ToolScore
Video Gen 5 min read

Our Video Pipeline Runs Local — Here's Why

≈$0.03/hour per take the number that matters here

By Stan K.

We publish video through a closed local pipeline — script, voice, visuals, music, edit, subtitles, thumbnail — on consumer GPUs, with zero video-gen API bills. The honest reasons it wins (consistency, iteration cost, no 5-second ceiling) and the honest reason it can't do everything yet: local video models still don't fit a 3060.

We run a production video pipeline end to end: an agent writes the script, local TTS voices it, image generation renders the visuals, music and assembly and subtitles and the thumbnail close the loop. The whole chain runs on consumer GPUs in the same box that renders our thumbnails. The video-generation API bill for all of this is $0 — not because we are dogmatic about local, but because the economics and the output both came out better. Here is the honest breakdown, including what local cannot do for us yet.

What actually runs on the box

StageWhat runsMarginal cost
ScriptAn agent on a flat-fee coding planincluded in the plan
VoiceLocal TTS (Kokoro-class)electricity
VisualsFLUX / Qwen-Image on a 3060 12GB, ComfyUI workflowselectricity
MusicA local library, no streaming licensesone-time
Edit, subtitles, thumbnailffmpeg chain + in-pipeline image genelectricity

A 170 W card draws ≈$0.03/hour at the US average rate. A rejected take — wrong composition, wrong palette, sixth attempt at the same shot — costs a re-roll, not a credit. Compare that with the paid side of the market in The Real Cost of a Usable 5-Second Clip, where a $0.35 sticker becomes $2–4 per usable clip once the keeper rate is counted.

Why local wins for what we make

Three reasons, in order of how much they actually matter:

  1. Controllable output. A video product needs the same aesthetic across an episode — same palette, same character weight, same composition language, 40 shots in a row. Video-gen models give you a fresh roll of the dice per clip; a locked image-gen workflow (fixed LoRA, fixed prompt skeleton, fixed seeds where it matters) gives you a look. Consistency is the product.
  2. Zero marginal cost per iteration. Iteration is where video dies on paid APIs: every tweak is a new clip at full price, and the 5–10-takes-per-keeper reality multiplies it. Locally, iteration is our default mode — the tenth variant of a shot costs the same $0.03 of electricity as the first.
  3. No “just a 5-second clip” ceiling. Video gen sells seconds. Our episodes run minutes. Composing longer sequences from stills, motion passes and ffmpeg animation has no unit length at all — the render takes as long as the render takes.

None of that is “local is free” — the hardware is a real up-front cost, amortized over everything else the box does. That math is its own article: One GPU Beats Ten Subscriptions, and the hardware side lives in the Local GPU AI leaderboard.

Why we still don’t run local video models

Honesty section, because this is where local-video hype needs a cold shower. The open video models are real — but none of them fits our 12GB cards at official specs:

ModelOfficial requirement12GB reality
Wan 2.2 T2V/I2V-A14B80 GBcommunity GGUF Q4–Q5 (~9–11 GB files) — no confirmed 3060 report
Wan 2.2 TI2V-5B24 GB (a 4090)not officially supported below that
HunyuanVideo 1.514 GB minimum with offloadclose, but above 12 GB
Mochi 1~60 GB (under 20 GB in ComfyUI)still above a 3060

So for the shots that genuinely need generated motion — specific physical movement we can’t compose from stills — local is not yet our answer. When a shot truly calls for it, we treat it as a per-project purchase on a PAYG API (FLUX 3 or Gemini Omni, no subscription), budgeted at the 5–10× rejection rate. What we don’t do is keep a monthly video-gen plan for a handful of shots a month — that’s the exact trap list in Video Gen Fine-Print Traps.

The rule we landed on

If a shot can be composed — a generated still with camera motion, layered scenes, a screencast, a slide build — compose it locally at electricity prices. If it truly needs generated motion, buy it per project on pay-as-you-go, priced at 5–10 takes per keeper. What you should never do is pay a monthly subscription for shots you can compose yourself.

The takeaway

Video gen APIs sell motion by the second; most of what a published video needs is composition, consistency and iteration — three things a local pipeline is structurally better at. If your shot list is 90% composable, start with You Probably Don’t Need a Video Gen API. If it isn’t, price the difference honestly in The Real Cost of a Usable 5-Second Clip before you commit to a plan. The full market board — paid and local — is the Video generation hub.

Numbers change — the math doesn't
Token prices and plan limits shift monthly; the structural arguments in this guide age slowly. For current per-tool numbers see the live rankings.