ToolScore
Video Gen 5 min read

You Probably Don't Need a Video Gen API

90% of shots compose the number that matters here

By Stan K.

Screencasts, slide builds, thumbnail compositions and ffmpeg animation over generated stills cover 90% of published video — at pennies per frame instead of $2–4 per usable 5-second clip. When you genuinely need generated motion, and the local models with clean licenses (Wan 2.2, LTX-2.5, Mochi 1, LongCat, Kandinsky) and honest VRAM numbers.

Before you add a video-generation API key to the project, audit the shot list. Most published video — tutorials, product walkthroughs, explainers, listicles, commentary — is built from shots that never need generated motion: screencasts, slide builds, talking-head frames, title cards, thumbnail compositions, B-roll that is really a still with camera movement. The video-gen invoice covers the remaining 10%, and that 10% is the only part worth paying for.

The 90%: shots you compose, not generate

Four patterns cover nearly everything:

  1. Screencast — capture the screen, zoom on the action. Zero generation.
  2. Slide visuals — generated stills (or charts, or diagrams) with build-in animation. The still costs pennies; the motion is free.
  3. Thumbnail composition — generated image, text layer, color grade. Pure image gen.
  4. Animated stills — a generated frame pushed through an ffmpeg chain: Ken Burns pans, zooms, parallax layers, crossfades, shake, light sweeps.

Pattern 4 is our own answer to B-roll. Instead of paying a video model per 5-second roll of the dice, the pipeline generates a still with image gen (FLUX / Qwen-Image on a 3060) and animates it with ffmpeg — any duration, any resolution, no 5-second ceiling, perfectly repeatable. The cost is pennies per frame: the card draws ≈$0.03/hour, so an evening of iteration over a dozen shots rounds to the price of a coffee. The full walkthrough of that pipeline is in Our Video Pipeline Runs Local.

Compare the unit economics. One usable 5-second clip from a paid API is $2–4 once the keeper rate is counted (the math). One animated still locally is effectively electricity — and the hundredth iteration costs the same as the first, which is what makes iteration culture possible at all.

When you actually need video gen

Two honest cases, and only two:

  • Motion you cannot fake — liquid pouring, cloth folding, crowds moving, a specific physical gesture with specific timing. ffmpeg cannot animate what a still doesn’t contain.
  • Scenes that must read as live action — real-world texture and handheld imperfection that composition gives away.

If your shot list has a few of these per month, buy them per project on PAYG (FLUX 3, Gemini Omni, Seedance direct — no subscription) and budget the 5–10× rejection rate up front. What doesn’t survive contact with reality is the monthly plan: burning plan credits on shots you could have composed is the quiet leak, and the subscription traps are cataloged in Video Gen Fine-Print Traps.

The local route: clean licenses, honest VRAM

If you want generated motion without an API, the open models are real — but read the license before the benchmark, and the VRAM column before both. The verified map:

ModelLicenseHonest VRAM
Wan 2.2 (T2V/I2V 14B)Apache 2.080 GB official; community GGUF Q4–Q5 (~9–11 GB files) stretch toward 12 GB, no confirmed 3060 report
Wan 2.2 TI2V-5BApache 2.024 GB — a 4090-class card; under 9 min per 720p clip
LTX-2.5Apache (code); weights under LTX Community license19B with native audio+video; fp8/fp4 quants — revenue-capped at $10M/yr
Mochi 1Apache 2.0~60 GB official, under 20 GB via ComfyUI
LongCat-VideoMIT13.6B dense; no verified hardware figure published yet
Kandinsky 6.0 ProMIT5-second clips with audio; weights gated
The license map has a ban in it

MiniMax-H3’s community license excludes the USA (plus the EU, UK and Korea) — for US work it is unusable, whatever its benchmark score. HunyuanVideo 1.5 is the mirror image: the community license excludes the EU, UK and South Korea, but the USA is explicitly allowed — the early “banned in the US” rumor had it backwards. Read the LICENSE file before you render a single client frame.

The structural advantage of this table isn’t the price — it’s that a rejected take costs watts, not credits, so the brutal video keeper rate stops being a budget line. The trade is hardware: honest numbers put current-gen local video at a 4090-class card or a community quant you test yourself, which is why our own pipeline composes stills instead of generating motion today.

The decision rule

  1. Compose first. If the shot is a still plus motion, a screencast, a slide or a thumbnail — it never needed video gen.
  2. Buy motion per project. For the few shots that truly need generated motion, PAYG APIs at the 5–10× rejection rate — no subscriptions.
  3. Go local only with a license read and a VRAM check. Apache/MIT with honest hardware numbers, or don’t bother yet.

The full market board — paid per-clip prices and local license traps side by side — is the Video generation hub, and the hardware that runs the compose-first pipeline is ranked in the Local GPU AI leaderboard.

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.