Our recent finding on Diffusion Alignment: a reward model in pixel space can be easily transferred to score noisy diffusion latents directly — at small finetuning cost, via stitching. This makes Faster & Better for both Training & Inference Alignment. Meet StitchVM👇 1/

🚀 Excited to share my @GoogleDeepMind student researcher project: Dual-Rate Diffusion✨ ⚡ A simple construction that speeds up both regular diffusion and distilled models by interleaving a heavy context encoder with a light conditional denoiser. 🧵👇

For alignment you need V, but is hard to compute. Most methods try to approximate with 1) Tweedie, which is biased 2) MC roll-outs, which is slow with high var. Training V was often neglected since it's hard. We beg to differ. StitchVM enables this! Led by @gohyojun3 👇

Want a lighter yet stronger Point Transformer? Meet LitePT ✨ LitePT is a lightweight, high-performance 3D point cloud architecture for a wide range of point cloud processing tasks. Our smallest variant LitePT-S, features 3.6× fewer parameters, 2× faster runtime and 2× lower memory footprint than PTv3, while already matching or outperforming it across a range of benchmarks. 💻Code: github.com/prs-eth/LitePT 🌐Project page: litept.github.io/ 📰Paper: arxiv.org/abs/2512.13689 with Damien Robert, @jianyuan_wang , Sunghwan Hong, Jan Dirk Wegner, Christian Rupprecht, and Konrad Schindler

Combining video diffusion and 3D feedforward models by simply stiching them together in latent space - very cool idea! Make sure to check out this novel work from my collagues at Google and ETH!

Want to leverage the power of SOTA 3D models like VGGT & Video LDMs for 3D generation? Now you can! 🚀 Introducing VIST3A — we stitch pretrained video generators to 3D foundation models and align them via reward finetuning. 📄 arxiv.org/abs/2510.13454 🌐 gohyojun15.github.io/VIST3A

🎺Meet VIST3A — Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video Generator. ➡️ Paper: arxiv.org/abs/2510.13454 ➡️ Website: gohyojun15.github.io/VIST3A/ Collaboration between ETH & Google with Hyojun Go, @DNarnhofer, Goutam Bhat, @fedassa, and Konrad Schindler.

Even the SOTA VideoLLMs see videos in 1 fps, and you CANNOT perceive fine-grained motion 💃 with this frequency 🥲 📣 Presenting Video Parallel Scaling (VPS), an inference-time strategy that lets VideoLLMs see more frames by scaling compute in the parallel-axis 🤩

Excited to share that 3 papers are accepted to #ICCV2025 at EverEx 🎉 📌 SteerX: arxiv.org/abs/2503.12024 📌 VideoRFSplat: arxiv.org/abs/2503.15855 📌 CapeLLM: arxiv.org/abs/2411.06869 See you in Hawaii 🌴 👇 link to some threads

🚨Introducing VideoRFSplat📽️, a feed-forward text-to-3DGS generative model with high-quality scene-level results without post-optimization (e.g. SDS) Led by collaborators at EverEx AI - @gohyojun3, @bypark___, @namhyelin99, Byung-Hoon gohyojun15.github.io/VideoRF… A 🧵 👇 1/n

SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering @bypark___, @gohyojun3, @namhyelin99, Byung-Hoon Kim, @hyungjin_chung, Changick Kim tl;dr: MV-DUSt3R and MonST3R->geometric reward functions->geometric consistency arxiv.org/abs/2503.12024

3D consistent videos are hard to generate 🙁 What if we could steer them to be consistent during generation? Introducing SteerX🛞, a plug-and-play sampling method that works with *any* video diffusion to make videos physically plausible🤩 w/ @bypark___ @gohyojun3 @namhyelin99

SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis TL;DR: SplatFlow is a unified framework that combines a latent-space multi-view generator and a Gaussian Splatting Decoder to enable efficient 3D generation, editing, and inpainting directly from text prompts. Abstract (excerpt): SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts, thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks-including object editing, novel view synthesis, and camera pose estimation-within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks.