RT @YasutakaFuruka1: MapTracker: Online Consistent Vector HD Mapping by my students (Jiacheng Chen, Yuefan Wu, and Jiaqi Tan) and colleague….

RT @YasutakaFuruka1: Thank you for posting @_akhaliq .🚀 Excited to share Compressive Light-Field Tokens (CLiFT) — a new scene representatio….

RT @ShivamDuggal4: Compression is the heart of intelligence.From Occam to Kolmogorov—shorter programs=smarter representations. Meet KARL: K….

RT @MattNiessner: Stunning voice model by @synthesiaIO:. *𝐄𝐗𝐏𝐑𝐄𝐒𝐒-𝐕𝐨𝐢𝐜𝐞*. -> new SotA that perseveres identity, accent, expressiveness w/o….

RT @n_karaev: ⚡️Today we’re releasing SpatialTrackerV2, the first feedforward model for dynamic 3D reconstruction and point tracking in the….

RT @emmanuel_2m: All of these 3D objects were generated entirely via AI (each from a single image) and the results are reaching exceptional….

RT @gabriberton: I can't stress enough how useful this trick has been for me in all these years. It reduces GPU memory by N equal the numbe….

More results: We adapt to dynamic light changes (Img 1) and difficult expressions (Img 2) whereas portrait animation methods might find it difficult. Comparing to single image recon, the extra reference view helps improve identity coherence and reduce artifacts (Img 3) (7/7)

We remove such distortions and improve occluded regions by leveraging an additional frontal reference image of the person. We use this reference to produce an warping that undistorts the raw reconstruction, and we use this reference image to help recover occluded regions (6/n)

The core challenge is achieving temporally consistent reconstructions despite the varying head pose. As shown here, capturing the user from the sides can produce stretching/distortion and artifacts, making the lady’s face rounder. (5/n)

2. Coherent3D improves temporal consistency for single-view 3D portrait reconstruction, while retaining more authentic expression changes and dynamic lighting/shoulder poses than mesh-driven portrait animation methods. The core challenge is … (3/n)

More results

Insight: prior works found it difficult to solve focal length+3D translation+3D human from a single image. We realize that it’s solvable by first accurately estimating metric pelvis depth, then 3D mesh, and finally 3D translation and focal.

1. BLADE solves for both 3D human pose and camera, which is often neglected in prior works but very important for close range images with severe perspective distortion. BLADE achieves both accurate 2D alignment and 3D pose (2/n)

At CVPR2025, I presented 2 of our papers:. 1. BLADE: Single-view Body Mesh Learning through Accurate Depth Estimation. 2. Coherent 3D Portrait Video Reconstruction via Triplane Fusion. 1/n

RT @iam_NCJ: What if a Transformer could render?.Not text → image. But mesh → image — with global illumination. No rasterizers. No ray-tra….

RT @taiyasaki: This paper is awesome. 1) it fine tunes a video model on just 2 GPUs, opening academic research on the topic.2) the trainin….

RT @youngjoongkwon: I’ll be recruiting Ph.D. and Master’s students for Fall 2026. I’ll also be at CVPR this year-happy to connect there!.

RT @y0b1byte:

RT @BrianRoemmele: BOOM!. FREE Text to multi-guest podcast AI. Open Source Nari Labs with Dia-1.6b just beat podcast-style clips on Google….