Yaofang Liu

🚀 ​​Pusa V1.0 Release Can you believe training a SOTA level Image-to-Video model with only $500 training cost? No way? But yes, we made it! And we achieved much more beyond that. We’re thrilled to release ​​Pusa V1.0​​—a paradigm shift in video generation, redefining video diffusion efficiency. With our novel ​​Vectorized Timestep Adaptation (VTA)​​ based on our prior FVDM work: 🔥 ​​Key Features: ✅Unprecedented Efficiency: - Surpasses Wan-I2V-14B with ≤ 1/200 of the training cost ($500 vs. ≥ $100,000) - Trained on a dataset ≤ 1/2500 of the size (4K vs. ≥ 10M samples) - Achieves a VBench-I2V score of 87.32% with 10 inference steps (vs. 86.86% for Wan-I2V-14B with 50 steps) ✅ Comprehensive Multi-task Support: VTA fully preserves Text-to-Video from the base model Wan-T2V, and after finetuning, Pusa V1.0 extends to the following all in a zero-shot way (no task-specific training): - Image-to-Video - Start-End Frames - Video completion/transitions - Video Extension - And more... ✅Complete Open-Source Release: - Full codebase and training/inference scripts - Model weights and dataset for Pusa V1.0 - Paper/ Tech Report with Detailed and Comprehensive Methodology 💡 ​​Scientific breakthrough​​: VTA enables granular temporal control via frame-level noise adaptation—no task-specific training needed. 🌍 ​​Fully open-sourced​​: • Codebase: github.com/Yaofang-Liu/Pusa-… • Project Page: yaofang-liu.github.io/Pusa_W… • Technical report: github.com/Yaofang-Liu/Pusa-… • Model weights: huggingface.co/RaphaelLiu/Pu… • Dataset: huggingface.co/datasets/Raph… [1/n]

hhhh I had similar tests. Even Astra tends to slack off and generates much worse scenes than DeepSeek, and way too slow…

For my coding tasks so far: DeepSeek 4.1 Flash 1017 (main) + Codex GPT-6 Astra (subagent) >> Astra as main. Much faster, cheaper, and better outputs

小打小闹，还是受限于没有自己的终端。传统终端厂商也还是有机会的，在模型能力逐渐收敛，体验日渐趋同的当下，伐毛换髓，彻底重构使用和交互逻辑，从体验上弯道超车是完全有可能的 我们也在尝试做一个很有意思的东西，相信用户体验绝对远超muse, instinct，openclaw, 乃至chatgpt...

看到这个的震惊程度不亚于小时候第一次使用计算器，发现速度和准确性远超自己🙃 习惯就好…

We looked at this months ago. Public since May, last commit about five months ago, a few dozen stars at the time. Funny to see it rediscovered today as the end of an entire industry 😂

Large-model agents like #GPT6 #Astra can now build editable 3D scenes directly from visual observations. The reconstructed scenes look plausible but can be significantly wrong! A curved corner becomes a square one. A passage has the wrong width. Distances drift. For embodied agents, these aren’t just visual imperfections—they change where and how a robot can move. Introducing AWSM (pronounced “awesome”): grounding agentic reconstruction in geometric evidence—IMU-informed scale cues, depth estimates, and camera poses—to build more faithful, editable worlds for mapping and simulation. Our multi-robot demo puts a reconstructed map to work: four robots follow predefined routes to reception and line up. The bigger opportunity? Worlds that phygital (physical + digital) agents can navigate, remember, interact with, and learn inside. From real spaces to worlds phygital agents can use. Reconstruction is just the beginning. Welcome discussions and collaboration 🤗 Blog: phygital-ai.github.io/agenti… Code (open-source): github.com/wentingw/AWSM Huge Credit to @wenting_is_here

GPT-6 Astra building 3D scenes is exciting. Physical agents using those worlds could be transformative

Release Flying-Agent harness code, an agent with a drone body. We think drones could be among the first practical bodies for AI agents. For this v0.1 version, describe the photo you want, the agent figures out where to fly, how to frame, and when to take the shot. Photography is just the beginning. More real-world tasks ahead. Still early, with efficiency, interaction, and applications to explore. Sharing to meet fellow builders and trade ideas. Huge Credit to Ziru @LiuZiru111 github.com/LZR-S/Flying-Agen…

nitter.cf/liuziru111/status/2098… this direction certainly will be more and more popular, many applications will emerge

we reduced GPT-6 Astra's scene reconstruction error by incorporating geometric constraints (video + IMU, and estimated depth). The reconstruction now seems near-usable for robots navigation and object search. I guess this kind of agentic scene modeling method could become a critical part of future physical agents

瞎琢磨了会儿，突然有点理解字节为什么这么重视 AI 了。 张一鸣十几年前看清了信息分发的逻辑，把推荐做成了时代的入口。如今，AI 开始改写入口背后的权力关系。 信息过剩之后，稀缺的是注意力；注意力被算法接管之后，稀缺的是自主性。 1.0 搜索时代：人找信息。 用户提出问题，平台组织答案。效率低。 2.0 推荐时代：信息找人。 算法从行为中预测欲望。人类在匮乏环境中形成的好奇心、地位焦虑和即时奖励偏好，撞上了无限供给的内容工业。异化人性，信息茧房。 3.0 AI 时代：人委托Agent。 AI 有机会理解情境、承接意图，帮助人协调即时欲望与长期意愿。AI可以辅助用户重新定义：看什么、为什么看、看到什么时候。当然，AI由谁设计和掌控很关键。A/这种极端公司很可怕。 当 AI 从平台的“流量捕手”变成用户的“认知代理”，古典 Feed 流平台将被彻底降维打击。

I bet drones will be the first AI agent bodies to catch on with consumers. Guys, what do you think?

Huge credit to @LiuZiru111! Check out the real drone demo here nitter.cf/liuziru111/status/2098…