion-T2Pose-Generation/blob/main/RESEARCH_README.md" target="_blank">→ Research Report</a>
            <a href="https://github.com/mjsushanth/CLIP-Conditioned-Diffusion-T2Pose-Generation" target="_blank">→ GitHub Repository</a>
        </div>
    </div>
</body>
</html>