meakywang (@wangfengbin) profile picture
sea
Joined July 2017
Photo by meakywang (@wangfengbin)Photo by meakywang (@wangfengbin)
meakywang retweeted
Make InstantHMR support in ComfyUI @ComfyUI . github.com/jtydhr88/ComfyUI-… 3D human bodies from photos, video, or a live webcam / phone feed. About 2.5 ms per person, 40+ fps live, on the same MHR body as SAM 3D Body, so it plugs straight into the native nodes. github.com/jtydhr88/ComfyUI-…
We release InstantHMR v1.0: Real-Time Human Mesh Recovery on Your Phone. It recovers the full 3D body of every person in an image in 5.5 ms per person on a Galaxy S23 NPU, and 1 ms on an RTX 4070, the fastest of the 11 methods we measured, on GPU, NPU and CPU. 40.0 mm PA-MPJPE on 3DPW, 48.8 mm on EMDB and 41.7 mm on RICH. Code under Apache-2.0, weights for research and for commercial use. 🧵 💻 github.com/Naturalpad-studio… 🤗 huggingface.co/momolesang/In… 📄 hal.science/hal-05854387
7
23
3
208
9,439
meakywang retweeted
终于成了 在免费的Google Colab上面部署 DeepSeek-V4-Pro + Qwen3.5 9B破甲大模型
来了个大的 Hugging Face 上最近杀出一个核弹级无审查模型,直接被挤爆了 DeepSeek-V4-Pro + Qwen3.5 9B 的混血完全体,体积只有 5.6GB,上线就被狂下 1.8 万次。 它的测试结果极其残暴: · 拒答率暴跌:实测 100 个刁钻测试提示词,原版直接 100% 拒答,它硬生生只拒了 3 个 · 顶配推理蒸馏:用 DeepSeek-V4-Pro 约 25 万条高质量数据蒸馏,逻辑与数学推理能力完全拉满 · 零损耗去审查:底层用 Heretic 的 ARA 算法消融限制,彻底卸掉道德锁,智商分毫不损 · 普通电脑本地直跑:Q4_K_M 的 GGUF 格式仅 5.6GB,显存要求极低,核显本和轻薄本也能秒开 完全开源免费,Apache 2.0 协议,想在本地跑一个完全听话、推理极强的大模型的兄弟建议直接马住! 模型体验:huggingface.co/xtools-at/Dee… 你平时在本地跑大模型,遇到拒答最严重的是写代码还是角色扮演? #AI #大模型 #开源
Readers added context they thought people might want to know
The author claims this model caused Hugging Face to crash from overload, but to date this model only has 772 downloads. This is an oddly sensationalist claim, and given the risk of running models locally as weights can contain dangerous/deceptive behavior, beware running this. huggingface.co/xtools-at/Deep…
25
52
1
347
45,838
meakywang retweeted
Learn the core concepts of Copernicus in Houdini in this new tutorial series by Peter Arcara. Intro to Copernicus Concepts covers procedural texture creation, pattern generation, geometry rasterization, and custom simulations. A beginner-friendly way to get comfortable working in the COPs context. Watch now: youtube.com/playlist?list=PL… #sidefxhoudini #copernicus #procedural #texturing
3
14
91
3,352
meakywang retweeted
🛠️ You can now give Claude Code or Codex a complete set of professional Blender workflows, from modeling and rigging to animation, texturing, optimization, and game-engine export. Meet Blender Agent Skills, an open-source collection of 28 specialized skills that teach AI coding agents how to work with Blender 5.2 and newer. And the important distinction is that these aren't generic prompts telling the model to "be a 3D artist." They include executable scripts, technical instructions, and validation procedures tested in Blender itself. 🎮 What can an agent do? ▪️ Create hard-surface and organic 3D models. ▪️ Generate procedural geometry using Geometry Nodes. ▪️ Configure realistic PBR materials. ▪️ Perform UV unwrapping and texture baking. ▪️ Rig characters with Rigify. ▪️ Create facial shape keys and animation controls. ▪️ Retarget motion-capture animations. ▪️ Prepare hairstyles and game-ready hair cards. ▪️ Optimize meshes with LODs and polygon budgets. ▪️ Export assets for Unity, Unreal, Godot, Roblox, and the web. 🧠 How does it work? The project provides specialized agent skills for individual stages of a 3D production pipeline. For example, if you ask: "Create a stylized game character, add a walk cycle, and export it to Godot." The agent can load the relevant modeling, rigging, animation, and export instructions instead of improvising the entire workflow. Each skill contains guidance about Blender APIs, common errors, required parameters, and validation steps. 🔍 Quality assurance matters One particularly useful component is Blender Asset QA. It can detect issues such as: ▪️ Broken or inverted normals. ▪️ Floating mesh parts. ▪️ Z-fighting. ▪️ Missing textures. ▪️ Invalid skin weights. ▪️ Problematic export settings. The repository reports 67 automated Blender tests, covering the included scripts and workflows. 🔌 Supported agents ▪️ Claude Code. ▪️ OpenAI Codex. ▪️ Cursor. ▪️ Gemini CLI. ▪️ GitHub Copilot. ▪️ OpenCode. The skills can run Blender in the background using Python, or interact with an open Blender session through a compatible MCP server. 💻 Availability ▪️ MIT license. ▪️ Blender 5.2 LTS or newer. ▪️ Cross-platform installation scripts. ▪️ No proprietary Blender automation service required. However, using an external coding model may still require a subscription or API credits. ⚠️ Limitations These are workflows and tools, not a specialized 3D foundation model. The agent's results still depend on the underlying model, the task, and the available Blender environment. Complex character modeling, production rigging, and visual quality continue to require human inspection. The bigger opportunity: AI agents don't necessarily need to learn every 3D workflow from scratch. They can use tested, reusable procedures to turn familiar software into a more accessible production tool. 💻 github.com/EL4CTEO/blender-s… #Blender3D #ClaudeCode #CodingAgents #GameDev
4
1
31
2,267
meakywang retweeted
HMマップ用のEeveeノードをアップデートしました。Shader to RGBではなく、法線とライトの方向(Dot Product)を使用してライティングを計算するようになりました。 Updated the Eevee nodes for the HM map. Now they calculate lighting via normals and light direction (Dot Product) instead of Shader to RGB. The old version was simpler, but finicky. The new one is a bit more complex under the hood, but it works much more predictably. The shader file is available for free on my pages linked in my profile! 旧バージョンはシンプルでしたが、少し扱いづらい部分がありました。新バージョンは内部の構造がやや複雑になったものの、より予測しやすく安定して動作します。シェーダーファイルはプロフィールのリンク先(各プラットフォーム)にて無料で配布中です! teapaw.org/Weytant_V/status/20849… #b3d #3DCG #Shader #Anime3D #WutheringWaves
2
173
2
1,777
26,495
meakywang retweeted
Arcane 2DFX demoreel by Stéphane Chung.
1
34
380
5,568
meakywang retweeted
MOSS Image派生の画像生成モデル「FineView 欢视 | MOSS 4」が更新。 最新版「fêtesnap」では超写実的なスナップ写真風の質感を強化。映画的な明暗比や成人向け描写に対応し、テキスト描画や構図制御も両立。 #MOSSImage #ComfyUI URLはリプ⬇️
1
1
25
2,785
🎬 Turn ordinary videos into editable 3D character animations, fix sliding feet, and export directly to Blender, Unity, or Unreal. Meet RGB MoCap, an open-source Windows application that brings multiple AI motion capture models into one practical animation workflow. Most video-to-motion systems have a frustrating problem: even when the overall movement looks correct, characters often slide across the floor, float above it, or place their hands incorrectly. RGB MoCap lets you fix these problems after reconstruction, without running the AI model again. 🧠 Multiple motion capture engines The application supports: ▪️ ExMR: Full-body reconstruction with detailed finger and foot motion. ▪️ FlowHMR: Human motion reconstruction with articulated hands. ▪️ GVHMR: World-space motion estimation, with optional hand refinement. ▪️ MAMMA: Motion capture from multiple synchronized cameras. ▪️ PromptHMR: Joint reconstruction of multiple interacting people. ▪️ PEAR: Experimental real-time body and finger tracking from a webcam or video. The last two offline models require additional setup. PEAR is currently limited to live preview and doesn't support the normal export pipeline. 🦶 Fix the details that AI gets wrong After motion reconstruction, you can: ▪️ Inspect the original footage alongside the reconstructed skeleton. ▪️ Identify when feet or palms should remain in contact with a surface. ▪️ Manually edit contact intervals on a timeline. ▪️ Apply corrections to reduce foot sliding and floating. ▪️ Compare the original and corrected animations. ▪️ Reapply corrections without repeating expensive AI inference. This is particularly useful for walking, running, crouching, and interactions with the environment. 🎮 Export actual animation assets RGB MoCap supports: ▪️ FBX: Export character animations for Blender, Unity, and Unreal workflows. ▪️ BVH: Skeletal animation with Generic/Blender and 3ds Max Biped profiles. ▪️ VRM: Retarget motion to humanoid avatar rigs. ▪️ PMX → VMD: Export motion for MikuMikuDance. ▪️ Blender projects: Save animation, source footage, and camera information together. You can also retarget captured motion to existing Mixamo characters without rerunning reconstruction. 🤖 Works with coding agents The repository includes AGENTS.md and a dedicated guide for Codex and other coding agents. An agent can help install dependencies, configure models, troubleshoot errors, and prepare the application. However, this isn't a fully autonomous MCP animation pipeline. The supported workflow still centers on the graphical Windows application. 💻 Requirements ▪️ Windows 64-bit. ▪️ Compatible NVIDIA GPU and OpenGL 4.3. ▪️ Blender and FFmpeg, installed automatically by the setup script. ▪️ Approximately 42 GiB for the three main model runtimes and their supporting files. ▪️ Additional storage for optional models, projects, and caches. The application's own code is MIT-licensed, but individual AI models have separate licenses and usage restrictions. The interesting part: AI motion capture is becoming less about generating a rough skeleton and more about producing animation that artists can actually correct, retarget, and use. For indie game developers and 3D animators, that's a meaningful improvement. 💻 GitHub: github.com/toyxyz/toyxyz_RGB… #MotionCapture #GameDev #3DAnimation #OpenSource
11
57
5
495
45,335
WIP 간단하게 몇장만 그려서 만드는 SDF 메이커
7
274
16
2,482
98,212
meakywang retweeted
K/DA Firestarter x Fortiche ‘icons in motion’ video gave a behind the scenes look at a storyboard animatic, paired here with the final animation. #kda
10
764
30
6,333
94,940
meakywang retweeted
While League of Legends fans are enjoying K/DA's new FIRESTARTER music video, we asked @ForticheProd's director Marietta Ren and animation supervisor Rémy Terreaux about the work on the animation, VFX, and camera, found out what 3D software the studio uses, and learned what technique the team implemented for the first time. "The 3D effects are designed to look like paint; we add textures and color nuances that would be difficult to achieve in 2D without spending a great deal of time on them. We then blend these with 2D effects that reshape the overall form of the fire and make it more graphic. We switch back and forth between the two techniques, with the goal of making it nearly impossible to tell the difference between them. This is particularly true for the wave-shaped fire and the big explosion in the middle of the clip. All the FX artists came up with ideas to refine this fire. It’s a true collaborative effort!" See how the MV compares to Arcane: 80.lv/articles/take-a-look-b…
118
2
799
22,607
meakywang retweeted
おもしろいから真似してみた。 minimax H3 (ローカル) RGBMoCapで動画からモーションとって blenderでカメラつけてUIごとキャプってCharacter-Swap-LoRA使った。 似たような感じにはなったけどクオリティが低い😅 #生成AI
次世代AI? あんまりblender触らないのでちょっと面倒だった… #生成AI
22
145
11
1,511
127,734
meakywang retweeted
spatial outpainting
45
136
27
1,678
103,501
Ah!! We've been trying to get this working for months. Basically, we can now capture a person and the objects they're interacting with in 3D from regular video. Motion for both, plus approximate object shape. I've been excited about this for a while because so much of what people do involves stuff. If you want a robot to learn how to use a tool, or to understand someone's tennis swing, you probably want to know what the tool or racket is doing too. Same goes for props in animation and everyday tasks in occupational therapy. There's a lot here. Still imperfect, but this week it started working well enough that I wanted to show you.
50
72
8
985
42,056
meakywang retweeted
🎬 MiniMax H3 做连续剧情,人物和场景经常前后对不上?这个插件专门为视频连续性做了一套工作室。 ComfyUI-H3-Continuity-Studio:把分镜时间轴、视觉记忆、故事编写和连续性检查整合进 ComfyUI。 重点看 4 个能力: 🎞️ 1. 可视化分镜时间轴 把长视频拆成多个镜头片段,每段独立设置提示词、时长和生成任务,还能管理镜头之间的衔接关系。 🧠 2. Visual Memory 视觉记忆 可以把前面镜头的重要画面带到后续镜头,即使换了机位、离开原场景,也能继续引用之前的视觉信息。 ✍️ 3. AI 故事编写 接入 OpenAI 兼容的本地或云端模型,让 AI 帮你规划故事、编写分镜和镜头提示词。 🔍 4. 自动连续性检查 生成完成后,Vision 模型可以对照人物设定、参考图片和前后镜头,检查人物、道具、场景及镜头衔接是否出现明显偏差。 发现问题,可以单独重新生成对应镜头,不必整条视频推倒重来。 适合谁用? 使用 ComfyUI + MiniMax H3 制作连续剧情、AI 短片、电影分镜和多场景纪录片的创作者。 🔗 GitHub: github.com/KadoBOT/ComfyUI-H… 🎬 官方演示视频 ⚠️ 两个实际门槛: 需要新版 ComfyUI 和已安装的 H3 模型,示例工作流不会自动下载模型权重。 AI 连续性检查依赖单独配置的 Vision 模型;它能检查人物、道具和场景差异,但不能保证镜头运动、构图和画面细节完全一致。 #ComfyUI #MiniMaxH3 #AI视频 #人物一致性 #开源工具
11
28
197
10,483
美少女土豆姐 @poteto 是全网最会用 Grok Bot 的人之一。这段时间,我从她的访谈和长文里学了不少实用技巧。 后来想了想,与其这样碎片化地接收,不如把她所有的演讲、访谈、直播、长文,还有三百多条推文,整合成一部成体系的《Grok Bot 方法论》👇
131
455
57
2,358
210,820
meakywang retweeted
影と社 \ Shadow and Shrine Seedance 2.5 Prompt in the replies. 👇
26
57
10
590
49,836
大学教授可能没想到, GitHub 上已经有人把私人家教开源了... 这个项目叫 DeepTutor,由香港大学数据智能实验室开发,目前在 GitHub 上已经获得超过 4.8 万颗 Star! 它不是普通的 AI 聊天机器人,而是一套完整的 AI 个性化学习系统。 你可以把教材、论文、PDF 扔进去,它就能围绕你的资料建立知识库,讲解知识点、解答问题,甚至自动生成练习题,检查你到底学会了没有。 更夸张的是,它支持深度研究、交互式可视化、沉浸式阅读,还能根据学习进度规划掌握路径。 如果你学高数,它可以辅助拆解公式; 如果你学外语,它可以陪你练习; 如果你研究论文,它还能结合资料帮你梳理知识体系。 它甚至支持长期记忆、接入 QQ 和微信等聊天渠道,让 AI 老师不再局限于一个网页。 项目采用 Apache 2.0 开源协议,支持本地部署,不过调用部分 AI 模型可能产生费用。 未来的教育, 也许是每个人都有一个真正了解自己的 AI 老师: github.com/HKUDS/DeepTutor
14
152
7
822
37,388
meakywang retweeted
I've added MAMMA support to RGBMocap, an AI motion capture tool, enabling multi-camera input. #mocap github.com/toyxyz/toyxyz_RGB…
9
40
4
356
14,015
Quick tutorial on how I made the butterfly specter shader (I made this during lunch so quality may vary) #b3d #shader
15
124
4
1,332
29,340