Feiteng
@FeitengLi
Followers
728
Following
489
Media
158
Statuses
953
Speech & LLM & RL & Video 原生算法，写过后端，今年想写前端。 🤣🥵🤖😺 公众号 Generative AI 知乎 https://t.co/MXuiNWWVOO
Shanghai
Joined November 2016
I reproduced the results of VALL-E, the demo is here To avoid be misused, well-trained models and services will not be provided.
1
3
25
当你还在玩图像视频生成的时候….
What if you could not only watch a generated video, but explore it too? 🌐. Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt. From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵
0
0
1
RT @OpenBMB: 🚨 MiniCPM-V 4.0 is here! 🚨.✨Key features:.🏗️4.1B parameters.🧠Matches GPT-4.1-mini-20250414 on image understanding tasks in Ope….
0
48
0
RT @Alibaba_Qwen: 🚀 Meet Qwen-Image — a 20B MMDiT model for next-gen text-to-image generation. Especially strong at creating stunning graph….
0
624
0
RT @NagaSaiAbhinay: PR seems to be deleted but weights here: Interestingly they distilled the two transformers sep….
huggingface.co
0
1
0
Wan2.2-Lightning: Distill Wan2.2 Family into 4 Steps. 1. only 4 steps without the need of CFG trick, leading to x20 speed-up.2. The distilled model delivers visuals on par with the base model in most scenarios, sometimes even better.
github.com
1
0
3
🎬 High-Res Videos, Low Memory Costs.在笔记本电脑上生成384×384的视频？TITAN-Guide让你梦想成真！他们的低成本训练策略在保持空间-时间一致性的同时，大幅削减了VRAM的使用量。查看帧插值魔法（图9）和风格迁移演示（图8），这些演示均保留了文本对齐。代码和模型现可访问
0
0
0
🔧 TITAN-Guide 是如何工作的.TITAN-Guide 使用前向自动微分（forward AD）来估计梯度，而无需反向传播！这种巧妙的方法通过迭代将视频潜在变量（latents）调整到 t=0，从而避免了内存膨胀。它支持在消费级 GPU 上生成高分辨率（384×384）的视频，并在对齐得分上优于
1
0
0
🚀 Memory-Efficient Video Diffusion Breakthrough.TITAN-Guide 重新定义了文本到视频的生成指导！通过用正向梯度（forward gradients）替换反向传播（backpropagation），它将 GPU 内存使用量减少了 50%，同时在音视频对齐、风格迁移和帧插值方面取得了
1
0
0
📄 TITAN-Guide: Taming Inference-Time AligNment for Guided Text-to-Video Diffusion Models. 📄 论文: 🤖 GitHub Pages:
arxiv.org
In the recent development of conditional diffusion models still require heavy supervised fine-tuning for performing control on a category of tasks. Training-free conditioning via guidance with...
1
1
3
No More Redundant Steps!.🎯 全球计划将冗余操作减少了一半！在BabyAI（图6）中，PilotRL的自适应指导帮助代理避免了无目的的步骤。现在大型语言模型（LLMs）可以用更少的错误解决复杂任务。#AIResearch #ReinforcementLearning
0
0
0
Adaptive Global Planning = Better Coordination!.🧠 AdaPlan vs. ReAct：性能提升12.76%！通过将全局规划器和执行器整合到一个统一的模型中，PilotRL确保了紧密的协调和动态计划调整。详见图2。#MachineLearning #LLM
1
0
0
LLM Agents Just Got Smarter!.🚀 PilotRL 在代理任务中比 GPT-4o（3.60%）和 GPT-4o-mini（55.78%）表现更优秀！我们的三阶段渐进式强化学习（Reinforcement Learning）+ AdaPlan 理念实现了自适应全局规划，提升了长期决策能力。开源模型现在可以与闭源模型相匹敌。#AI #ReinforcementLearning
1
0
0
📄 PilotRL: Training Language Model Agents via Global Planning-Guided Progressive Reinforcement Learning. 📄 论文:
arxiv.org
Large Language Models (LLMs) have shown remarkable advancements in tackling agent-oriented tasks. Despite their potential, existing work faces challenges when deploying LLMs in agent-based...
1
0
0
🧩 Modular Agent Framework.如何在没有付费API的情况下构建代理？认知内核-Pro使用Python代码作为动作空间，并为子代理（网络、文件、工具）提供统一接口。其分层设计使得大型语言模型（LLM）和视觉语言模型（VLM）能力的无缝集成成为可能。查看图3中的框架架构！#AgentFoundationModels #TechTwitter
1
0
0
📄 Cognitive Kernel-Pro: A Framework for Deep Research Agents and Agent Foundation Models Training. 📄 论文: 💻 代码:
github.com
Contribute to Tencent/CognitiveKernel-Pro development by creating an account on GitHub.
2
0
2