#开源
6 tutorials, projects and news items tagged 开源.
The AI news site AIHOT has open-sourced its engine under the MIT license. This covers what the stack includes, what it leaves out, the costs to watch, and a prompt that has your AI agent adapt it to your own industry.
Intermediate tutorial: install the open-source PhoneGo, turn on Wireless debugging and Accessibility, add your own model API key, run a login-free task first, then have the agent move information between Xiaohongshu (RedNote) and Ctrip. Covers why it beats screenshot-only phone agents and where it stops before payment. Includes 2 copy-ready prompts.
The open-source ComfyUI plugin ComfyUI-H3-FaceRefine fixes faces that fall apart in the distance in MiniMax H3 video. It tracks the face on every frame, crops and enlarges it, has H3 redraw it at low strength, then pastes only the face back. It can follow one person in a crowd, split by shot, and keep lip sync. Covers how it works, the plugins and models you need, running it on 12 GB of VRAM, and common pitfalls, plus a prompt for your AI agent to install it and run a test.
Several open-source projects on GitHub are built for making videos with Meta Muse, and each one covers a different stage. muse-video-skill turns an idea into a script, storyboard, and call instructions, muse-video-prompt-skill handles prompts, stitching, and frame checks for Muse's built-in video, muse-video-prompt asks questions to build a prompt, and muse-video-prompts is a library of 204 public prompts. The article compares their licenses and stated limits and includes a prompt you can give your AI agent to help you pick one.
The OpenDesign team open-sourced OpenMotion, an Apache-2.0 alternative to CapCut and Claude Motion. Your AI agent breaks a reference video into shots, rebuilds it in HTML with your brand, copy, and assets, and you fine-tune and export an MP4 in the built-in editor, with support for more than 20 agents including Codex and Claude Code.
Microsoft's TRELLIS.2 and Tencent's Hunyuan3D-2.1 can both make textured 3D models from one image, but they are heavy. TRELLIS.2 needs a 24GB NVIDIA card and was only tested on Linux, and Hunyuan3D-2.1 needs 29GB for shape plus texture.
