Cartographe des enfers💽 Tier 0 ⚜️FDL Saint Empire Numérique 🧙♂️🧙🧙♀️ZKWizard (∩` ﾛ ´)⊃━炎炎炎炎炎 ♠️Cryptomancer.xdc♠️ 🐉DragonSlayer.xdc🐲 🌐local studio.depin
Terre
Joined November 2021
- Tweets20.2K
- Following429
- Followers3.7K
- Likes98.6K
FocalZero retweeted
Un laboratorio chino acaba de humillar a media industria del video.
Subes una foto y un audio, sale un avatar hablando en sincro. Open source y gratis.
Lo que antes era agencia, cámara y edición ahora es un repo.
Se llama LongCat-Avatar.
FocalZero retweeted
Si tienes Qwen3.5-4B corriendo local, hay una forma de cambiarle el comportamiento con solo dos neuronas.
En la capa 14, neurona 8949 sube +1.
En la capa 13, neurona 7956 baja -1.
Eso solo, en el down_proj, y el modelo baja bastante las negativas y empieza a responder más “sin filtro”. No es fine-tuning ni abliteration completa. Es steering directo sobre dos MLP neurons.
La dosis se escala así:
dose × mediana(‖residual‖ saliendo de la capa L, tokens 1:) / ‖W_down[:, n]‖
Hay gente que ya lo probó y reporta que baja el refusal rate fuerte (de ~90% a ~12% en algunos tests), aunque también pierde factualidad y a veces se pone repetitivo.
Si ya tienes el 4B en local (Ollama, llama.cpp, vLLM o lo que uses), prueba esos dos números y cuéntame qué sale. Es de lo más raro que he visto en mechinterp últimamente.
El autor tiene un explorador para probarlo en vivo, te dejo el link en comentarios
FocalZero retweeted
OKAY, THIS IS BIG. HERMES DESKTOP IN YOUR BROWSER IS NOW OFFICIAL. 👀
A few weeks ago, Bear put the real Hermes Desktop in a web browser.
Now it's merged into Hermes itself.
The full Desktop workspace, served straight from your Hermes gateway. No separate server. No relay.
What it does:
▸ Chat, files, terminal and profiles, right in a browser tab
▸ Same conversations as your CLI, Desktop app and messaging apps
▸ Touch-friendly layouts for phones and tablets
▸ One command to open it: hermes webapp
▸ Settings → Web Access serves it on your network or your own URL, behind Nous sign-in
▸ hermes webapp --sign-out signs every browser out at once
And here's what makes this so interesting...
Any device with a browser can now be your Hermes Desktop. Your laptop, a work machine, even your phone.
Community build to official feature. That's how Hermes keeps growing.
Repo + setup in the first reply. 🔥
Huge credit to Bear @BearHuddleston
FocalZero retweeted
More reference control. Fewer denoising steps. More places to run H3.
The latest community contributions span 8-step reference-conditioned video + audio, 4-step distillation, sparse attention, and reference-token compression - alongside serving and workflow integrations across GPU servers, TPUs, and Apple Silicon.
What matters is what this enables: more affordable experimentation, finer creative control, and more choices about where and how video generation runs. Open models become more useful when the community can improve the methods, adapt the runtime, and build workflows around them.
Credit to @haoailab @yisun0618 @lanfeng491270 @guocheng_qian @Alissonerdx @ddalcu @tazmannner @HaochengXiUCB @huggingface @Jhuaplin, and the PTD, DMAD, SparkH3, TorchTPU and FlashML teams - including Zhengming Yu and zechengtang on GitHub.
11 additions to the H3 ecosystem index. Thank you to everyone building, testing, and sharing.
github.com/MiniMax-AI/awesom…
Open weights. Shared progress. MiniMax H3 is moving fast.
We built H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on.
Recent highlights:
• FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon.
• Sol-H3 — NVIDIA’s SANA team: 15 seconds of 768p video + audio in 6.6 seconds on 8×B300, in the team’s warm-inference benchmark.*
• VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released.
• PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI.
• LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio.
Behind every release are people training, optimizing, quantizing, testing and sharing. Special thanks to:
@haoailab @nuvalab @NVIDIAAI @xieenze_jr @HaochengXiUCB @ArashVahdat @julberner @LightX2V @ComfyUI
And to the individual contributors pushing the work forward:
@haozhangml @cxlcl1 @lawrence_cjs @yitongli165665 @haopengl33 @songhan_mit @shanasaimoe
Thank you for building with H3 and helping make it faster, more accessible, and more useful for the community. Powerful models go further when we build together. Keep pushing H3. Excited to see what comes next. 🚀
Explore the ecosystem: github.com/MiniMax-AI/awesom…
FocalZero retweeted
AIアニメ制作を始めてからの軌跡を、2分の動画にまとめてみました✨
ChatGPTとの対話だけで、素材生成から編集まで完結させています。
制作を始めた当初の悩み、今感じている迷い、そしてこれからの抱負を素直に語りました。
Topview ( @TopviewAIJP )などのプラグインを使って、画像・動画・音楽を生成し、結合やテロップ入れもChatGPT側で行いました。
すべてクラウド上で作業しているので、ローカルにソフトを入れたりファイルをダウンロードしたりする必要はありませんでした。
FocalZero retweeted
動画生成モデルMiniMax-H3の新規マージモデルが公開。
独自手法PRISMによりRef2VA微調整版2種を統合。SVDによる構造ノイズ分離や測地線補間を用い、8ステップ高速版の特性を段階的に融合。
#MiniMaxH3 #Ref2VA
URLはリプ⬇️
FocalZero retweeted
I can't believe this is free. 🦊
Hermes just shipped its own browser that doesn't look like a bot.
It's called Camofox, an official Nous Research plugin, and it runs on your own machine.
▸ Fingerprint spoofing built in, on a Firefox fork
▸ Shows up as everyday Firefox, not a headless bot
▸ Its own browser in the background, so yours stays free
▸ Search, click, type, scroll and read pages
▸ Keeps your logins, one profile per Hermes profile
▸ No cloud browser, it all runs locally
▸ Already using Camofox? Nothing changes
Your machine. Your logins. Hermes in disguise.
FocalZero retweeted
好消息：ComfyUI 也有 Comfy Agent啦！
ComfyUI 大家不陌生吧？
一个免费开源的 AI 图像、视频、音频 生成工具。
节点图的方式让你把 AI 生图流程可视化地搭起来。
本月ComfyUI 团队又发布了 Comfy Agent,
一个专门为 ComfyUI 打造的 Agent。
以后在你的codex里也可以更好的自由创作啦
丝滑，nice！
😲 都马上2027年了，还有人没学过 ComfyUI 吗？
那还有必要学吗？
ComfyUI 开源项目目前都 142K个 Star ⭐️。
那它能干什么呢？
ComfyUI 是一个免费开源的 AI 图像、视频、音频 生成工具。
它用「节点图」的方式让你把 AI 生图流程可视化地搭起来。
节点图的每一步都高度可调，非常灵活，整个工作流可以保存为 JSON，可脚本化。
加上本月ComfyUI 团队又发布了 Comfy Agent,一个专门为 ComfyUI 打造的 Agent。
更是将搭建工作流的难度爆降，因为你只需用自然语言描述需求，Comfy Agent 就能帮你搭建、迭代工作流，并直接运行生成作品。
❓最后谁不需要学？
如果你只是偶尔生成几张封面图，又或者搞几张PPT，发发自媒体，那么还是直接使用GPT-Image 2.5吧
同样的视频也是也直接用 Seedance 就完事了。
FocalZero retweeted
Meet Oh-My-Hermes
I burned 200B tokens tuning my Hermes Agent setup, and now you can get all of it with a one-line install.
It comes with 150+ skills, a long-term memory system that lasts twice as long, graph engineering, and memory-management skills. It also includes subagent routing across model chains, plus optimizations like parallel tool calling and evals.
You don't need to do much, though. Just install it and give it a task. Everything is optimized to route automatically. Hermes Agent is excellent, and I love this technology.
I put a lot into making it really good, then open-sourced it so you can enjoy it right away. Meet oh-my-hermes.
GitHub: github.com/rlaope/oh-my-herm…
#hermes
FocalZero retweeted
ComfyUIで動画生成モデルMiniMax H3をタイムライン編集できる拡張機能が登場。
映像や音声をトラック配置し、ショットごとのプロンプト指定による絵コンテ作成が可能。生成プレビューやリテイク、音声同期機能も搭載。
#MiniMaxH3 #ComfyUI
URLはリプ⬇️
FocalZero retweeted
建议所有用 AI 写代码的人，在项目提示词里加一条规则：
“遇到不熟悉的功能，不要立即从零实现。先检索 GitHub 上维护活跃、许可证明确的同类开源项目，比较实现思路、依赖和已知问题。优先借鉴经过验证的设计，不要直接复制不兼容或来源不明的代码。”
这条规则帮我省了不少反复返工的时间。
Vibe Coding 不只是让 AI 会写代码。
还得让它知道什么时候应该先查资料。
FocalZero retweeted
太猛了，开源MOSS-Image在ComfyUI的下载量已经100W+了
看大家都在玩MOSS-Image图片模型，我也拿10个热门提示词实测了一把
3个免费玩的渠道、实测提示词和完整图片放在评论区，图片效果大家自己看
这次测试，尝试了好几种不同的玩法：
空姐、演唱会大屏、古建筑示意图、手相分析、直播、Cosplay、真人摄影、三视图、私房照、信息图等等
说实话，5B参数的开源模型能做到这样已经很不错了，已经可以和部分闭源模型媲美了
而且模型支持结构化的JSON Layout可以指定文字和视觉元素的位置，这个可以做图片内容的精准控制了
速速玩起来👇🏻👇🏻👇🏻
🤝 Paid partnership
FocalZero retweeted
Las clases de electrónica acaban de perder contra una app gratis.
Aprendes Arduino desde el navegador: montas el circuito sobre una placa en 3D, escribes el código, le das a ejecutar y lo ves funcionar. Sin comprar un solo componente.
16 lecciones, del primer LED a tu propio piano electrónico. En 12 idiomas.
Y no la hizo una escuela. La hizo una sola persona con Claude Opus 5.5.
