PageSourceSearch

https://guochengqian.github.io/project/magic123/

html guochengqian.github.io collected 2026-10-03 09:15:50 UTC 51,478 bytes, 406 lines download raw bytes

1<!doctype html><html lang=en-us><head><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta name=google-site-verification content="7MghMiEezWCdLZfbrgYR845yxCJj-g02bSAnoythP1A"><meta http-equiv=X-UA-Compatible content="IE=edge"><meta name=generator content="Source Themes Academic 4.5.0"><meta name=author content="Gordon Qian"><meta name=description content='<!DOCTYPE html>
2
3
4
5    
6
7    
8    
9    
10
11    
12    
13    Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
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            Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
54            
55            ICLR 2024
56            
57        
58
59        
60            
61                
62                    Guocheng
63                            Qian1,2 
64                
65                
66                    Jinjie
67                            Mai1
68                
69                
70                    Abdullah
71                            Hamdi3
72                
73                
74                    Jian
75                            Ren2
76                
77            
78            
79                
80                    Aliaksandr
81                            Siarohin2
82                
83                
84                    Bing
85                            Li1
86                
87                
88                    Hsin-Ying
89                            Lee2
90                
91                
92                    Ivan
93                            Skorokhodov1,2
94                
95            
96            
97                
98                    Peter
99                            Wonka1
100                
101                
102                    Sergey
103                            Tulyakov2
104                
105                
106                    Bernard
107                            Ghanem1
108                
109            
110            
111                
112                    1King Abdullah University of Science and
113                            Technology (KAUST)
114                
115                
116                    2Snap Inc.
117                
118                
119                    3Visual Geometry Group, University of
120                            Oxford
121                
122            
123        
124    
125
126    
127        
128            
129                
130                    
131                
132                Arxiv
133            
134            
135                
136                    
137                
138                Code
139            
140        
141    
142
143    
144    
145    
146
147
148    
149    
150        
151            
152                Abstract
153                
154                    
155                    We present "Magic123", a two-stage coarse-to-fine solution for high-quality, textured 3D meshes
156                    generation from a single unposed image in the wild using both 2D and 3D priors. In the first stage,
157                    we optimize a neural radiance field to produce a coarse geometry. In the second stage, we adopt a
158                    memory-efficient differentiable mesh representation to yield a high-resolution mesh with a visually
159                    appealing texture. In both stages, the 3D content is learned through reference view supervision and
160                    novel views guided by both 2D and 3D diffusion priors. We introduce a single tradeoff parameter
161                    between the 2D and 3D priors to control exploration (more imaginative) and exploitation (more
162                    precise) of the generated geometry. Additionally, We employ textual inversion and monocular depth
163                    regularization to encourage consistent appearances across views and to prevent degenerate solutions,
164                    respectively.
165                    Magic123 demonstrates a significant improvement over previous image-to-3D techniques, as validated
166                    through extensive experiments on synthetic benchmarks and diverse real-world images.
167                    
168                '><link rel=alternate hreflang=en-us href=https://guochengqian.github.io/project/magic123/><meta name=theme-color content="#b76531"><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/academicons/1.8.6/css/academicons.min.css integrity="sha256-uFVgMKfistnJAfoCUQigIl+JfUaP47GrRKjf6CTPVmw=" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.5.1/css/all.min.css integrity="sha512-DTOQO9RWCH3ppGqcWaEA1BIZOC6xxalwEsw9c2QQeAIftl+Vegovlnee1c9QX4TctnWMn13TZye+giMm8e2LwA==" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.2.5/jquery.fancybox.min.css integrity="sha256-ygkqlh3CYSUri3LhQxzdcm0n1EQvH2Y+U5S2idbLtxs=" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/9.15.6/styles/github.min.css crossorigin=anonymous title=hl-light><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/9.15.6/styles/dracula.min.css crossorigin=anonymous title=hl-dark disabled><link rel=stylesheet href="https://fonts.googleapis.com/css?family=Newsreader:500,600,700%7CManrope:400,500,600,700,800%7CIBM+Plex+Mono:400,500,600&display=swap"><link rel=stylesheet href=/css/academic.min.c06754a7cf79d154c937a526cfcdc457.css><link rel=manifest href=/manifest.webmanifest><link rel=icon href="/img/icon.ico?v=202607072"><link rel=apple-touch-icon href="/img/icon.ico?v=202607072"><link rel=canonical href=https://guochengqian.github.io/project/magic123/><meta property="twitter:card" content="summary_large_image"><meta property="og:site_name" content="Gordon Qian"><meta property="og:url" content="https://guochengqian.github.io/project/magic123/"><meta property="og:title" content=" | Gordon Qian"><meta property="og:description" content='<!DOCTYPE html>
169
170
171
172    
173
174    
175    
176    
177
178    
179    
180    Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
181    
182    
183    
184    
185
186    
187    
188    
189
190
191    
192
193    
194    
195    
196    
197    
198    
199    
200
201    
202    
203    
204    
205    
206    
207    
208
209    
210    
211
212
213
214
215    
216
217    
218        
219            
220            Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
221            
222            ICLR 2024
223            
224        
225
226        
227            
228                
229                    Guocheng
230                            Qian1,2 
231                
232                
233                    Jinjie
234                            Mai1
235                
236                
237                    Abdullah
238                            Hamdi3
239                
240                
241                    Jian
242                            Ren2
243                
244            
245            
246                
247                    Aliaksandr
248                            Siarohin2
249                
250                
251                    Bing
252                            Li1
253                
254                
255                    Hsin-Ying
256                            Lee2
257                
258                
259                    Ivan
260                            Skorokhodov1,2
261                
262            
263            
264                
265                    Peter
266                            Wonka1
267                
268                
269                    Sergey
270                            Tulyakov2
271                
272                
273                    Bernard
274                            Ghanem1
275                
276            
277            
278                
279                    1King Abdullah University of Science and
280                            Technology (KAUST)
281                
282                
283                    2Snap Inc.
284                
285                
286                    3Visual Geometry Group, University of
287                            Oxford
288                
289            
290        
291    
292
293    
294        
295            
296                
297                    
298                
299                Arxiv
300            
301            
302                
303                    
304                
305                Code
306            
307        
308    
309
310    
311    
312    
313
314
315    
316    
317        
318            
319                Abstract
320                
321                    
322                    We present "Magic123", a two-stage coarse-to-fine solution for high-quality, textured 3D meshes
323                    generation from a single unposed image in the wild using both 2D and 3D priors. In the first stage,
324                    we optimize a neural radiance field to produce a coarse geometry. In the second stage, we adopt a
325                    memory-efficient differentiable mesh representation to yield a high-resolution mesh with a visually
326                    appealing texture. In both stages, the 3D content is learned through reference view supervision and
327                    novel views guided by both 2D and 3D diffusion priors. We introduce a single tradeoff parameter
328                    between the 2D and 3D priors to control exploration (more imaginative) and exploitation (more
329                    precise) of the generated geometry. Additionally, We employ textual inversion and monocular depth
330                    regularization to encourage consistent appearances across views and to prevent degenerate solutions,
331                    respectively.
332                    Magic123 demonstrates a significant improvement over previous image-to-3D techniques, as validated
333                    through extensive experiments on synthetic benchmarks and diverse real-world images.
334                    
335                '><meta property="og:image" content="https://guochengqian.github.io/img/avatar.jpg"><meta property="twitter:image" content="https://guochengqian.github.io/img/avatar.jpg"><meta property="og:locale" content="en-us"><meta property="article:modified_time" content="2026-06-02T14:28:55-07:00">
335<script type=application/ld+json>{"@context":"https://schema.org","@type":"Article","mainEntityOfPage":{"@type":"WebPage","@id":"https://guochengqian.github.io/project/magic123/"},"headline":"","datePublished":"0001-01-01T00:00:00Z","dateModified":"2026-06-02T14:28:55-07:00","author":{"@type":"Person","name":"Gordon Qian"},"publisher":{"@type":"Organization","name":"Gordon Qian","logo":{"@type":"ImageObject","url":"https://guochengqian.github.io/img/icon.ico"}},"description":"\u003c!DOCTYPE html\u003e\n\u003chtml\u003e\n\n\u003chead\u003e\n    \u003cscript id=\"MathJax-script\" async src=\"https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js\"\u003e\u003c/script\u003e\n\n    \u003c!-- Google tag (gtag.js) --\u003e\n    \u003cscript async src=\"https://www.googletagmanager.com/gtag/js?id=G-XB3PR2Y1TQ\"\u003e\u003c/script\u003e\n    \u003cscript\u003e\n        window.dataLayer = window.dataLayer || [];\n\n        function gtag() {\n            dataLayer.push(arguments);\n        }\n        gtag('js', new Date());\n\n        gtag('config', 'G-XB3PR2Y1TQ');\n    \u003c/script\u003e\n\n    \u003cmeta charset=\"utf-8\"\u003e\n    \u003cmeta name=\"viewport\" content=\"width=device-width, initial-scale=1.0, shrink-to-fit=no\"\u003e\n    \u003ctitle\u003eMagic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors\u003c/title\u003e\n    \u003clink rel=\"stylesheet\" href=\"https://cdnjs.cloudflare.com/ajax/libs/twitter-bootstrap/4.5.0/css/bootstrap.min.css\"\u003e\n    \u003clink href='https://fonts.googleapis.com/css?family=Source+Sans+Pro:300,400,500,600' rel='stylesheet'\n        type='text/css'\u003e\n    \u003clink rel=\"stylesheet\" href=\"assets/css/Highlight-Clean.css\"\u003e\n    \u003clink rel=\"stylesheet\" href=\"assets/css/styles.css\"\u003e\n\n    \u003clink rel=\"apple-touch-icon\" sizes=\"180x180\" href=\"/dragon.png\"\u003e\n    \u003clink rel=\"icon\" type=\"image/png\" sizes=\"32x32\" href=\"/dragon-32x32.png\"\u003e\n    \u003clink rel=\"icon\" type=\"image/png\" sizes=\"16x16\" href=\"/dragon-16x16.png\"\u003e\n\n\n    \u003clink rel=\"manifest\" href=\"/site.webmanifest\"\u003e\n\n    \u003cmeta property=\"og:site_name\"\n        content=\"Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors\" /\u003e\n    \u003cmeta property=\"og:type\" content=\"video.other\" /\u003e\n    \u003cmeta property=\"og:title\"\n        content=\"Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors\" /\u003e\n    \u003cmeta property=\"og:description\"\n        content=\"Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors, 2023.\" /\u003e\n    \u003cmeta property=\"og:url\" content=\"https://guochengqian.github.io/project/magic123/\" /\u003e\n    \u003cmeta property=\"og:image\" content=\"https://guochengqian.github.io/project/magic123/assets/images/pull.jpg\" /\u003e\n    \u003c!-- todo change the sample image here  --\u003e\n\n    \u003cmeta property=\"article:publisher\" content=\"https://guochengqian.github.io/project/magic123/\" /\u003e\n    \u003cmeta name=\"twitter:card\" content=\"summary_large_image\" /\u003e\n    \u003cmeta name=\"twitter:title\"\n        content=\"Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors\" /\u003e\n    \u003cmeta name=\"twitter:description\"\n        content=\"We combine 2D and 3D priors to generate photo-realistic 3D objects from a single image.\" /\u003e\n    \u003cmeta name=\"twitter:url\" content=\"https://guochengqian.github.io/project/magic123/\" /\u003e\n    \u003cmeta name=\"twitter:image\" content=\"https://guochengqian.github.io/project/magic123/assets/images/pull.jpg\" /\u003e\n    \u003c!-- \u003cmeta name=\"twitter:site\" content=\"\" /\u003e --\u003e\n\n    \u003cscript src=\"assets/js/video_comparison.js\"\u003e\u003c/script\u003e\n    \u003cscript type=\"module\" src=\"https://unpkg.com/@google/[email protected]/dist/model-viewer.min.js\"\u003e\u003c/script\u003e\n\u003c/head\u003e\n\n\u003cbody\u003e\n\n    \u003cstyle\u003e\n        .project-links {\n            margin-top: 0.25rem;\n            margin-bottom: 2rem;\n        }\n        .project-link-item {\n            margin-bottom: 0.25rem;\n        }\n        .project-link-icon {\n            height: 64px;\n            width: auto;\n            display: inline-block;\n        }\n        .hero-gif {\n            width: 100%;\n            height: auto;\n            display: block;\n            margin: 14px auto 0;\n        }\n        .example-generated-objects,\n        .example-generated-objects h2,\n        .example-generated-objects p {\n            text-align: center;\n        }\n        .example-generated-objects .examples-grid .row {\n            justify-content: center;\n        }\n        .example-generated-objects .examples-grid .col-sm-1 {\n            justify-content: center !important;\n            text-align: center;\n        }\n        .example-generated-objects .examples-grid img,\n        .example-generated-objects .examples-grid video,\n        .tradeoff-media img,\n        .tradeoff-media video,\n        .quantitative-results-media img {\n            width: 100%;\n            max-width: 100%;\n            height: auto;\n            display: block;\n            margin-left: auto;\n            margin-right: auto;\n        }\n        .tradeoff-media {\n            width: 100%;\n        }\n        .tradeoff-media img,\n        .tradeoff-media video {\n            margin-bottom: 16px;\n        }\n        .tradeoff-media \u003e img,\n        .tradeoff-media \u003e video {\n            width: 100% !important;\n            max-width: none !important;\n            min-width: 100%;\n            height: auto !important;\n            max-height: none !important;\n            display: block;\n            box-sizing: border-box;\n            object-fit: contain;\n        }\n        .footer .footer-icons {\n            display: flex;\n            justify-content: center;\n            align-items: center;\n            gap: 16px;\n            text-align: center;\n            margin-bottom: 10px;\n        }\n        .footer .icon-link {\n            display: inline-flex;\n            align-items: center;\n            justify-content: center;\n            min-width: 36px;\n            min-height: 36px;\n            font-size: 18px;\n            text-decoration: none;\n        }\n        .footer .content {\n            text-align: center;\n        }\n        @media (max-width: 575px) {\n            .project-link-icon {\n                height: 52px;\n            }\n            .project-links {\n                margin-bottom: 1.5rem;\n            }\n        }\n    \u003c/style\u003e\n\n    \u003cdiv class=\"highlight-clean\" style=\"padding-bottom: 10px;\"\u003e\n        \u003cdiv class=\"container\" style=\"max-width: 100%;\"\u003e\n            \u003cbr\u003e\n            \u003ch1 class=\"text-center\"\u003e\u003cb\u003eMagic123\u003c/b\u003e: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors\u003c/h1\u003e\n            \u003cbr\u003e\n            \u003ch2 class=\"text-center\"\u003e\u003cb\u003eICLR 2024\u003c/b\u003e\u003c/h2\u003e\n            \u003cbr\u003e\n        \u003c/div\u003e\n\n        \u003cdiv class=\"container author-container\"\u003e\n            \u003cdiv class=\"row authors\"\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://guochengqian.github.io/\"\u003eGuocheng\n                            Qian\u003csup\u003e1,2\u003c/sup\u003e \u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://cemse.kaust.edu.sa/people/person/jinjie-mai\"\u003eJinjie\n                            Mai\u003csup\u003e1\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://abdullahamdi.com/\"\u003eAbdullah\n                            Hamdi\u003csup\u003e3\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://alanspike.github.io/\"\u003eJian\n                            Ren\u003csup\u003e2\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n            \u003c/div\u003e\n            \u003cdiv class=\"row authors\"\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca class=\"text-center\"\n                            href=\"https://aliaksandrsiarohin.github.io/aliaksandr-siarohin-website/\"\u003eAliaksandr\n                            Siarohin\u003csup\u003e2\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca class=\"text-center\" href=\"https://cemse.kaust.edu.sa/people/person/bing-li\"\u003eBing\n                            Li\u003csup\u003e1\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"http://hsinyinglee.com/\"\u003eHsin-Ying\n                            Lee\u003csup\u003e2\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://universome.github.io/\"\u003eIvan\n                            Skorokhodov\u003csup\u003e1,2\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n            \u003c/div\u003e\n            \u003cdiv class=\"row authors\"\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://peterwonka.net/\"\u003ePeter\n                            Wonka\u003csup\u003e1\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"http://www.stulyakov.com/\"\u003eSergey\n                            Tulyakov\u003csup\u003e2\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch5\u003e\u003ca href=\"https://www.bernardghanem.com/\"\u003eBernard\n                            Ghanem\u003csup\u003e1\u003c/sup\u003e\u003c/a\u003e\u003c/h5\u003e\n                \u003c/div\u003e\n            \u003c/div\u003e\n            \u003cdiv class=\"row affiliations\"\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch6 class=\"text-center\"\u003e\u003ca class=\"text-center\"\u003e\u003csup\u003e1\u003c/sup\u003eKing Abdullah University of Science and\n                            Technology (KAUST)\u003c/a\u003e\u003c/h6\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch6 class=\"text-center\"\u003e\u003ca class=\"text-center\"\u003e\u003csup\u003e2\u003c/sup\u003eSnap Inc.\u003c/a\u003e\u003c/h6\u003e\n                \u003c/div\u003e\n                \u003cdiv class=\"col\"\u003e\n                    \u003ch6 class=\"text-center\"\u003e\u003ca class=\"text-center\"\u003e\u003csup\u003e3\u003c/sup\u003eVisual Geometry Group, University of\n                            Oxford\u003c/a\u003e\u003c/h6\u003e\n                \u003c/div\u003e\n            \u003c/div\u003e\n        \u003c/div\u003e\n    \u003c/div\u003e\n\n    \u003cdiv class=\"container align-items-center project-links\"\u003e\n        \u003cdiv class=\"row text-center mx-auto justify-content-center\"\u003e\n            \u003cdiv class=\"col-sm-4 col-md-3 project-link-item\"\u003e\n                \u003ca href=\"https://arxiv.org/abs/2306.17843\" target=\"_blank\" class=\"link-block w-inline-block\"\u003e\n                    \u003cimg src=\"assets/images/arxiv_paper.png\" alt=\"paper\" class=\"img-fluid project-link-icon\" /\u003e\u003c/a\u003e\n                \u003cbr\u003e\n                \u003cdiv class=\"text-block-2\"\u003e\u003cstrong class=\"bold-text-nerf_v2\"\u003eArxiv\u003c/strong\u003e\u003c/div\u003e\n            \u003c/div\u003e\n            \u003cdiv class=\"col-sm-4 col-md-3 project-link-item\"\u003e\n                \u003ca href=\"https://github.com/guochengqian/Magic123\" target=\"_blank\" class=\"link-block w-inline-block\"\u003e\n                    \u003cimg src=\"assets/images/github.png\" alt=\"github\" class=\"img-fluid project-link-icon\" /\u003e\u003c/a\u003e\n                \u003cbr\u003e\n                \u003cdiv class=\"text-block-2\"\u003e\u003cstrong class=\"bold-text-nerf_v2\"\u003eCode\u003c/strong\u003e\u003c/div\u003e\n            \u003c/div\u003e\n        \u003c/div\u003e\n    \u003c/div\u003e\n\n    \u003c/div\u003e\n    \u003c!-- \u003cdiv class=\"banner\"\u003e\n      \u003cvideo class=\"video lazy\"\n          poster=\"https://dreamfusion-cdn.ajayj.com/sept28/banner_1x6_customhue_A.jpg\"\n          autoplay loop playsinline muted\u003e\n        \u003csource data-src=\"https://dreamfusion-cdn.ajayj.com/sept28/banner_1x6_customhue_A.mp4\" type=\"video/mp4\" style=\"width: 100%; height: auto;\"\u003e\u003c/source\u003e\n      \u003c/video\u003e\n    \u003c/div\u003e --\u003e\n    \u003cvideo src=\"assets/videos/pull.mp4\" autoplay loop playsinline muted class=\"hero-gif\" type=\"video/mp4\"\u003e\u003c/video\u003e\n\n\n    \u003chr class=\"divider\" /\u003e\n    \u003cdiv class=\"container\" style=\"max-width: col-md-8;\"\u003e\n        \u003cdiv class=\"row\"\u003e\n            \u003cdiv class=\"col-sm-12\"\u003e\n                \u003ch2\u003eAbstract\u003c/h2\u003e\n                \u003cp\u003e\n                    \u003c!-- \u003cstrong\u003e --\u003e\n                    We present \"Magic123\", a two-stage coarse-to-fine solution for high-quality, textured 3D meshes\n                    generation from a single unposed image in the wild using both 2D and 3D priors. In the first stage,\n                    we optimize a neural radiance field to produce a coarse geometry. In the second stage, we adopt a\n                    memory-efficient differentiable mesh representation to yield a high-resolution mesh with a visually\n                    appealing texture. In both stages, the 3D content is learned through reference view supervision and\n                    novel views guided by both 2D and 3D diffusion priors. We introduce a single tradeoff parameter\n                    between the 2D and 3D priors to control exploration (more imaginative) and exploitation (more\n                    precise) of the generated geometry. Additionally, We employ textual inversion and monocular depth\n                    regularization to encourage consistent appearances across views and to prevent degenerate solutions,\n                    respectively.\n                    Magic123 demonstrates a significant improvement over previous image-to-3D techniques, as validated\n                    through extensive experiments on synthetic benchmarks and diverse real-world images.\n                    \u003c!-- \u003c/strong\u003e --\u003e\n                \u003c/p\u003e"}</script>
335<title>| Gordon Qian</title></head><body id=top data-spy=scroll data-offset=70 data-target=#TableOfContents class=theme-claude><aside class=search-results id=search><div class=container><section class=search-header><div class="row no-gutters justify-content-between mb-3"><div class=col-6><h1>Search</h1></div><div class="col-6 col-search-close"><a class=js-search href=#><i class="fas fa-times-circle text-muted" aria-hidden=true></i></a></div></div><div id=search-box><input name=q id=search-query placeholder=Search... autocapitalize=off autocomplete=off autocorrect=off spellcheck=false type=search></div></section><section class=section-search-results><div id=search-hits></div></section></div></aside><nav class="navbar navbar-light fixed-top navbar-expand-lg py-0 compensate-for-scrollbar" id=navbar-main><div class=container><a class=navbar-brand href=/>Gordon Qian</a>
336<button type=button class=navbar-toggler data-toggle=collapse data-target=#navbar aria-controls=navbar aria-expanded=false aria-label="Toggle navigation">
337<span><i class="fas fa-bars"></i></span></button><div class="collapse navbar-collapse" id=navbar><ul class="navbar-nav mr-auto"><li class=nav-item><a class=nav-link href=/#about><span>Home</span></a></li><li class=nav-item><a class=nav-link href=/#news><span>News</span></a></li><li class=nav-item><a class=nav-link href=/#focused-areas><span>Focus Areas</span></a></li><li class=nav-item><a class=nav-link href=/#featured><span>Publications</span></a></li><li class=nav-item><a class=nav-link href=/#experience><span>Experience</span></a></li></ul><ul class="navbar-nav ml-auto"><li class=nav-item><a class=nav-link href=/blog/><span>Blog</span></a></li><li class=nav-item><a class="nav-link js-search" href=#><i class="fas fa-search" aria-hidden=true></i></a></li><li class=nav-item><a class="nav-link js-dark-toggle" href=#><i class="fas fa-moon" aria-hidden=true></i></a></li></ul></div></div></nav><article class="article article-project"><div class="article-container pt-3"><h1></h1><div class=article-metadata><span class=article-date>Last updated on
338Jun 2, 2026</span><div class=share-box aria-hidden=true><ul class=share><li><a href="https://twitter.com/intent/tweet?url=https://guochengqian.github.io/project/magic123/&amp;text=" target=_blank rel=noopener class=share-btn-twitter><i class="fab fa-twitter"></i></a></li><li><a href="https://www.facebook.com/sharer.php?u=https://guochengqian.github.io/project/magic123/&amp;t=" target=_blank rel=noopener class=share-btn-facebook><i class="fab fa-facebook-f"></i></a></li><li><a href="mailto:?subject=&amp;body=https://guochengqian.github.io/project/magic123/" target=_blank rel=noopener class=share-btn-email><i class="fas fa-envelope"></i></a></li><li><a href="https://www.linkedin.com/shareArticle?url=https://guochengqian.github.io/project/magic123/&amp;
338title=" target=_blank rel=noopener class=share-btn-linkedin><i class="fab fa-linkedin-in"></i></a></li><li><a href="https://web.whatsapp.com/send?text=%20https://guochengqian.github.io/project/magic123/" target=_blank rel=noopener class=share-btn-whatsapp><i class="fab fa-whatsapp"></i></a></li><li><a href="https://service.weibo.com/share/share.php?url=https://guochengqian.github.io/project/magic123/&amp;title=" target=_blank rel=noopener class=share-btn-weibo><i class="fab fa-weibo"></i></a></li></ul></div></div></div><div class=article-container><div class=article-style><!doctype html><html><head>
338<script id=MathJax-script async src=https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js></script>
vendor: 63 bytes, line 338
338<script async src="https://www.googletagmanager.com/gtag/js?id=
338G-XB3PR2Y1TQ
vendor: 11 bytes, line 338
338"></script>
338<script>window.dataLayer=window.dataLayer||[];function gtag(){dataLayer.push(arguments)}gtag("js",new Date),gtag("config","G-XB3PR2Y1TQ")</script>
338<meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1,shrink-to-fit=no"><title>Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors</title><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/twitter-bootstrap/4.5.0/css/bootstrap.min.css><link href='https://fonts.googleapis.com/css?family=Source+Sans+Pro:300,400,500,600' rel=stylesheet type=text/css><link rel=stylesheet href=assets/css/Highlight-Clean.css><link rel=stylesheet href=assets/css/styles.css><link rel=apple-touch-icon sizes=180x180 href=/dragon.png><link rel=icon type=image/png sizes=32x32 href=/dragon-32x32.png><link rel=icon type=image/png sizes=16x16 href=/dragon-16x16.png><link rel=manifest href=/site.webmanifest><meta property="og:site_name" content="Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors"><meta property="og:type" content="video.other"><meta property="og:title" content="Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors"><meta property="og:description" content="Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors, 2023."><meta property="og:url" content="https://guochengqian.github.io/project/magic123/"><meta property="og:image" content="https://guochengqian.github.io/project/magic123/assets/images/pull.jpg"><meta property="article:publisher" content="https://guochengqian.github.io/project/magic123/"><meta name=twitter:card content="summary_large_image"><meta name=twitter:title content="Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors"><meta name=twitter:description content="We combine 2D and 3D priors to generate photo-realistic 3D objects from a single image."><meta name=twitter:url content="https://guochengqian.github.io/project/magic123/"><meta name=twitter:image content="https://guochengqian.github.io/project/magic123/assets/images/pull.jpg">
338<script src=assets/js/video_comparison.js></script>
338<script type=module src=https://unpkg.com/@google/[email protected]/dist/model-viewer.min.js></script>
338</head><body><style>.project-links{margin-top:.25rem;margin-bottom:2rem}.project-link-item{margin-bottom:.25rem}.project-link-icon{height:64px;width:auto;display:inline-block}.hero-gif{width:100%;height:auto;display:block;margin:14px auto 0}.example-generated-objects,.example-generated-objects h2,.example-generated-objects p{text-align:center}.example-generated-objects .examples-grid .row{justify-content:center}.example-generated-objects .examples-grid .col-sm-1{justify-content:center!important;text-align:center}.example-generated-objects .examples-grid img,.example-generated-objects .examples-grid video,.tradeoff-media img,.tradeoff-media video,.quantitative-results-media img{width:100%;max-width:100%;height:auto;display:block;margin-left:auto;margin-right:auto}.tradeoff-media{width:100%}.tradeoff-media img,.tradeoff-media video{margin-bottom:16px}.tradeoff-media>img,.tradeoff-media>video{width:100%!important;max-width:none!important;min-width:100%;height:auto!important;max-height:none!important;display:block;box-sizing:border-box;object-fit:contain}.footer .footer-icons{display:flex;justify-content:center;align-items:center;gap:16px;text-align:center;margin-bottom:10px}.footer .icon-link{display:inline-flex;align-items:center;justify-content:center;min-width:36px;min-height:36px;font-size:18px;text-decoration:none}.footer .content{text-align:center}@media(max-width:575px){.project-link-icon{height:52px}.project-links{margin-bottom:1.5rem}}</style><div class=highlight-clean style=padding-bottom:10px><div class=container style=max-width:100%><br><h1 class=text-center><b>Magic123</b>: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors</h1><br><h2 class=text-center><b>ICLR 2024</b></h2><br></div><div class="container author-container"><div class="row authors"><div class=col><h5><a href=https://guochengqian.github.io/>Guocheng
339Qian<sup>1,2</sup></a></h5></div><div class=col><h5><a href=https://cemse.kaust.edu.sa/people/person/jinjie-mai>Jinjie
340Mai<sup>1</sup></a></h5></div><div class=col><h5><a href=https://abdullahamdi.com/>Abdullah
341Hamdi<sup>3</sup></a></h5></div><div class=col><h5><a href=https://alanspike.github.io/>Jian
342Ren<sup>2</sup></a></h5></div></div><div class="row authors"><div class=col><h5><a class=text-center href=https://aliaksandrsiarohin.github.io/aliaksandr-siarohin-website/>Aliaksandr
343Siarohin<sup>2</sup></a></h5></div><div class=col><h5><a class=text-center href=https://cemse.kaust.edu.sa/people/person/bing-li>Bing
344Li<sup>1</sup></a></h5></div><div class=col><h5><a href=http://hsinyinglee.com/>Hsin-Ying
345Lee<sup>2</sup></a></h5></div><div class=col><h5><a href=https://universome.github.io/>Ivan
346Skorokhodov<sup>1,2</sup></a></h5></div></div><div class="row authors"><div class=col><h5><a href=https://peterwonka.net/>Peter
347Wonka<sup>1</sup></a></h5></div><div class=col><h5><a href=http://www.stulyakov.com/>Sergey
348Tulyakov<sup>2</sup></a></h5></div><div class=col><h5><a href=https://www.bernardghanem.com/>Bernard
349Ghanem<sup>1</sup></a></h5></div></div><div class="row affiliations"><div class=col><h6 class=text-center><a class=text-center><sup>1</sup>King Abdullah University of Science and
350Technology (KAUST)</a></h6></div><div class=col><h6 class=text-center><a class=text-center><sup>2</sup>Snap Inc.</a></h6></div><div class=col><h6 class=text-center><a class=text-center><sup>3</sup>Visual Geometry Group, University of
351Oxford</a></h6></div></div></div></div><div class="container align-items-center project-links"><div class="row text-center mx-auto justify-content-center"><div class="col-sm-4 col-md-3 project-link-item"><a href=https://arxiv.org/abs/2306.17843 target=_blank class="link-block w-inline-block"><img src=assets/images/arxiv_paper.png alt=paper class="img-fluid project-link-icon"></a><br><div class=text-block-2><strong class=bold-text-nerf_v2>Arxiv</strong></div></div><div class="col-sm-4 col-md-3 project-link-item"><a href=https://github.com/guochengqian/Magic123 target=_blank class="link-block w-inline-block"><img src=assets/images/github.png alt=github class="img-fluid project-link-icon"></a><br><div class=text-block-2><strong class=bold-text-nerf_v2>Code</strong></div></div></div></div></div><video src=assets/videos/pull.mp4 autoplay loop playsinline muted class=hero-gif type=video/mp4></video><hr class=divider><div class=container style=max-width:col-md-8><div class=row><div class=col-sm-12><h2>Abstract</h2><p>We present "Magic123", a two-stage coarse-to-fine solution for high-quality, textured 3D meshes
352generation from a single unposed image in the wild using both 2D and 3D priors. In the first stage,
353we optimize a neural radiance field to produce a coarse geometry. In the second stage, we adopt a
354memory-efficient differentiable mesh representation to yield a high-resolution mesh with a visually
355appealing texture. In both stages, the 3D content is learned through reference view supervision and
356novel views guided by both 2D and 3D diffusion priors. We introduce a single tradeoff parameter
357between the 2D and 3D priors to control exploration (more imaginative) and exploitation (more
358precise) of the generated geometry. Additionally, We employ textual inversion and monocular depth
359regularization to encourage consistent appearances across views and to prevent degenerate solutions,
360respectively.
361Magic123 demonstrates a significant improvement over previous image-to-3D techniques, as validated
362through extensive experiments on synthetic benchmarks and diverse real-world images.</p></div></div></div><div class=container style=max-width:col-md-8><div class="row captioned_videos"><div class=col-sm-12><img src=assets/images/magic123_pipeline.png alt="Image description" class=img-fluid><h6 class=caption style=text-align:left><b>Magic123 pipeline.</b> Magic123 is a two-stage
363coarse-to-fine framework for high-quality
3643D generation from a reference image. Magic123 is guided by the reference image, constrained by the
365monocular depth estimation from the image, and driven by a joint 2D and 3D diffusion prior to dream
366up novel views. At the coarse stage, we optimize an Instant-NGP neural radiance field (NeRF) to
367reconstruct a coarse geometry. At the fine stage, we initialize a DMTet mesh from the NeRF output
368and optimize a high-resolution mesh and texture directly. Textural inversion is used in both stages
369to generate object-preserving geometry and view-consistent textures.</h6></div></div></div><hr class=divider><div class="container example-generated-objects" style=max-width:col-md-8><div class="row text-center"><div class=col-sm-12><h2>Example generated objects</h2><p>Magic123 generates photo-realistic 3D objects from a single unposed image.</p></div></div><div class="container align-items-center examples-grid"><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title>Input</div></div><div class=col-sm-2><img src=assets/videos/reference/bear.png alt="Bear Image" class=img-fluid></div><div class=col-sm-2><img src=assets/videos/reference/dragon.png alt="Dragon Image" class=img-fluid></div><div class=col-sm-2><img src=assets/videos/reference/horse.png alt="Horse Image" class=img-fluid></div><div class=col-sm-2><img src=assets/videos/reference/teapot.png alt="Sparrow Image" class=img-fluid></div></div><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title><a href=https://github.com/openai/shap-e>Shap-E</a></div></div><div class=col-sm-2><video src=assets/videos/shape_e/bear.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/shape_e/dragon.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/shape_e/horse.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/shape_e/teapot.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title><a href=https://ku-cvlab.github.io/3DFuse/>3D Fuse</a></div></div><div class=col-sm-2><video src=assets/videos/3d_fuse/bear.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/3d_fuse/dragon.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/3d_fuse/horse.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/3d_fuse/teapot.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title>
369<a href=https://vita-group.github.io/NeuralLift-360/>Neural Lift</a></div></div><div class=col-sm-2><video src=assets/videos/neural_lift/bear.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/neural_lift/dragon.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/neural_lift/horse.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/neural_lift/teapot.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title><a href=https://lukemelas.github.io/realfusion/>Real Fusion</a></div></div><div class=col-sm-2><video src=assets/videos/realfusion/bear.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/realfusion/dragon.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/realfusion/horse.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/realfusion/teapot.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div><div class=row><div class="col-sm-1 d-flex align-items-center justify-content-start p-0"><div class=method-title><b>Magic123</b><br><b>(Ours)</b></div></div><div class=col-sm-2><video src=assets/videos/magic123/bear.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/magic123/dragon.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/magic123/horse.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div><div class=col-sm-2><video src=assets/videos/magic123/teapot.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div></div></div><hr class=divider><div class=container style=max-width:col-md-8><div class=row><div class=col-sm-12><h2>Tradeoff between 2D and 3D priors in Magic123</h2><p>We compare single image reconstructions for three cases: a cactus (common object), two stacked
370donuts (less common object), and a dragon statue (uncommon object).
371Magic123 with only 2D prior (on the right) favors geometry exploration, generating 3D content
372with
373more imagination but potentially lacking 3D consistency. Magic123 with only 3D prior prioritizes
374geometry exploitation, resulting in precise yet potentially simplified geometry with reduced
375details. Magic123 thus proposes to use both 2D and 3D prior and introduces a tradeoff parameter
376\( \lambda_{2D/3D} \) to control the geometry exploration and exploitation. We provide a
377balanced point
378\( \lambda_{2D/3D}=1 \) , with which Magic123 consistently offers
379identity-preserving 3D content with fine-grained geometry and visually appealing texture.</p></div></div><div class="row justify-content-center" style=padding:0;margin:0><div class="col-sm-8 text-center tradeoff-media"><img src=assets/images/priors.jpeg alt="Image description" class=img-fluid><div style=height:30px></div><img src=assets/images/2d_3d.png alt="Image description" class=img-fluid>
380<video src=assets/videos/human_examples.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video>
381<video src=assets/videos/lambda_2d.mp4 autoplay loop playsinline muted type=video/mp4 style=width:100%;height:auto></video></div></div></div></div><hr class=divider><div class=container style=max-width:col-md-8><div class=row><div class=col-sm-12><h2>Effects of the coarse-to-fine pipeline</h2><p>Qualitative comparisons of the coarse and fine stages between Magic123 with only 2D prior \(
382\lambda_{2D}=1,\lambda_{3D}=0 \),
383Magic123 with only 3D prior \( \lambda_{2D}=0,\lambda_{3D}=40 \), and Magic123 \(
384\lambda_{2D}=1,\lambda_{3D}=40 \).</p></div></div><div class="row justify-content-center" style=padding:0;margin:0><div class="col-sm-12 text-center"><img src=assets/images/coarse_2_fine.png alt="Image description" class=img-fluid><div style=height:10px></div></div></div></div><hr class=divider><div class=container style=max-width:col-md-8><div class=row><div class=col-sm-12><h2>Quantitative Results</h2><p>To demonstrate the effectiveness of the proposed Magic123, we evaluate its performance on the
385NeRF4 and RealFusion15 datasets. We conduct a comprehensive and quantitative comparison with
386multiple baselines for both datasets, as shown in Table I. Notably, our method achieves Top-1
387performance across all the metrics when compared to previous SOTA approaches. This remarkable
388performance demonstrates the superiority of Magic123 and its ability to generate high-quality 3D
389representations.</p></div></div><div class="row justify-content-center" style=padding:0;margin:0><div class="col-sm-12 text-center quantitative-results-media"><img src=assets/images/results.jpeg alt="Image description" class=img-fluid></div></div></div><hr class=divider><footer class=footer><div class=container><div class="content has-text-centered footer-icons"><a class=icon-link href=https://arxiv.org/abs/2306.17843 aria-label="Magic123 paper"><i class="fa-solid fa-file-pdf" aria-hidden=true></i>
390</a><a class=icon-link href=https://github.com/guochengqian/Magic123 aria-label="Magic123 GitHub"><i class="fab fa-github" aria-hidden=true></i></a></div><div class="columns is-centered"><div class="column is-8"><div class=content><p>This website is licensed under a <a rel=license href=http://creativecommons.org/licenses/by-sa/4.0/>Creative
391Commons Attribution-ShareAlike 4.0 International License</a>.</p><p>We borrow the <a href=https://dreamfusion3d.github.io/>source code</a> for our
392website.
393We sincerely appreciate DreamFusion authors for their awesome templates.</p></div></div></div></div></footer>
393<script src=assets/js/yall.js></script>
393<script>yall({observeChanges:!0})</script>
393<script src=assets/js/scripts.js></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/twitter-bootstrap/4.5.0/js/bootstrap.bundle.min.js></script>
393</body></html></div></div></article>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery/3.4.1/jquery.min.js integrity="sha256-CSXorXvZcTkaix6Yvo6HppcZGetbYMGWSFlBw8HfCJo=" crossorigin=anonymous></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery.imagesloaded/4.1.4/imagesloaded.pkgd.min.js integrity="sha256-lqvxZrPLtfffUl2G/e7szqSvPBILGbwmsGE1MKlOi0Q=" crossorigin=anonymous></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery.isotope/3.0.6/isotope.pkgd.min.js integrity="sha256-CBrpuqrMhXwcLLUd5tvQ4euBHCdh7wGlDfNz8vbu/iI=" crossorigin=anonymous></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.2.5/jquery.fancybox.min.js integrity="sha256-X5PoE3KU5l+JcX+w09p/wHl9AzK333C4hJ2I9S5mD4M=" crossorigin=anonymous></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/9.15.6/highlight.min.js integrity="sha256-aYTdUrn6Ow1DDgh5JTc3aDGnnju48y/1c8s1dgkYPQ8=" crossorigin=anonymous></script>
393<script src=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/9.15.6/languages/r.min.js></script>
393<script>hljs.initHighlightingOnLoad()</script>
393<script>const search_config={indexURI:"/index.json",minLength:1,threshold:.3},i18n={no_results:"No results found",placeholder:"Search...",results:"results found"},content_type={post:"Posts",project:"Projects",publication:"Publications",talk:"Talks"}</script>
393<script id=search-hit-fuse-template type=text/x-template>
394      <div class="search-hit" id="summary-{{key}}">
395      <div class="search-hit-content">
396        <div class="search-hit-name">
397          <a href="{{relpermalink}}">{{title}}</a>
398          <div class="article-metadata search-hit-type">{{type}}</div>
399          <p class="search-hit-description">{{snippet}}</p>
400        </div>
401      </div>
402      </div>
403    </script>
403<script src=https://cdnjs.cloudflare.com/ajax/libs/fuse.js/3.2.1/fuse.min.js integrity="sha256-VzgmKYmhsGNNN4Ph1kMW+BjoYJM2jV5i4IlFoeZA9XI=" crossorigin=anonymous></script>
403<script src=https://cdnjs.cloudflare.com/ajax/libs/mark.js/8.11.1/jquery.mark.min.js integrity="sha256-4HLtjeVgH0eIB3aZ9mLYF6E8oU5chNdjU6p6rrXpl9U=" crossorigin=anonymous></script>
403<script src=/js/academic.min.b89d1011e4e05f2d9ac6fafcb0f556d1.js></script>
403<script>(function(){function n(e){var t=e.previousElementSibling;return t&&t.classList.contains("video-placeholder")?t:null}function t(e){e.style.display="block";var t=n(e);t&&(t.style.display="none")}function e(e){e.muted=!0,e.defaultMuted=!0,e.playsInline=!0,e.setAttribute("muted",""),e.setAttribute("playsinline",""),e.setAttribute("webkit-playsinline",""),t(e);var n=e.play();n&&typeof n.catch=="function"&&n.catch(function(){})}function s(n){if(n.dataset.mobileAutoplayReady==="true")return;n.dataset.mobileAutoplayReady="true",n.addEventListener("loadeddata",function(){e(n)}),n.addEventListener("canplay",function(){e(n)}),n.addEventListener("play",function(){t(n)}),n.readyState>=2&&e(n)}function o(){var e=document.querySelectorAll(".publication-image-container .video-content");e.forEach(s)}document.addEventListener("DOMContentLoaded",function(){o(),document.addEventListener("visibilitychange",function(){if(document.hidden)return;document.querySelectorAll(".publication-image-container .video-content").forEach(e)}),["touchstart","pointerdown"].forEach(function(t){document.addEventListener(t,function n(){document.querySelectorAll(".publication-image-container .video-content").forEach(e),document.removeEventListener(t,n)},{passive:!0})})})})()</script>
403<div class=container><footer class=site-footer><p class=powered-by>© 2026 Gordon Guocheng Qian. Thanks for the template from wowchemy-hugo-themes.
404<span class=float-right aria-hidden=true><a href=# class=back-to-top><span class=button_icon><i class="fas fa-chevron-up fa-2x"></i></span></a></span></p></footer></div><div id=modal class="modal fade" tabindex=-1 role=dialog aria-hidden=true><div class=modal-dialog><div class=modal-content><div class=modal-header><h5 class=modal-title>Cite</h5><button type=button class=close data-dismiss=modal aria-label=Close>
405<span aria-hidden=true>&#215;</span></button></div><div class=modal-body><pre><code class="tex hljs"></code></pre></div><div class=modal-footer><a class="btn btn-outline-primary my-1 js-copy-cite" href=# target=_blank><i class="fas fa-copy"></i> Copy
406</a><a class="btn btn-outline-primary my-1 js-download-cite" href=# target=_blank><i class="fas fa-download"></i> Download</a><div id=modal-error></div></div></div></div></div></body></html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.