Your AI-Generated Video Isn't Production-Ready: The Post-Processing Pipeline Nobody Implements

You generated a video with Kling, Veo, or Sora and dropped it directly onto your website. You've probably doubled your page weight, obliterated your LCP, and given your visitors an experience you wouldn't tolerate yourself.

At Mintec, we stress-tested every major AI video tool during the first half of 2026: Veo 3.1, Kling 3.0, Sora 2, Seedance 2.0, and Runway Gen-4. We put them all through production-grade testing — and there's an uncomfortable truth none of them put on their landing pages: no AI-generated video is web-production-ready as it comes out of the tool.

This isn't a knock on the tools. They were designed for content creators posting to social platforms, not for web teams who need to maintain healthy Core Web Vitals. The gap between what these tools export and what a professional website needs is enormous — and closing it requires a post-processing pipeline that most agencies and in-house teams still aren't running.

This article is that pipeline.

The problem in numbers: what AI video actually weighs

We tested every tool with the same prompt: "drone flying over a tropical forest at sunrise, 30 seconds, slow motion." The results tell the real story:

A single Kling 3.0 4K clip weighs 18.2 MB for 30 seconds. For context: the recommended total page weight for a modern website is roughly 1.5-2 MB. One unoptimized AI video multiplies that by 10.

But raw weight isn't the only problem. Native exports from these tools:

• No poster frames — the browser has no static image to show while the video loads, forcing LCP to wait for the actual first video frame
• Legacy codecs — most export H.264 in standard MP4 containers, ignoring AV1 which would cut weight in half
• No streaming metadata — without duration, bitrate, or resolution in file headers, some players and CDNs have to guess basic properties
• No quality variants — a single file tries to serve every device, from an iPhone 16 Pro Max to a low-end Android with 3GB of RAM

The 4-step pipeline we use at Mintec

After burning hundreds of dollars in API credits and dozens of hours testing, we landed on this pipeline. It's not optional — it's the minimum standard for publishing AI-generated video on a professional website.

Step 1: Transcode to AV1 (or HEVC as fallback)

The most impactful step. AV1 reduces video weight by 50-56% vs H.264 at equivalent visual quality. We use SVT-AV1 (software) for server-side pipelines or hardware encoders (NVIDIA NVENC Ada, Intel Arc) for real-time processing.

For 30 seconds of 1080p video:

• SVT-AV1 preset 8: ~25 fps encoding — processes in real time or better
• SVT-AV1 preset 5: ~10 fps — better compression but 2.5x slower

Rule of thumb: preset 8 for daily production pipelines, preset 5 only for hero videos rendered once.

Not every device supports AV1 in hardware. 88% of devices certified since 2021 do (Intel 11th Gen, Apple M3, NVIDIA RTX 30, Snapdragon 8 Gen 1). For the rest, keep HEVC (x265 slow) as automatic fallback via canPlayType() capability detection.

Step 2: Generate a poster frame

The poster frame is the first video frame, exported as an AVIF image (~150-300KB). It loads immediately (native lazy loading) while the actual video downloads in the background.

In our luxury brand portal project, this single technique dropped LCP from 6.2s to 1.2s — the highest-impact optimization we measured. Full details in our Astro 6 content architecture case study.

How to do it:

# Extract first frame with FFmpeg
ffmpeg -i ai_video.mp4 -vframes 1 poster.%d.avif

Make that poster the immediately visible element. Load the video on demand with loading="lazy" or use Server Islands if you're on Astro 6.

Step 3: Metadata packaging for streaming

A video without streaming metadata is like a web page without meta tags — it works, but every intermediary system (CDN, player, crawler) has to guess basic properties.

Minimum metadata every file needs before production:

• Exact duration in seconds
• Resolution (width × height)
• Codec (av01.0... / hev1...)
• Average bitrate (kbps)
• Frame rate (fps)
• Rotation and orientation (critical for vertical videos)

Package with -movflags +faststart so metadata lives at the start of the file (moov atom), enabling progressive streaming without downloading the full video first:

ffmpeg -i ai_video.mp4 -c:v libsvtav1 -preset 8 -movflags +faststart output_av1.mp4

Step 4: CDN with multi-platform variants

One file doesn't serve everyone. We create 3 variants per video:

• High (AV1): Modern devices with hardware support and fast connections
• Medium (HEVC): Devices without AV1 support (pre-2020)
• Low (H.264): Universal fallback for very old browsers or slow connections

Selection via a simple capability detection function:

function getBestVideoSource(variants) {
  if (document.createElement('video').canPlayType('video/mp4; codecs="av01.0.05M.08"')) {
    return variants.av1;
  }
  if (document.createElement('video').canPlayType('video/mp4; codecs="hev1"')) {
    return variants.hevc;
  }
  return variants.h264;
}

Combined with Server Islands for content architecture at scale, the video loads deferred and the correct variant is selected without client-side JavaScript overhead.

When to skip steps (and when not to)

Not every AI-generated piece needs the full pipeline. Our decision framework:

The general rule: if the video appears on a website's main page, hero section, or as directly embedded content, the full pipeline is mandatory. For everything else, apply the "how much would my site lose if this video is 10x heavier than it needs to be?" test.

What changes when you implement the pipeline

Across our projects, applying this full pipeline delivers consistent results:

• Weight reduction: 62-68% (from 5-18 MB to 1.9-5.8 MB)
• LCP improvement: 0.7-2.5 seconds of direct improvement (depending on whether video is the hero element)
• CDN savings: ~60% less bandwidth served
• Browser coverage: 100% of browsers served with the correct variant (vs ~88% if you use AV1 alone)

These numbers aren't theoretical. They come from our performance budget framework for synthetic media, developed specifically to measure and control this impact.

AI-generated video is the future. The pipeline is the present.

AI video tools improve every month. Kling ships 4K, Seedance solves character consistency, Veo generates native audio. But none of them are building the post-processing pipeline needed to make that video work on the real web.

That work falls to web and production teams. And the sooner we integrate it as a standard step — on par with image optimization or CSS minification — the sooner we stop seeing sites that choke because someone uploaded a 20MB video straight from an AI tool.

AI-generated video isn't the problem. Serving it raw is. And that problem is solved with a pipeline, not with miracles.