---
title: "One Sentence In, Stereo Video Out: 6 Jobs, 90–790 Seconds, an Unfilmed Actress"
date: 2026-08-27
description: "E7: a local H3 video bench — 6 logged jobs, 90 to 790 seconds on one 4090, a director skill with a motion budget, and a heroine with no footage yet."
category: "AI & Compute"
tags:
  - "One Man One Legion"
  - "AI Agents"
  - "Text to Video"
  - "ComfyUI"
  - "Local AI"
  - "Automation"
verifiedHardware: "Workstation · MiniMax H3 on one RTX 4090D via ComfyUI 0.30 · evidence = h3-video skill + 6 job logs + 11 MP4s on disk, h3-guide.md, h3_t2v.py, h3-prompt-director v1.3 (test-log-001, C001.json, r2v_pipeline), MP4 box scans · recounted 2026-08-27 · timestamps Asia/Shanghai"
keyTakeaways:
  - "A locally hosted MiniMax H3 text-to-video bench ran 6 logged jobs between Aug 6 and Aug 14, 2026 — 11 MP4s on disk including 5 deployment-day tests — with render times of 90s (3s clip), 138s (5s, three identical runs), and 790s (10s): doubling the clip length cost 5.7x the render on an offload-bound 4090."
  - "The pipeline delivers itself: launched in background mode, the script posts the finished video back to the chat group (5 of 6 jobs did) while the agent replies with an estimate and exits — an agent that waits synchronously on a 90-second-to-13-minute job times out."
  - "A director skill (v1.3) caps every prompt with a motion budget — 1 primary motion, 2 secondary, 2 environmental, 1 camera move, 1 expression — validated by an 8-cell reference-to-video test matrix whose findings (swap the seed, not the prompt; write the light source in) are codified as rules; the automated result tracker, built to log every run, has zero entries."
measuredOn: 2026-08-27
faq:
  - q: "Does the model really generate audio?"
    a: "The generation graph carries a dedicated audio VAE and an audio-decode node alongside the video VAE, and every shipped MP4 scanned on disk contains an audio track. The workshop's own docs describe that track as stereo; the file-level check confirms the audio is there, not the channel count."
  - q: "Why run video generation locally at all?"
    a: "The deployment notes are blunt: a local model has no content filter, which is the stated purpose of the build. The price is the render ladder — a 24GB card running the model through weight offload, so a 10-second clip costs 13 minutes. The single hard line, written into the skill itself: child sexual abuse material is refused on the spot, and the filtering agent is the filter."
  - q: "So is the character-continuity system in production?"
    a: "No — and the shop's own records say so. The character bible is written, three scenes are staged, and the 8-cell test matrix ran on Aug 10, but the automated generation-history file has never been created: zero recorded runs. The tests were hand-logged. Built is not shipped, a distinction this series keeps relearning."
---
*[Episode 7 of One Man One Legion](/posts/2026-08-27-one-man-one-legion-fleet-audit/) — the fourth stop in the workshops arc, after [the image workshop](/posts/2026-08-27-one-man-legion-ep1-image-workshop/), [the text factory](/posts/2026-08-27-one-man-legion-ep9-text-factory/), and [the quality gate](/posts/2026-08-27-one-man-legion-ep11-quality-gate/). An earlier dispatch measured [InfiniteTalk 14B on this same dual-GPU rig](/posts/2026-08-26-infinitetalk-14b-dual-gpu-measured/); this is the other video bench — the one that takes orders from a chat window.*

A text-to-video model answers a question every owner eventually asks: **what stands between "a model that can generate video" and "a workshop I can order from chat"?** This bench's answer took shape in a single day, then spent a week teaching its owner what the model actually listens to. The deployment guide's first line is the whole ambition, dated Aug 6, 2026: one sentence in, an MP4 out, without ever opening the generator's web interface.

## Getting the template was the first craft

ComfyUI ships new-model templates as packed group nodes that the automation API refuses to accept — and gives no export button. The workaround, recorded in the guide as a reusable trick: load the template once in the UI, queue one run, then pull the executed job's history and grab the workflow JSON the backend expanded on its own. That becomes a 17-node API template — 18 nodes for reference-to-video — with five fields parameterized: prompt, duration, resolution, seed, steps. One quantization trap was mapped the same day: the text encoder ships as int8 weights because the 4090 doesn't support the model's other quantization format.

## Six jobs, three price points

| # | Date | Prompt | Clip | Render | Size |
|---|---|---|---|---|---|
| 1 | Aug 6 | skeleton test: cat on a windowsill | 3s | 90s | 0.2 MB |
| 2 | Aug 6 | two stray dogs fighting at a village entrance | 3s | 90s | 0.5 MB |
| 3 | Aug 7 | boxy Chinese luxury SUV, slow drive | 5s | 138s | 1.3 MB |
| 4 | Aug 7 | same SUV, second take | 5s | 138s | 1.6 MB |
| 5 | Aug 7 | same SUV, commercial phrasing | 5s | 138s | 1.3 MB |
| 6 | Aug 14 | FPV drone chase of a tight formation | 10s | 790s | 2.1 MB |

The ladder is the story: 3 seconds cost 90, 5 seconds cost 138 — three takes of one commercial, timing identical to the second — and 10 seconds cost 790. Doubling the clip from 5 to 10 seconds multiplied render time by 5.7. The card runs the model through weight offload, and the shop's own waiting message says so: minutes to tens of minutes, please be patient. Across all six jobs, 31 seconds of finished video cost 1,384 seconds of GPU time — a 44.6:1 render tax. Total inventory: 11 MP4s on disk, the 6 job outputs plus 5 deployment-day tests.

## The pipeline delivers itself

A video takes 90 seconds to 13 minutes; any agent that waits synchronously times out. So the launch command returns instantly, the agent replies with a time estimate and exits, and the generation script — not the agent — posts the finished video back to the chat group on completion, captioned with its prompt and seed. 5 of the 6 jobs did exactly that, each leaving a delivery receipt in its log; the first skeleton test was not sent. Nobody polls. The video arrives because the script, not a person and not an agent, is holding it.

## The only red line

The deployment notes are unusually blunt about motive: a local model has no content filter, and that is the purpose of building it. Deployment day included four named probes — baseline, horror, war, weapon — and all four rendered. The skill's standing rule draws exactly one hard line: child sexual abuse material is refused on the spot. Everything else is the agent's judgment call, and the agent is the filter.

## The director

By Aug 10 the bench had a second skill: a director that turns one sentence into a storyboard. Its core doctrine is subtraction. Every prompt carries a motion budget — 1 primary motion, at most 2 secondary, 2 environmental, 1 camera move, at most 1 expression change — and the rule of thumb reads: brainstorm ten actions, keep the 3 that matter, delete the other 7. Describe physics, not choreography; write a continuous state, not timestamped events. The bench also keeps a character bible: a 30-year-old Asian woman photographer with five written visual anchors, from a mole near her left eyebrow to a silver film camera that must appear in every frame.

The director's claims weren't taken on faith. An 8-cell test matrix ran the same day at 15 steps: 3 clean passes, 2 failures, 2 retries markedly improved, 1 cell left unmarked. The findings are now codified rules — a local eye defect is cured by changing the seed, not the prompt; a dark scene needs the light source written in; original photos preserve identity better than generated intermediate images. The next day a second model was put through the same discipline, pitfall table and all.

## The honest column

The pipeline ships with a result tracker that records every run's seed, parameters, quality notes, and identity drift — the feedback loop this series keeps praising in other workshops. Its history file has never been created: zero entries. The actress is cast, her wardrobe locked, three scenes staged. She has not filmed a shipped frame. Built is not shipped — and this time the shop's own empty log file is the evidence.

*[Back to the fleet audit.](/posts/2026-08-27-one-man-one-legion-fleet-audit/)*