15s of MiniMax-H3 video + audio, generated in ~72s on one RTX 5090. Our baseline took ~26 minutes. We wrote up how LightX2V got it down to 72s 👇 768×1344, steady-state benchmark. MP4 export and network transfer excluded.

Your RTX 5090 has a new side quest 👀 Real-time AI video generation with EdenLive 5B, powered by LightX2V. We measured ~24.7 inference FPS at 1280×704 on one card after warmup. Grab the code from the original post, give it a spin, and show us what you build!

LightX2V powers real-time EdenLive on a single RTX 5090 ⚡ Our team built the acceleration that makes this setup possible. 🥳 Proud to play a key role in bringing EdenLive to the community, and big thanks to the VAST team for open-sourcing it! Got an RTX 5090? Grab the code and try EdenLive locally with LightX2V. We’d love to hear how it runs for you! 🙌 👉 Try it with LightX2V: github.com/ModelTC/LightX2V/… 🌱 Explore EdenLive: github.com/VAST-AI-Research/…

“Stay With Me~” hits different under neon lights 🌆🎶 Made with the LightX2V Studio API, combining our singing avatar model with MiniMax-H3 for the story scenes. We handed the script, shot planning, transitions and API calls over to Codex. Turns out making an AI music video can involve a lot less clicking!!! #AIMusicVideo #AIAvatar #LightX2V

⚡ Meet LightX2V APP — an open-source macOS app for running Qwen-Image-2.1-Turbo locally, with optional Prompt Enhancement. 🛠️ Powered by LightX2V for image generation and llama.cpp for Prompt Enhancement. 🎨 Create images from text, refine them with references, and keep your creations on your Mac. Tested on an M4 Pro with 24GB unified memory. 💻 Download the app, explore the code, and make it your own: github.com/ModelTC/LightX2V-… #LightX2V #llamacpp #QwenImage #LocalAI #OpenSource #macOS

Run Qwen-Image-2.1-Turbo locally on your Mac 💻 Here’s what we measured on an Apple M4 Pro with 24GB unified memory, using 8 inference steps: • 512 resolution: 38s total, 3.3s per denoising step • 1K: 118s total, 13.1s per step • 2K: 622s total, 72.7s per step These timings exclude prompt enhancement. In our tests, prompt quality made a big difference, so we recommend keeping enhancement enabled. It adds about a minute via llama.cpp on this setup. Want to try local AI image generation? LightX2V APP runs on Apple Silicon, with Qwen-Image-2.1-Turbo and built-in prompt enhancement: 💫💫Get the Mac app： github.com/ModelTC/LightX2V-… Building your own workflow? The LightX2V inference framework supports Qwen-Image-2.1 on NVIDIA, AMD, Apple Silicon, and other chips: 💫💫Explore LightX2V：github.com/ModelTC/LightX2V Both repos are open source. 🔥🔥 Give them a try and tell us how they run on your setup. Feedback and contributions are always welcome 🙌

Run Qwen-Image-2.1-Turbo locally on your Mac 🍎 LightX2V now supports AI image generation on Apple Silicon via MPS — merged into main. Running it on your Mac? Share your chip, memory, and results. We’d love to hear how it goes. #LightX2V #QwenImage github.com/ModelTC/LightX2V/…

Try it online now! 🎬 Create with MiniMax H3, Wan 2.2, SekoTalk, Qwen-Image, SwiftVR, and more in LightX2V Studio. No installation needed. Come see what you can make! 👉 x2v.light-ai.top/en

Github: github.com/ModelTC/LightX2V

Scaling the optimized setup to 8× RTX 5090 brought generation time down to 12.52s for the same clip. The recipe: 29 → 4 steps, sparse attention with matching LoRAs, FP8, kernel and VAE optimizations, plus multi-GPU parallelism. The blog breaks down each stage.