# White House Funk

A parody remake of a whole 2-minute music video, done end to end by codex through this MCP. Eight people
are recast as eight US political figures. Each replacement is the whole person: face, hair, build and
outfit. Everyone else, the motion, the camera and the soundtrack come from the source.

| file | what it is |
|---|---|
| `uptown_funk.mp4` | the source, 121.8 s, 2923 frames at 24 fps |
| `task.txt` | the request codex received, word for word. `@SRC@` stands for the source path |
| `chars/` | the character library: one sheet and one photo per character, plus `lib_whf.json` (descriptions and outfits) |
| `run.sh` | copies `chars/` into `$REMAKE_HOME/chars/`, then runs `scripts/codex_exec.sh` with the task |
| `white-house-funk.mp4` | the result, 1280×720, with the source soundtrack |
| `result.json` | what codex saved: the cast, a grade and a note per window, the seeds it picked, and the remaining flaws |

## Run it

Start render workers on the same `REMAKE_HOME`, then run the script from a source checkout (the `scripts/`
directory is not installed by pip):

```bash
viggle-remake-worker &            # one per GPU; see the main README
examples/white-house-funk/run.sh
```

## What the run above took

The run used codex `gpt-6.1-terra` at reasoning effort `high`, in pool image mode, with two render workers.
Each worker needs a GPU with about 94 GiB free at bf16 (see [Hardware](../../README.md#hardware)); on a smaller GPU
you would offload or quantize the model yourself.

- **Time:** 86 min of wall-clock time for 22 windows. Most of it was spent waiting on renders.
- **Image calls:** 81 of the 160 the task allowed. Of those, 13 failed: image-safety refusals on some of the political figures, and group pictures that came back at the wrong aspect ratio. The agent repainted around them.
- **Tokens:** 250 K.
- **API spend:** $0. Pictures come from the ChatGPT subscription's image allowance.
- **Seeds:** every window rendered at seed 0. Window 17 was re-rendered at seed 1, and the seed-1 take was kept. Window 18 got a pinned frame to fix a cut that came too early.

codex graded the result `BAD` overall, because of one window:

- 12.6–13.1 s: a passing woman's fur coat drops out.
- 14.9–16.1 s: the strobe timing drifts.

It noted smaller flaws elsewhere: a white sleeve at about 47 s, gold necklaces left on some suits at 64–68 s,
background figure counts that drift, and some likeness drift in the fastest cuts. A second run will
differ in detail, since codex paints fresh keyframes each time.