The Engine Room, ComfyUI

Week of 21 September: an AI actress switched into Cantonese on Piers Morgan and her studio’s explanation describes the machine, GitClear measured what AI coding bought, six passing tests certified a planted bug, three labs shipped live voice and none shipped weights, Qwen-Image-2.1 changed its licence, 85.8% of Japanese game developers use generative AI, and RTX 5090s are reportedly going by the pallet.

A 31-second photoreal film of a consistent AI presenter, made on one RTX 5090 with no cuts: reference pictures given jobs, a locked voice track, three cheap seed previews then one upscaled finish, each clip extending the last. The five ways it broke, the picture that fixed each one, the plate rules, the LoRA retrain, and the bill.

Qwen released a new image model today at 7.1 billion parameters, roughly a third the size of everything else in the line, with ComfyUI support the same afternoon. Every earlier Qwen image release was Apache 2.0. This one is research and evaluation only, and we read the licence file rather than the thread about it.

Somebody asked today how you count the real cost of AI video once the failed generations are included. We published a 16 minute episode this morning, we keep every take we render, so here are our actual numbers: what we made, what we used, what the retries cost, and the line item nobody prices.

Turn a photograph into an anime character, put that character in six worlds, and make a short film with no dialogue. All of it local, in ComfyUI, on one gaming card. Five things went wrong on the way and all five had the same fix.

YuE2 landed last week, and what happened next is the best argument for open weights we have seen in a while. Quantised to 8 GB by day two. The missing encoder trained from scratch by a stranger on day five. A LoRA trainer the same day. One thing almost nobody is mentioning, though, and it matters if you plan to earn anything.

Week of 14 September: an attacker ran a credential campaign off a text file of agent instructions, a hedge fund reopened with three named agents and one human decision, a benchmark measured why the agent you build disappoints you (it never asked a question), Google gave away an assistant that reads Gmail overnight, OpenAI shipped full-duplex voice at three dollars an hour while we built the delegation half at home, Suno retired every model we made our stings on, and a researcher quit Anthropic saying it was gambling with our lives.

One group photograph, three people, and we wanted two of them gone without the third one changing. Five free image editors running on our own card took the job in turn. The first thing that happened was the room came back empty. What fixed it was grammar, and the rest of what we learned is below, with the five workflow files.

Every identity benchmark we can find scores the face. We put one full-length figure through five local image editors with an explicit instruction to keep her proportions, scaled the results so the heads match, and measured. Every engine kept her face. Four of five changed her body. Here is the method, the tool, and what the control test says about who this happens to.

We took a stock Asus TUF gaming laptop with an RTX 3070 and 8 GB of video memory, and benched eight local models overnight across ten tasks. The result that surprised us: a 30 billion parameter model with two thirds of itself off the card beat a 14 billion parameter model that fit better, by nearly four times. File size does not predict speed. All 103 measured rows are free below, failures included.