Keep on to blur preview images; turn off to show them clearly
微软新3D建模大模型 Trellis.2 实测 微软刚刚发布了新模型 trellis 2, 4B 这是一个通过图片就能创建3D模型的大模型, 它使用了一个叫"基于稀疏体素的 3D VAE 流匹配变换器"的新方法, 从官网看效果很惊艳, 但是打开它的demo, 就能发现一些问题, 模型有孔洞, 可以看这个官方demo, 雕像细节处有很多镂空的部分. 我连续测试了几张图片均能复现这个问题. 我的猜测是可能跟这个模型支持透明度有关, 官方介绍说"TRELLIS 能够重建并生成具有复杂拓扑结构、锐利特征和完整物理基础渲染（PBR）材质的任意 3D 资产——包括透明度和半透明效果。" 但这个透明度如果控制不好, 就变成模型孔洞了 总结, 这个模型用于粗略建模是可以的, 我实际运行中他的速度很快, 毕竟模型只有4B, 但是细节是需要手动修正的, 尤其是提到的孔洞和幻觉问题, 很难直接用于3D打印.
A coder, road bike rider, server fortune teller, electronic waste collector, co-founder of KCORES, ex-director at IllaSoft, KingsoftOffice, Juejin.
RT @0xThiru: Had a great call with @VittoStack last week about dev resources for new SDKs and Protocols like x402, ERC- 8004 > devs are so…
Product & Devs Growth @Cyfrin | Ex @Alchemy | Created @cyfrinupdraft and @AlchemyLearn | Robotics | Making web3 mainstream
I think «Pax Americana» officially ended in March 18, 2021, at Anchorage, Alaska, when Blinken and Sullivan were preening about rules-based international order and moralizing on HK/Xinjiang/Tibet, and got hit with this. It wasn't received very seriously at the time.
very interesting in retrospect The perception of dominance was so strong https://t.co/4BeyG4GeA4
RT @YueYuanwen: Want a lighter yet stronger Point Transformer? Meet LitePT ✨ LitePT is a lightweight, high-performance 3D point cloud arch…
Founder and CEO of https://t.co/5MjtfpwEU3 | Your guide to radiance fields | Host of the podcast @ViewDependent | FTP: 279 | discord: https://t.co/lrl64WGvlD
RT @ilyasu: Dec 10: Cambridge, MA cancels Flock contract. Dec 18: Flock finds the killer of an MIT professor.
GP @a16z — Building American Dynamism 🇺🇸 — Anthropologist — Formerly Founder/CEO @OpenDNS — Lokah Samastah Sukhino Bhavantu
![Using Gaussian Splats to Create High-Fidelity Facial Geometry and Texture
• We propose two key modifications of Gaussian Splatting to enable accurate triangulated surface reconstruction:
- Soft constraints encourage Gaussians to be more tightly coupled to the underlying mesh, allowing Gaussian perturbations to more accurately drive mesh deformation.
- Segmentation annotations supervise the Gaussians, ensuring they do not attempt to explain regions of the target image they shouldn't be associated with.
• We present a method to disentangle albedo textures from lighting and normals. The PCA coefficients of the textured mesh are optimized to capture as much albedo color as possible while minimizing the contributions from the relightable Gaussians used to capture differences between the synthetic rendering of the mesh and the target image.
• Our method avoids the need for controlled capture setups and instead requires only commodity hardware and a limited number of views. The flexibility of our approach allows for joint training using data from different capture setups; for example, we combine images from our capture method with those obtained from so-called flashlight capture (see e.g. [44]).
• Finally, and perhaps most importantly, we illustrate that the geometry obtained via our pipeline is accurate enough to use with a view-dependent neural texture. We propose a novel Gaussian Splatting approach to view-dependent neural textures, enabling the use of high visual fidelity Gaussian Splatting on any asset in a scene without needing to modify any other asset or any aspect (geometry, lighting, renderer, etc.) of the graphics pipeline.](/_next/image?url=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FG8g-0Y0akAQLVmx.jpg&w=3840&q=75 "Using Gaussian Splats to Create High-Fidelity Facial Geometry and Texture
• We propose two key modifications of Gaussian Splatting to enable accurate triangulated surface reconstruction:
- Soft constraints encourage Gaussians to be more tightly coupled to the underlying mesh, allowing Gaussian perturbations to more accurately drive mesh deformation.
- Segmentation annotations supervise the Gaussians, ensuring they do not attempt to explain regions of the target image they shouldn't be associated with.
• We present a method to disentangle albedo textures from lighting and normals. The PCA coefficients of the textured mesh are optimized to capture as much albedo color as possible while minimizing the contributions from the relightable Gaussians used to capture differences between the synthetic rendering of the mesh and the target image.
• Our method avoids the need for controlled capture setups and instead requires only commodity hardware and a limited number of views. The flexibility of our approach allows for joint training using data from different capture setups; for example, we combine images from our capture method with those obtained from so-called flashlight capture (see e.g. [44]).
• Finally, and perhaps most importantly, we illustrate that the geometry obtained via our pipeline is accurate enough to use with a view-dependent neural texture. We propose a novel Gaussian Splatting approach to view-dependent neural textures, enabling the use of high visual fidelity Gaussian Splatting on any asset in a scene without needing to modify any other asset or any aspect (geometry, lighting, renderer, etc.) of the graphics pipeline.")
Using Gaussian Splats to Create High-Fidelity Facial Geometry and Texture • We propose two key modifications of Gaussian Splatting to enable accurate triangulated surface reconstruction: - Soft constraints encourage Gaussians to be more tightly coupled to the underlying mesh, allowing Gaussian perturbations to more accurately drive mesh deformation. - Segmentation annotations supervise the Gaussians, ensuring they do not attempt to explain regions of the target image they shouldn't be associated with. • We present a method to disentangle albedo textures from lighting and normals. The PCA coefficients of the textured mesh are optimized to capture as much albedo color as possible while minimizing the contributions from the relightable Gaussians used to capture differences between the synthetic rendering of the mesh and the target image. • Our method avoids the need for controlled capture setups and instead requires only commodity hardware and a limited number of views. The flexibility of our approach allows for joint training using data from different capture setups; for example, we combine images from our capture method with those obtained from so-called flashlight capture (see e.g. [44]). • Finally, and perhaps most importantly, we illustrate that the geometry obtained via our pipeline is accurate enough to use with a view-dependent neural texture. We propose a novel Gaussian Splatting approach to view-dependent neural textures, enabling the use of high visual fidelity Gaussian Splatting on any asset in a scene without needing to modify any other asset or any aspect (geometry, lighting, renderer, etc.) of the graphics pipeline.
Paper: https://t.co/HsrhcZGcAF