Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD–CPU–GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., ∼100M Gaussians) and standard in-memory training (e.g., ∼11M Gaussians).</p>\n<p>Website & More Visualizations: <a href=\"https://sponge-lab.github.io/TideGS\" rel=\"nofollow\">https://sponge-lab.github.io/TideGS</a></p>\n<p>GitHub Repo: <a href=\"https://github.com/sponge-lab/TideGS\" rel=\"nofollow\">https://github.com/sponge-lab/TideGS</a></p>\n","updatedAt":"2026-05-20T09:29:46.854Z","author":{"_id":"64f1ec9e5894f48e7270dacd","avatarUrl":"/avatars/5aca8dfec4a31d49984dfec23541f067.svg","fullname":"Chaojian Li","name":"Chaojian","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8059031367301941},"editors":["Chaojian"],"editorAvatarUrls":["/avatars/5aca8dfec4a31d49984dfec23541f067.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.20150","authors":[{"_id":"6a0d589365eb30f20d962db9","name":"Chonghao Zhong","hidden":false},{"_id":"6a0d589365eb30f20d962dba","name":"Linfeng Shi","hidden":false},{"_id":"6a0d589365eb30f20d962dbb","name":"Hua Chen","hidden":false},{"_id":"6a0d589365eb30f20d962dbc","name":"Tiecheng Sun","hidden":false},{"_id":"6a0d589365eb30f20d962dbd","name":"Hao Zhao","hidden":false},{"_id":"6a0d589365eb30f20d962dbe","name":"Binhang Yuan","hidden":false},{"_id":"6a0d589365eb30f20d962dbf","name":"Chaojian Li","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/64f1ec9e5894f48e7270dacd/j_hk0GMrwgImtfuWf86U7.mp4"],"publishedAt":"2026-05-19T00:00:00.000Z","submittedOnDailyAt":"2026-05-20T00:00:00.000Z","title":"TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization","submittedOnDailyBy":{"_id":"64f1ec9e5894f48e7270dacd","avatarUrl":"/avatars/5aca8dfec4a31d49984dfec23541f067.svg","isPro":false,"fullname":"Chaojian Li","user":"Chaojian","type":"user","name":"Chaojian"},"summary":"Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD-CPU-GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., approximately 100M Gaussians) and standard in-memory training (e.g., approximately 11M Gaussians).","upvotes":2,"discussionId":"6a0d589365eb30f20d962dc0","projectPage":"https://sponge-lab.github.io/TideGS/","githubRepo":"https://github.com/sponge-lab/TideGS","githubRepoAddedBy":"user","ai_summary":"TideGS enables training 3D Gaussian Splatting with over one billion primitives on a single GPU by managing parameters across SSD-CPU-GPU hierarchy through block-virtualization, asynchronous pipeline, and differential streaming techniques.","ai_keywords":["3D Gaussian Splatting","out-of-core training","SSD-CPU-GPU hierarchy","block-virtualized geometry","hierarchical asynchronous pipeline","trajectory-adaptive differential streaming","working-set cache","parameter management"],"githubStars":12,"organization":{"_id":"6a0d590d26d3ca51300cd87d","name":"sponge-lab","fullname":"Sponge Computing Lab at HKUST","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/64f1ec9e5894f48e7270dacd/u-IyQjixVrkLo7bFS28X_.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64f1ec9e5894f48e7270dacd","avatarUrl":"/avatars/5aca8dfec4a31d49984dfec23541f067.svg","isPro":false,"fullname":"Chaojian Li","user":"Chaojian","type":"user"},{"_id":"689b5067e4aff5e9a74905d3","avatarUrl":"/avatars/3c82c90137c75198bf05c7023822042a.svg","isPro":false,"fullname":"Chonghao Zhong","user":"zchnanguan7","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a0d590d26d3ca51300cd87d","name":"sponge-lab","fullname":"Sponge Computing Lab at HKUST","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/64f1ec9e5894f48e7270dacd/u-IyQjixVrkLo7bFS28X_.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.20150.md"}">
TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization
Abstract
TideGS enables training 3D Gaussian Splatting with over one billion primitives on a single GPU by managing parameters across SSD-CPU-GPU hierarchy through block-virtualization, asynchronous pipeline, and differential streaming techniques.
AI-generated summary
Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD-CPU-GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., approximately 100M Gaussians) and standard in-memory training (e.g., approximately 11M Gaussians).
Community
Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD–CPU–GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., ∼100M Gaussians) and standard in-memory training (e.g., ∼11M Gaussians).
Website & More Visualizations: https://sponge-lab.github.io/TideGS
GitHub Repo: https://github.com/sponge-lab/TideGS
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2605.20150 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.20150 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.20150 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.