🌟 We have released the code for MagicClay. Generative mesh editing with SDS! The best part? It’s implemented as a Threestudio plugin, so it’s easy to install. Project page:

Moving to bluesky -->

What do you want to know?.

This work was led by Bo Sun, in collaboration with @AigermanNoam, Chen Song, @QixingHuang , and me. Check out our paper for more details: 7/.

We show generalizable results where TutteNet learns human reposing on SMPL data, and generalizes to SDF and real NeRF. 6/

The fast Jacobian computation facilitates elastic deformation. Here, we show examples of elastic NeRF deformation by manually choosing handles and target positions. We optimize the L2 loss on handles and As-Rigid-As-Possible loss on the other parts. 5/

Compared with other injective methods, TutteNet is more expressive and robust, with easy and fast Jacobian (gradient from deformed points to input points) computation. 4/

To parameterize 3D injective deformation, we propose to compose 2D injective deformations, i.e. Tutte embedding from different 3D orientations. We can parameterize Tutte Embedding (2D deformation) by discretizing the 2D domain in a coarse mesh and treating the Laplacian weights

The key motivation to learn injective deformation is deforming implicit neural fields, e.g. NeRF, SDF. Taking NeRF as an example, to query the point color and density in deformed space, we have to find where it is in the original space, i.e. find its inverse. This process

Injectivity is important for 3D deformations, as matter is not supposed to collapse on itself. The naive way of learning 3D deformation is a neural net that parameterizes all transformations of R^3 to itself without concern for the injective ones. Why is injectivity important,

This work was led by PhD candidate Amir Barda, in collaboration with @bermano_h @AigermanNoam Vova Kim, and me. Check out our paper for more details: 7/.

However, the un-selected parts remain identical. This means UV maps and animation rigs are preserved and can be transferred. 6/

During optimization, the mesh triangulation evolves to match the SDF in the editable region. 5/

Half-meshes because meshes are ubiquitous and easiER to manipulate. 4/.

Half-implicit, because implicits are more robust to noisy gradients, like those from SDS. We check this with a multi-view reconstruction experiment on synthetic data. Injecting varying amount of noise in the ground truth renderings, the neural implicits are more stable in the

The key to MagicClay is the hybrid representation: half-SDF, half-Mesh? Why two representations? 2/

If you're considering a PhD in graphics, this is a great opportunity! I had the chance to collaborate with Noam quite a bit, and I wholeheartedly recommend working with him.

🌟 tldr, 3D neural Deformation meets Score-Distillation Sampling ! Great job on MeshUp @hywkim ✨!.

Looking forward to chat at the posters or Dm me if you want to chat about research/collaboration/etc :).

Learning Continuous 3D Words for Text-to-Image Generation, @ChengTim0708 · @gadelha_m · @thibaultgroueix · Matthew Fisher · Radomir Mech · Andrew Markham · Niki Trigoni, Wed at 5pm #190. pdf: web: code:

NOPE: Novel Object Pose Estimation from a Single Image, @vannguyen_ng · @thibaultgroueix · Georgy Ponimatkin · Yinlin Hu · Renaud Marlet · Mathieu Salzmann · @VincentLepetit2, Thursday at 5 p.m, #321.web: code: