用行动赢得尊重

Alan Hou

Notes on coding, AI, and technology.

记录编程、人工智能与技术探索。

Back to Pixels: Lifting Latent Video Diffusion into Pixel Space

重返像素空间：摆脱 3D VAE 桎梏的原生像素级视频扩散模型 PixVid

Roblox, UC Merced, and UPenn researchers introduce PixVid, lifting pretrained latent video models (Wan2.1-14B) directly into native pixel space. By replacing the 3D VAE with a convolutional detailer head, predicting clean video x_0 instead of velocity, and aligning noise schedules, PixVid breaks the multi-frame block latency barrier. It achieves 2x resolution scaling (960x1664) with zero extra DiT tokens, enables frame-accurate depth editing under fast motion, and unlocks true next-frame responsiveness for real-time interactive world models.

Nov 14, 2026 Read article

Alan Hou

Back to Pixels: Lifting Latent Video Diffusion into Pixel Space

Recent Posts / 近期文章

Unbundling Multimodal Vision Encoding: Why Pretrained Priors Are Optional but Dedicated In-Fusion Capacity Is Key

Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies

A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets

BiVerify: Closing the Loop Between Backward and Forward Verification for Mathematical Self-Correction

GenRM-Flow: Video Generators Inherently Function as Process-Aware Reward Models in Flow Matching

Unpacking the Compute Efficiency of Strong-to-Weak OPD Against GRPO

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Benchmarks as Software: Terminal-Bench 2.1 and the Reality of Benchmark Maintenance Debt

Spectral and Subspace Dynamics of LoRA: Why Reinforcement Learning Demands Different Optimization Than SFT

Series / 系列文章

Nobel Prize in Economics (1969–2025)

Odoo 12 Development Cookbook

Odoo 14 Development Cookbook

TouchDesigner Tutorials

Mastering Python Scripting for System Administrators

Odoo 12 Development Essentials

Hung-yi Lee ML 2025

Dwarkesh Podcast Notes

Andrej Karpathy LLM Notes

freeCodeCamp