Official Kuaishou tech account, providing real-time updates on the latest Kuaishou technology practices.
The paper introduces VidEmo, a multimodal video foundation model that uses a two‑stage emotion‑clue‑guided reasoning framework and a large emotion‑centric dataset (Emo‑CFG) to achieve state‑of‑the‑art performance on facial attribute, expression, and fine‑grained emotion tasks, surpassing Gemini 2.0.
This article presents a novel paradigm—Latent Reward Model (LRM) and Latent Preference Optimization (LPO)—that repurposes diffusion models as noise‑aware latent reward models for step‑level preference optimization, addressing the shortcomings of pixel‑level reward models, introducing multi‑preference consistent filtering, and demonstrating significant performance and efficiency gains on benchmarks such as PickScore and T2I‑CompBench++.
Kwai has open‑sourced its new flagship multimodal model Keye‑VL‑671B‑A37B, which upgrades visual perception, cross‑modal alignment and complex reasoning, achieving top scores on image, video, and mathematical reasoning benchmarks while detailing its architecture, three‑stage pre‑training, post‑training strategies, and future multimodal agent plans.
Flow‑GRPO introduces online reinforcement learning into flow‑matching models by converting deterministic ODE sampling to stochastic SDE sampling and reducing denoising steps, raising SD‑3.5‑Medium's GenEval accuracy from 63% to 95%—surpassing GPT‑4o—and demonstrating strong gains in complex composition, text rendering, and human‑preference alignment across multiple generative tasks.
This article details a new research pipeline that leverages large‑scale human preference data, a multi‑dimensional video reward model, and specialized alignment algorithms to dramatically improve video generation quality, motion fidelity, and text‑video consistency, with open‑source code and benchmarks for reproducibility.
This article explains how Kuaishou’s UniDex replaces traditional term‑based inverted indexes with model‑driven semantic posting lists and how the end‑to‑end UniSearch framework generates video IDs directly from queries, delivering higher relevance, lower latency, and significant online performance gains.
CoTyle introduces a novel open‑source framework that generates unique image styles from a numeric style code, eliminating the need for reference images, lengthy prompts, or LoRA modules, and demonstrates superior style consistency compared to existing solutions like Midjourney.
This article explains the over‑optimization problem in GRPO‑based flow models, analyzes why importance‑ratio clipping fails, and introduces GRPO‑Guard with RatioNorm and cross‑step gradient balancing, showing through extensive experiments that it stabilizes training and improves image quality across multiple diffusion backbones and tasks.
The paper identifies imbalanced concept distributions as the main obstacle to arbitrary concept‑combination in text‑to‑image/video generation, proposes the token‑level IMBA Distance and a lightweight IMBA Loss that adaptively re‑weights training tokens, and demonstrates through extensive experiments and a new Inert‑CompBench benchmark that this loss dramatically improves compositional ability without extra data.
KaiFG, Kuaishou's self‑built AI Feature Generator, unifies fragmented feature extraction frameworks, replaces slow C++ compilation cycles with Python‑level development, and achieves near‑C++ performance through Codon‑based compilation, reference‑counted memory management, and aggressive LLVM optimizations, dramatically shortening iteration time.
