The full build behind EP39: three install commands, a reference generated from the installed packages so the agent stops guessing, motion tokens, word-synced components, cams and graphics trading places on sentence beats, which local model writes the graphics, and the six things that went wrong on the way.
How our graphics know the exact frame a word is spoken. A free, local motion-graphics kit built on Remotion, driven by Whisper word timestamps, written by an AI agent from a plain text kit. What we installed, what we built, the model that wrote the props in half a second, and the three things that went wrong.
A published episode scored 86 failures against its own script. Here is every class of mistake we found making local AI video with LTX 2.5, SageAttention, Krea 2 and ComfyUI, what actually caught each one, and the checks we now run before anything ships.
Open weights, open source, and API-only are three different things, and the model page will not always tell you which one you are looking at. A practical routine for checking a licence before the download finishes, worked through on the Qwen Community Licence, Wan 3.0, and Apache 2.0.
How we made our AI presenter walk between sets in one continuous shot on a single RTX 5090: the walk-set plate design, guide strength 0.7 to 0.5, the 47% midpoint problem, generate-long-cut-short, chroma key at 0.28, and the honest numbers including the advice of ours that testing killed.
Three open-weight models, one RTX 5090, and the same jobs: a real bug, playable games, tool calling, long documents. Measured numbers, two defaults that were quietly breaking the results, and the finding we published wrong and had to correct.
We trained a character LoRA on Krea 2 and benched it against our production image stack. What ports, what breaks, the prompting rules that differ from Z-Image, and the LoRA lessons from 230 test renders.
For five weeks we would not say how our presenter’s talking footage is made. Here is the whole chain: LongCat for the body, MuseTalk for the mouth, Duix to clean up what that leaves, and MuseTalk again for the visemes. Why the order matters more than any single model.
We benched LTX-2.5 for two full rounds and got confident, wrong answers, because the official templates ship an LLM prompt rewriter that is on by default and buried inside a subgraph. Here is how to find it, how to actually turn it off, and what the model does once your prompts reach it intact.
The bridge nobody documents: connect a local Ollama brain to a real-time 3D VRM face with voice, lip-sync and expressions. The four services, the ports they talk on, and the two traps that cost us days, the thinking-mode stall and the open-mic mistake.