Official Kuaishou Account
Fast.ai's Kwaipilot team unveiled OASIS, a 1.3B‑parameter code‑embedding model that, using only 5 million tokens, outperforms larger OpenAI embeddings across CodeSearchNet, CoSQA and AdvTest benchmarks, thanks to repository‑level program analysis, synthetic data generation, and a fused loss function.
This article details a comprehensive set of techniques—including data‑ and tensor‑parallel overlap, context‑parallelism, activation rematerialization, and a performance‑driven cost model—that dramatically improve large‑language‑model training efficiency on ultra‑large GPU clusters while preserving model quality.
Researchers from Kwai’s large-model team present a novel training system that combines pipeline-parallel-aware activation offloading with a compute-memory balanced checkpointing strategy, enabling lossless acceleration of large language models, achieving up to 42.7% MFU on 256 NVIDIA H800 GPUs while reducing memory usage.
Fast‑forwarding image‑to‑video generation, the article introduces I2V‑Adapter, a lightweight plug‑in for Stable Diffusion‑based video diffusion models that converts a single static image into a coherent video without altering the original T2V architecture, and details its design, frame‑similarity prior, experimental results, and real‑world applications.
Kwai (Kuaishou) has eight papers accepted at CVPR 2024 covering multi‑dimensional human preference for text‑to‑image generation, short‑video quality assessment, efficient video quality assessment, compressed video enhancement, conditional unsigned distance fields, universal cross‑domain retrieval, perception‑oriented frame interpolation, and test‑time energy adaptation.
DVIS introduces a decoupled video instance segmentation framework that splits the task into segmentation, tracking, and refinement modules, achieving state-of-the-art performance across VIS, VPS, and VSS benchmarks while maintaining low computational overhead, and demonstrates robustness in both online and offline settings.
This paper introduces HairStep, a novel intermediate representation combining Strand Maps and Depth Maps, and demonstrates how it reduces domain gap and improves single‑view 3D hair reconstruction accuracy across multiple algorithms, supported by new annotated datasets (HiSa, HiDa) and fair evaluation metrics.
At NVIDIA GTC 2023, Kuaishou engineers unveiled cutting‑edge solutions ranging from video quality assessment and enhancement, 3D digital‑human live streaming, a custom TensorRT‑based performance framework, large‑scale recommendation model acceleration, to multimodal massive‑model deployment for short‑video scenarios.
The article explores the rapid evolution of virtual digital humans, highlighting recent innovations, industry applications like Kuaishou's V‑Star program, academic research breakthroughs, and the technical and commercial challenges that must be overcome for these AI‑driven avatars to become mainstream.
This article explains how to leverage server‑side GPUs for hardware‑accelerated H.264 encoding and decoding with FFmpeg, covering installation, key API calls, format conversion to OpenGL textures, multi‑process considerations, and performance optimizations for cloud‑rendered visual effects.
