Day 04: When AI video tools refuse violent scenes, try open models on your own machine

I asked an AI to film a man being hanged. Runway refused twice. A street shot vanished three times without a word.

So we took the murder home to my Mac. One open model got the hanging right. The other rendered for 35 minutes, then the spy climbed down from the beam and strolled over to the chessboard.

By bedtime my hero had come out Hulk-sized in four shots, I'd judged 22 clips and approved zero, and Claude had broken my first Premiere project by overwriting files I had open. Worst day on the film so far, and tomorrow an AI starts grading my script.

Day 04 at a glance

Hosted AI video services would not render the violent shots my R-rated film needs, so we tested open models on my Mac. MiniMax H3 made 3 seconds in about 10 minutes and got the story beat right. Wan 2.2 took 35 minutes for 5 seconds and got it badly wrong.

I set a $100 cap on rented GPUs and spent $0. The film went into Premiere for the first time, and promptly broke.

Why won't Runway render the violent scenes?

DEATH MATCH is meant to be R-rated (the whole project lives on the Making Death Match series page). The Penthouse scene ends with Victor's men killing a captured spy. Later there's a tournament where people die, and a love scene between John and Lucy.

Runway's filter refused the spy shots over and over. Shot 08-10 was rejected twice. So was 08-18.

A street shot in Scene 6 vanished three times with no error at all. We later worked out those were silent rejections. None of the big hosted video services will show a hanging, and sex is out entirely.

So the question for the day was simple. Could we make those shots somewhere else and keep the look?

I set the ground rules first. Claude could rent cloud GPUs up to $100, logging every pod. But we'd test on my own machine first, with the models on the Studio drive. It had 1.9 TB free. My Mac had 76 GB.

The film is R, not NC-17. The violence should be brutal and specific. Intimate scenes can go about as far as Basic Instinct went, nothing explicit. Claude drafted an R-rating pass listing every violent beat.

How did MiniMax H3 do on my Mac?

H3 is an open-weights video model that was already installed. Claude rendered shot 08-18, the hanging spy's shadow on the wall, from the uncensored prompt.

Three seconds took about 10 minutes and peaked at 44 GB of memory.

The beat was right: the light flares, the silhouette kicks, the light drops back to orange. The look was softer and flatter than our Runway shots, closer to a pastel wash.

Why did Wan 2.2 ignore the noose?

Claude downloaded about 45 GB of Wan 2.2 image-to-video files to the Studio drive and drove them through ComfyUI. The first 5-second segment took 35 minutes.

Then the spy, who was supposed to be hanging from a beam, stood up and strolled over to the chessboard. I waited 35 minutes to watch Wan ignore the noose entirely. Victor's men are going to need better rope.

Claude's verdict was blunt: Wan follows the picture more than the words. H3 follows the words. So it re-rendered 08-10 with H3, and the beat landed: the spy hangs behind Victor, Boris punches him, and Victor keeps playing chess.

By evening H3 had an overnight queue of four more Penthouse death shots, about 25 minutes each. The queue failed instantly the first time.

There were two reasons. Three frames had never been through the outline-removal pass, so the files didn't exist. And H3 has to be launched from its own folder to find its shader file. Both fixes took two minutes.

Can a local model handle the love scene?

Midjourney won't draw nudity. Qwen Image Edit, also on my Mac, will, and its three frames for the 07-23 love scene were discreet and R level. But Qwen draws the ink outlines I don't want, and with John's character lock fed in it got worse. In Claude's words: "John came out as a generic heavyset man and the look went flat and cartoony." Those frames stay off this blog.

How did Claude break my first Premiere project?

I asked why there were no Premiere projects yet. Claude installed a small Premiere panel it called DM Bridge, which runs script commands dropped into a folder.

Through it, Claude built the Death Match project. There are nine scenes, each on its own sequence, with every storyboard panel in script order. Voice, effects and music sit on their own tracks, and gray "NEEDS PICTURE" slates fill every gap.

It refreshed 440 media items and leveled 341 audio clips.

I opened the project, feeling pretty pleased, and saw "lots of errors".

Claude had overwritten media files while Premiere had them open. It's like swapping the reels while the projector runs.

New rule: never overwrite a file Premiere is using. Add a new versioned file instead. If you've ever had to re-encode a video trim, you know editors don't like files changing under them.

Why did I approve zero of 22 clips?

Runway ran all day on two slots and kept making the same mistakes. John came out "hulk"-sized in four different shots, because some approved Midjourney frames of him were already too broad and Seedance made them broader. It's the prop lock lesson from Day 03 again: fix the input.

Lucy gave the camera a wide, fixed stare that belongs in a horror film. And the script feeding Runway counted failed jobs as running, so slots sat idle waiting on jobs that were already dead. I judged 22 clips and approved none. Midjourney, whose block from Day 02 was meant to end at 10:40, couldn't help: Claude's test browser failed Cloudflare's human check, and so did I.

What comes next?

Open models will shoot what the big services won't, but they don't look like my film yet. H3 got the story right and Wan kept the picture closer. Neither matched Midjourney's look, and the rest of the week kept coming back to that gap.

Tomorrow I watch the first cut, and it's a mess. The lip sync is wrong. And Jev, an AI, starts grading my script. After a week of me judging the machines, one of them gets to judge me.

FAQ

Will hosted AI video tools render violent scenes?
Not the ones we used. Runway refused our hanging shots repeatedly, sometimes silently, and none of the big hosted services will show a hanging.

Can open video models run on a Mac?
Yes. MiniMax H3 rendered 3 seconds in about 10 minutes on my Mac, peaking at 44 GB of memory. Wan 2.2 took 20 to 35 minutes per 5 seconds.

Which open model followed the prompt better, H3 or Wan 2.2?
H3 followed the words and got the story beat right. Wan 2.2 followed the input picture more, and in our test the spy left the beam and walked to the chessboard.

How much did Day 04 cost in rented GPUs?
Nothing. I set a $100 cap with a spend log, but everything that day ran locally.

Next: Day 05

While you wait for Day 05: play Halfsies

Halfsies is our iPhone game, live now on the App Store. "One cut. Two halves."

Three objects a day, one attempt at each: cut each one exactly in half by eye and the scale tells you how close you got. Nobody gets hanged, no filter refuses you, and you can't blame the AI when you miss. More on the Halfsies page. Download Halfsies on the App Store