Tianshu Kuai (@TianshuKuai) profile picture

CS PhD Student @UMontreal and @Mila_Quebec | MSc and BASc @UofT

Montreal
Joined July 2022
Our model can also generate yoga poses for humanoid meshes that were never seen by the model during training. [7/8]
1
70
We can train our method to generate humans in random yoga poses. [6/8]
1
1
40
We sample discretization-agnostic noise by solving a screened Poisson equation w.r.t. Gaussian iid noise, which we prove is invariant under discretization changes in our paper. [5/8]
1
1
44
How do we achieve triangulation agnosticism? We couple a discretization-agnostic neural network backbone with a novel discretization-agnostic noise model. [4/8]
1
1
50
To achieve accurate generation on high-resolution meshes, we devise the first triangulation-agnostic flow matching model for meshes. [3/8]
1
66
This work focuses on generation over existing meshes, i.e., generating values on each vertex. For instance, in this figure the generated values define new XYZ positions on each vertex. [2/8]
1
106
📢 Introducing our #SIGGRAPH2026 work, Matérn Noise for Triangulation-Agnostic Flow Matching on Meshes. 🌐: matern-fm.github.io/ 📜: arxiv.org/abs/2605.19305 [1/8]
1
6
1
16
1,304
Our method only requires a small set of unpaired low-quality images for fine-tuning. We start to outperform the pre-trained model in perceptual quality by a large margin even with only 100 images, and all metrics continue to improve as the data size increases. (6/7)
1
85
Our best fine-tuned model outperforms all baseline methods and can output artifact-free restorations. (5/7)
1
15
Our approach consistently improves pre-trained models of different architectures on out-of-distribution degradation types. (4/7)
1
21
We generate pseudo targets by leveraging a pre-trained diffusion model. This process involves low frequency content constrained denoising, followed by unconditional denoising. The generated pseudo targets are then used to fine-tune the pre-trained model. (3/7)
1
27
In this work, we introduce a framework for adapting a pre-trained face restoration model to work on inputs with unseen and out-of-distribution degradation types. Our unsupervised pipeline only requires a small collection of unpaired images with target degradation type. (2/7)
1
33