The Sessions

The project was built in long working sessions with an AI collaborator driving a local render pipeline. This is the honest shape of them.

Before: 2023–2025 — the obsession

March 2023: the source video runs as the only ad on the platform for a month, from an account that never posts again. Three years of posts, DMs and two podcast appearances turn it into a private mythology. 57 hand-made prototype stills. A spec forms in DMs: 38 segments, one style each, "as many differing styles as possible." "I am trying to execute."

Session 1 — the founding (July 2026)

Source re-acquired and inventoried shot by shot. ComfyUI stood up on a 6GB laptop card. The architecture invented: per-shot rendering, style templates, canny edge-lock. Three complete cuts rendered (v1 "the 2023 print", a fade coda, a combustion coda). The doctrines written down — including the load-bearing one: the accidents outperform the intentions.

Session 2 — the restoration (July 2026)

The lineart conduit solved the pen-line styles. Two video models taste-tested; a style LoRA trained on the 57 prototypes in fifteen minutes. A parts librarian built: every render carries its full recipe in its metadata, films assemble from JSON in seconds. Then the discovery that a $0.04- per-frame Chinese edit model preserves lettering through any style — and the whole film re-rendered through it for about $25. Cubism, "categorically impossible" under edge-lock, fell for $0.92.

Session 3 — the truth pass (July 2026)

Scene detection against the source found the shot table's own errors: a misidentified IPO (Toast, not Olo), phantom cuts, a sponsor hiding behind the Fearless Girl statue, and an off-by-one label slide that had silently mislabeled the film's hardest zone. A frame-exact inventory rebuilt — 65 true shots, 961 frames, zero overlap. The bells mapped. The maximal version assembled at 2048², exactly 961 frames, ending on true black — with the film's final image an accidental question mark of fire. Total cloud spend, all sessions: about $29.

The instruments

• SD1.5 + AnimateDiff + ControlNet — the 2023 print; its flicker is period
• a lineart conduit — for styles whose soul is a pen line
• a LoRA trained on the author's own prototypes — the house voice
• LTX-Video — dissolution textures; its drift became the coda's rising shadows
• CogVideoX — proved the successor's engines have the ceremony memorized
• Seedream 4.5 — the restoration engine; text survives the transform
• ffmpeg, scene detection, a bell-band onset map, and a librarian