OpenFest 2026

A Thousand Elephants and Sixteen Gigabytes — vorch: the imps in the box are open source, but we crank the handle
Език: Български

In Moving Pictures the pictures are painted by imps shut inside a box: unimaginably fast, understanding nothing of what they paint, with someone outside cranking the handle. Thirty-six years later, that turns out to be a technical description. vorch is a local orchestrator for AI video. A text idea goes in; out comes a folder of consistent frames and clips, ready for editing. The box of imps is ComfyUI (FLUX.2 Klein, Wan 2.2, ReActor). The handle is a deterministic Python state machine — not an agent framework. And the part of Gaspode, the only one who honestly says what is actually on the screen, is played by Gemma 4 (26B/4B active, multimodal, via llama.cpp): it writes the storyboard and the prompts, then looks at every generated frame and returns a structured JSON verdict — accept, or retry with a corrected prompt.
Dibbler wanted a thousand elephants. The card has 16303 MiB, narration and a render will not fit on it together, and that ceiling is the project's real architect. I will show the working system, the measured numbers, and the three problems no model solved.