New paper: AsymFlow๐ฅ
JiT x0-prediction is not enough for pixel generation. Better keep velocity in a low-rank subspace:
- 1.57 FID on ImageNet (best pixel flow model)
- Finetunes FLUX.2 klein into pixel space, beats the original on HPSv3/DPG/GenEval (#1 overall on HPSv3)
1/7
Preparing a major PR to improve @ThreeJS MaterialX support significantly. In order to perfect it, I've written an open source @MaterialXcg Fidelity Suite inspired by the @thekhronosgroup / @glTF3D fidelity tool:
https://t.co/WTURPMEHRe
Captured a vid of sampling a tiny diffusion model with the neural engine (A17 Pro) on my phone. Not very fast but pretty acceptable given everything happened on the iPhone locally
โ๏ธ Tiny diffusion model on @Apple iPhone Neural Engine (ANE)
Converted one of my Tiny checkpoints into a CoreML package, and this image was generated in 3.78 seconds with memory use under 140MB on my iPhone 15 Pro.
Local, private and energy-efficient ๐ฅ
โ ๐๐ข๐ง๐ฒ ๐๐ข๐๐๐ฎ๐ฌ๐ข๐จ๐ง ๐๐จ๐๐๐ฅ๐ฌ I trained before are fairly compact, the left on was from a ๐.๐ ๐๐ one and the right was from a ๐๐ ๐๐ one. It is fun to witness how a model learns the statistics which enables generative sampling of that photographic experience.
HDRify, the pure JavaScript implementation of EXR, HDR, and UltraHDR Jpgs, now has its own VSCode & Cursor extensions.
Preview (with tone mapping/exposure) and convert between HDR, EXR and UltraHDR Jpgs.
Link below.
It's been a long time coming but the latest version of three-mesh-bvh brings support for out-of-the-box Line & Point cloud BVH support! Now all your geometries can be fast ๐
1/3
#threejs#webgl#javascript
Splannequin: Freezing Monocular Mannequin-Challenge Footage with Dual-Detection Splatting
TL;DR: How to freeze a dynamic molecular scene when people inevitably move.
Contributions:
โข A Novel Problem Formulation and Benchmark: We are the first to formally address synthesizing high-fidelity, freeze-time videos from monocular MC footage, providing a new benchmark and evaluation protocol.
โข A Targeted Regularization Framework: We propose a novel method to identify and regularize hidden and defective Gaussians, the primary sources of temporal artifacts, anchoring them to reliable past or future states.
โข State-of-the-art Performance with Zero Inference Overhead: We improve visual quality and stability in existing methods without architectural changes. As the deformation runs only once for a target instant, we achieve inference speeds exceeding 280 FPS on an RTX 4090.
Every cape is one bug away from becoming a scarf. ๐ฆธโโ๏ธ
But one scientist finally fixed it - millions of ribbons, noodles, and fabrics twisting together. Human ingenuity at its best! What a time to be alive! Full video: https://t.co/kqrnLYt7Hf
My talk from @BetterSoftwareC last week is up on youtube. I present my findings on thread synchronization and job systems that I learned while parallelizing the physics solver. https://t.co/sWqeAojze9