Eungjune Shim (@ej__shim) profile picture

Machine Learning Engineer @itsclo3d (CLO Virtual Fashion) · 3D generation & geometry processing Linkedin : https://teapawt.net/t.co/XgKTK2THSC

Seoul
Joined April 2026
Photo by Eungjune Shim (@ej__shim)Photo by Eungjune Shim (@ej__shim)Photo by Eungjune Shim (@ej__shim)Photo by Eungjune Shim (@ej__shim)
High-fidelity thin-shell 3D in about a second on a consumer GPU — it even runs on a laptop CPU. No server needed. Introducing DiffGI: a fully differentiable geometry-image framework for end-to-end 3D generation. Accepted to #ECCV2026 🇸🇪 🧵👇
8
42
309
20,736
5/ Why so fast? Subpixel TSDF boundaries survive compression, so a 256×256 geometry image lives in a 32×32 latent — and diffusion runs on a tiny tensor: 0.8s on an A6000, 1.2s on an RTX 4070, 8.5s on a MacBook M4, CPU-only. In-browser WebGPU demo coming soon.
1
261
4/ What this buys us: the same garment recovered from our representation and pipeline (left) vs. a conventional geometry image (right). Occupancy tears thin shells and aliases at boundaries. DiffGI keeps them sharp and intact.
287
3/ Full pipeline: mesh → multi-chart TSDF geometry image → DiffGI-VAE compresses it into a compact 32×32 latent → a transformer-based latent diffusion model generates in that space, conditioned on a single image or labels.
3
322
2/ Our fix: replace the binary occupancy map with a continuous 2D TSDF, and extract the surface with a Differentiable Marching Squares module. 3D surface losses — including a normal rendering loss — now backprop end-to-end to the 2D latent. Boundaries become subpixel-accurate.
6
443
1/ Most 3D generators rely on watertight volumetric representations (SDF, occupancy). But garments and many real objects are thin shells — geometry those models fundamentally can't represent. Geometry images can. The catch: binary occupancy masks tear and alias at boundaries.
1
1
581