Original Essay Source
- Original Author: JK Shin (lablup CEO)
- Article: 《After Astra: What Loop Transformers Are Changing》 Facebook.
(September 4, 2026, OpenAI GPT-6 Astra Announcement and Hardware Ecosystem Analysis)
I obtained permission from the original author. Please note that the thumbnail and introduction are not related to the original author.
I wanted to try a new production method after constantly making music videos. I also explored the purity of text interpreted by AI and how it is expressed.
There's a full 12-minute version, but even magical girls can get bored! So, I summarized only the interesting parts. However, I didn't choose this myself. I threw in the whole text and asked it to compose the scenes on its own. Although I didn't interfere with the directing at all, "magical girl" as a keyword resulted in concentrated actions from the model.
I extracted 33 scenes and this time added subtitles to the original without any modifications, added background music, and used the Nvidia upscaler. All are open-source projects.
Astra seems to have a significant impact (..) on video generation, but since it's cheaper than token costs...it might be okay for a while.
Control became available from AE 5.5, and there was a lot of interest in Blender -> Hicksfield or Mimic workflows. This method has already been released and used, but it's great that someone else can do it for those who don't know how to use Blender (3D outsider). It's like making an app with Rust even though I barely know any coding languages...
Additional note: Since I introduced Minimax h3 several times in this project, let me introduce YuE2 3B. It's a great open-source tool for creating warm music. Someone who used Suono said that the 4.0 quality comes out. For someone like me who has been using open-source music generation, it's a project comparable to Minimax h3.
I will share a summary this weekend if no one else introduces it.
