High-fidelity thin-shell 3D generation in about a second on a consumer GPU — it even runs on a laptop CPU, no server needed.</p>\n<p>🌐 Project page: <a href=\"https://ejshim.github.io/diffgi/\" rel=\"nofollow\">https://ejshim.github.io/diffgi/</a></p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/CrySTgZx8cebX0xN6lTCR.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/CrySTgZx8cebX0xN6lTCR.png\" alt=\"tsdf_vs_occupancy_labeled\"></a></p>\n<p>DiffGI replaces binary occupancy maps in geometry images with a continuous 2D TSDF and a Differentiable Marching Squares module, making the whole pipeline end-to-end differentiable.</p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/ztJmWgtP4clyJ_SIOUJ6D.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/ztJmWgtP4clyJ_SIOUJ6D.png\" alt=\"tsdf_boundary\"></a></p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/EIBqVuguPh5O6p4Y9X9ej.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/EIBqVuguPh5O6p4Y9X9ej.png\" alt=\"method\"></a></p>\n<p>Interactive demo coming soon — happy to answer any questions!</p>\n","updatedAt":"2026-07-21T01:33:38.770Z","author":{"_id":"640c8d7a04c679553cf7395b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/640c8d7a04c679553cf7395b/jDHPHIe-at-VErzp45nPx.jpeg","fullname":"Eungjune Shim","name":"ejshim","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6448408365249634},"editors":["ejshim"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/640c8d7a04c679553cf7395b/jDHPHIe-at-VErzp45nPx.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.13365","authors":[{"_id":"6a58da35203b48f49f41e526","user":{"_id":"640c8d7a04c679553cf7395b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/640c8d7a04c679553cf7395b/jDHPHIe-at-VErzp45nPx.jpeg","isPro":false,"fullname":"Eungjune Shim","user":"ejshim","type":"user","name":"ejshim"},"name":"Eungjune Shim","status":"claimed_verified","statusLastChangedAt":"2026-07-20T16:05:32.308Z","hidden":false},{"_id":"6a58da35203b48f49f41e527","name":"Hansol Lee","hidden":false},{"_id":"6a58da35203b48f49f41e528","name":"Eunjung Ju","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/640c8d7a04c679553cf7395b/mv3AWzw3QoAmHlIkpkc8h.mp4"],"publishedAt":"2026-07-15T00:00:00.000Z","submittedOnDailyAt":"2026-07-21T00:00:00.000Z","title":"DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation","submittedOnDailyBy":{"_id":"640c8d7a04c679553cf7395b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/640c8d7a04c679553cf7395b/jDHPHIe-at-VErzp45nPx.jpeg","isPro":false,"fullname":"Eungjune Shim","user":"ejshim","type":"user","name":"ejshim"},"summary":"Existing 3D generative models predominantly rely on implicit volumetric representations, which enforce watertight topology and struggle to represent thin-shell and non-manifold geometries such as garments. Geometry image-based approaches offer a surface-centric alternative, but existing methods rely on discrete binary occupancy maps whose resolution-dependent boundary encoding causes staircase artifacts and information loss upon downsampling, while surface reconstruction remains a non-differentiable post-processing step disconnected from the learning pipeline. To address this, we propose Differentiable Geometry Image (DiffGI), an end-to-end 3D-to-2D mapping framework that seamlessly integrates surface representation and geometric optimization. DiffGI replaces binary maps with a continuous 2D Truncated Signed Distance Function (TSDF), which encodes boundary position at subpixel precision within a fixed grid resolution, eliminating resolution-dependent staircase artifacts even under aggressive downsampling. Building on this continuous field, we introduce a differentiable Marching Squares algorithm based on analytical linear interpolation, allowing gradients from 3D surface losses to propagate back to the 2D latent space. Leveraging this differentiable pipeline, we train a DiffGI-VAE augmented with a geometry-aware normal rendering loss to compress complex 3D surfaces into an ultra-compact 32X32 latent space, and instantiate a transformer-based latent diffusion model with a flow-matching objective on top of this space for conditional 3D generation. Extensive experiments on garment and object datasets demonstrate that our method achieves superior reconstruction fidelity and boundary precision compared to prior geometry-image and voxel-based approaches, while requiring significantly fewer computational resources.","upvotes":6,"discussionId":"6a58da35203b48f49f41e529","projectPage":"https://ejshim.github.io/diffgi/","organization":{"_id":"6892dc78bda76685fb48c341","name":"clovir21","fullname":"CLO Virtual Fashion","avatar":"https://www.gravatar.com/avatar/05170d06e0160b53bfd0aad3f90a2aaf?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"640c8d7a04c679553cf7395b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/640c8d7a04c679553cf7395b/jDHPHIe-at-VErzp45nPx.jpeg","isPro":false,"fullname":"Eungjune Shim","user":"ejshim","type":"user"},{"_id":"64ef5e3e925c75f3e5b70575","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ef5e3e925c75f3e5b70575/KchS7VINIBqZiV3Xjp3OZ.jpeg","isPro":false,"fullname":"Milton Montiel","user":"miltmont","type":"user"},{"_id":"68ab96fa745427e0c0ac0f8f","avatarUrl":"/avatars/3a1c4dd06cb1fb87bd759327660fb8f3.svg","isPro":false,"fullname":"Elgohary mohamed khedr","user":"elgohary47","type":"user"},{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user"},{"_id":"64295d1f4e073875f6a605ac","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64295d1f4e073875f6a605ac/HEK8dluqPUhiARW-0eBZG.jpeg","isPro":true,"fullname":"Zhiyuan Ma","user":"ZhiyuanthePony","type":"user"},{"_id":"6860c1ea7d7039a1f4548043","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6860c1ea7d7039a1f4548043/4y1qs9nDODDeJ82dv4dcx.jpeg","isPro":false,"fullname":"Z","user":"vereneesa","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6892dc78bda76685fb48c341","name":"clovir21","fullname":"CLO Virtual Fashion","avatar":"https://www.gravatar.com/avatar/05170d06e0160b53bfd0aad3f90a2aaf?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.13365.md","query":{}}">
DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation
Abstract
Existing 3D generative models predominantly rely on implicit volumetric representations, which enforce watertight topology and struggle to represent thin-shell and non-manifold geometries such as garments. Geometry image-based approaches offer a surface-centric alternative, but existing methods rely on discrete binary occupancy maps whose resolution-dependent boundary encoding causes staircase artifacts and information loss upon downsampling, while surface reconstruction remains a non-differentiable post-processing step disconnected from the learning pipeline. To address this, we propose Differentiable Geometry Image (DiffGI), an end-to-end 3D-to-2D mapping framework that seamlessly integrates surface representation and geometric optimization. DiffGI replaces binary maps with a continuous 2D Truncated Signed Distance Function (TSDF), which encodes boundary position at subpixel precision within a fixed grid resolution, eliminating resolution-dependent staircase artifacts even under aggressive downsampling. Building on this continuous field, we introduce a differentiable Marching Squares algorithm based on analytical linear interpolation, allowing gradients from 3D surface losses to propagate back to the 2D latent space. Leveraging this differentiable pipeline, we train a DiffGI-VAE augmented with a geometry-aware normal rendering loss to compress complex 3D surfaces into an ultra-compact 32X32 latent space, and instantiate a transformer-based latent diffusion model with a flow-matching objective on top of this space for conditional 3D generation. Extensive experiments on garment and object datasets demonstrate that our method achieves superior reconstruction fidelity and boundary precision compared to prior geometry-image and voxel-based approaches, while requiring significantly fewer computational resources.
Community
High-fidelity thin-shell 3D generation in about a second on a consumer GPU — it even runs on a laptop CPU, no server needed.
🌐 Project page: https://ejshim.github.io/diffgi/
DiffGI replaces binary occupancy maps in geometry images with a continuous 2D TSDF and a Differentiable Marching Squares module, making the whole pipeline end-to-end differentiable.
Interactive demo coming soon — happy to answer any questions!
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/2607.13365 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.13365 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.13365 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.