A client’s marketing lead sent us 47 generated video clips for a product launch page. Each one was 8 seconds, 1080p, between 9 and 14 MB, and none of them had a poster frame, a transcript, or any record of which prompt produced them. The brief was “pick the good ones”. That took two days, and about half of that time was spent working out which clips were variants of each other.

That’s the real cost of generative media in 2026. The generation step is nearly free and nearly instant. Everything downstream (selection, review, encoding, provenance, disclosure, storage) is where teams quietly lose a week per campaign. If you’re evaluating AI video tools right now, model quality is the least interesting part of the decision.

We’ve shipped this workflow on launch pages, product explainer libraries and a few hundred ad variants. Here’s what actually holds up.

Where generative media actually earns its keep

The honest split, from jobs we’ve billed: generative media wins on volume, variation and abstraction. It loses on specificity.

Good fits we use weekly:

• Ambient hero loops. Abstract motion, light, fabric, liquid, aerial textures. No product, no faces, no text. Nobody can tell and nobody cares.
• Ad variant volume. Thirty backgrounds for the same offer, so paid social has something to rotate before creative fatigue sets in.
• Section imagery on content pages. The images that used to be bland stock photos. This is a straight upgrade in brand coherence because you control palette and lighting across the whole set.
• Storyboards and animatics. This is the quietest, biggest win. Generating a 6-shot animatic in an hour means the client approves a direction before you book a real crew.

Bad fits, learned painfully:

• Your actual product. Unless you’re doing image-to-video from a real photograph with tight motion constraints, the model will redesign your hardware, invent ports, and shuffle your logo. Reference-image features in Runway Gen-4 and the Veo family help a lot, but “a lot” is not “reliably”.
• Legible interface screenshots. Text rendering has improved enormously. It still produces plausible gibberish at small sizes. Screenshot the real thing.
• Anything a regulator reads. Financial, medical, insurance. The approval overhead outweighs the saving.
• Testimonials and people who imply endorsement. Not a quality question. A trust question.

Choosing ai video tools without torching a quarter’s budget

Every vendor sells credits, and credits obscure the only number that matters. Work out your cost per approved second instead:

// Run this after your first real campaign, not before
const generations   = 412;   // total clips generated
const approved      = 38;    // clips that shipped
const spend         = 740;   // USD on credits
const avgClipSecs   = 8;

const acceptRate     = approved / generations;              // 0.092
const approvedSecs   = approved * avgClipSecs;              // 304
const costPerSecond  = spend / approvedSecs;                // ~2.43 USD
const costPerMinute  = costPerSecond * 60;                  // ~146 USD

// Compare against: stock licence, freelance motion designer day rate,
// and the half-day of internal review time you just spent.

Our observed acceptance rates: abstract motion lands near 1 in 3. Anything with a person doing a specific action is closer to 1 in 15. Camera-on-product work has run as bad as 1 in 30 before we gave up and shot it.

Practical selection advice as of early 2026: pick two tools, not six. One video model that handles your dominant use case plus one image model with strong style control, then live with them for a quarter. Multi-tool sprawl kills consistency because every model has a look, and mixing four looks on one page reads as cheap. For video, native audio generation (Veo 3 and similar) matters more than resolution, because sound design is the part marketing teams cannot do themselves. For images, prefer whichever model supports reference images, LoRAs or style references. Brand consistency beats single-image quality every time.

Check clip length limits before you plan a concept. Most models still cap at 8 to 10 seconds per generation. Your 30 second spot is three or four generations stitched in an editor, with continuity you have to engineer yourself.

The asset pipeline: naming, provenance, versioning

This is the part the listicles skip, and it’s the part that saves the week. Every generated asset gets a sidecar JSON file at the moment of download. No exceptions, no “we’ll tidy it later”.

{
  "id": "hero-loop-fabric-004",
  "file": "hero-loop-fabric-004.mp4",
  "created": "2026-01-14T09:22:11Z",
  "tool": "runway",
  "model": "gen-4-turbo",
  "mode": "image-to-video",
  "prompt": "slow drift across folded charcoal linen, soft rim light from left, no text, no people",
  "negative": "text, logo, hands, faces",
  "seed": 884412,
  "refs": ["refs/linen-swatch-02.png"],
  "duration_s": 8,
  "aspect": "16:9",
  "approved_by": "mk",
  "campaign": "q1-launch",
  "disclosure": "synthetic-video",
  "license_note": "brand-owned prompt, no third-party likeness"
}

Name files by intent, never by model output ID. gen4_0d8f2a1c.mp4 tells a designer nothing six weeks later. hero-loop-fabric-004.mp4 sorts, greps and survives a handoff.

Keep the sidecars in the repo even if the media sits in object storage. They’re small, they diff cleanly in pull requests, and when the client asks “can we get three more like the one on the pricing page”, you look up the seed and ship in twenty minutes instead of re-prompting from memory.

Keeping generated video from wrecking your page weight

Model output is encoded for quality, not for the web. A 9 second 1080p clip at 14 MB, dropped onto a hero section unchanged, will kill your Largest Contentful Paint and push your mobile bounce rate up.

Standard treatment, two commands:

# H.264 baseline for universal support, capped bitrate so a busy
scene can't balloon, faststart so it begins playing before full download
ffmpeg -i hero-loop-fabric-004.mp4 \
  -an \                        # ambient loops need no audio track at all
  -vf "scale=1920:-2,fps=25" \
  -c:v libx264 -crf 24 -maxrate 2.4M -bufsize 4.8M \
  -pix_fmt yuv420p -profile:v high -movflags +faststart \
  hero-loop-fabric-004.web.mp4

Poster frame at 1.5s, because frame 0 of a generated clip is often a fade-in
ffmpeg -ss 1.5 -i hero-loop-fabric-004.mp4 -frames:v 1 -q:v 3 \
  hero-loop-fabric-004.poster.jpg

That reliably takes 14 MB to around 1.6 MB. Add a WebM/AV1 variant if your traffic justifies it, but measure first: AV1 encoding is slow, and for an 8 second muted loop the extra saving is often under 400 KB.

Then the markup. Autoplaying video needs muted and playsinline or iOS refuses, and it needs to respect motion preferences:

<video class="hero-media" poster="hero-loop-fabric-004.poster.jpg"
       autoplay muted loop playsinline preload="none">
  <source src="hero-loop-fabric-004.web.webm" type="video/webm">
  <source src="hero-loop-fabric-004.web.mp4" type="video/mp4">
</video>

/ Users who asked for less motion get the poster frame, nothing more /
@media (prefers-reduced-motion: reduce) {
  .hero-media { display: none; }
  .hero { background-image: url("hero-loop-fabric-004.poster.jpg"); }
}

Note preload="none". The poster carries first paint, the video loads after. If you’re building hero sections on a framework that already handles this, the video background components in Canvas ship with the poster-plus-lazy-load pattern baked in, which saves you re-solving it per project. Either way, do not let a generated clip be your LCP element. See the landing page structural teardown for where media belongs in the load order.

Quality control: review before publish, at scale

Human review of 400 clips doesn’t scale, and asking a marketing team to “just check them” produces inconsistent results by clip 50. Build a rubric and score against it. Ours has five binary gates, and a clip fails if any single gate fails:

1. Artefact check. Morphing hands, flickering edges, impossible physics, text that isn’t a real word.
2. Brand check. Palette within tolerance, lighting direction consistent with the set, no competitor-adjacent visual language.
3. Claim check. Does the imagery imply a capability the product doesn’t have? This one catches more than you’d think.
4. Likeness check. No recognisable real person, no recognisable third-party logo or trade dress.
5. Loop check. First and last frame close enough for a clean cut, or the clip gets a crossfade.

Two reviewers, independently, on a 10% sample of each batch. If their verdicts disagree on more than one in five, the rubric is ambiguous and needs rewriting rather than more reviewers. We cover the general shape of this in building a lightweight quality harness, and it transfers almost directly from text to media.

One more thing: review on the device the asset ships to. Artefacts that are glaring on a 27 inch display vanish in a 9:16 feed placement at 400px wide. Reviewing everything at full resolution means rejecting clips that would have performed perfectly fine in context.

Disclosure, rights and the bits that bite later

The compliance picture firmed up considerably through 2025. Treat these as fixed requirements, not nice-to-haves.

Marking. The EU AI Act’s transparency obligations for synthetic content become applicable on 2 August 2026, requiring machine-readable marking and clear disclosure where content could mislead. Google’s models embed SynthID watermarks already. C2PA Content Credentials are the practical interoperable standard, and most major platforms now read them. Preserve that metadata through your pipeline, because a careless re-encode strips it. Check with exiftool after processing, not before.

Indemnity. Enterprise tiers from the big vendors offer output indemnification. Consumer tiers mostly don’t. If a client’s legal team is involved, that single line in the terms decides your tool choice more than quality does.

Disclosure placement. Our position: abstract ambient backgrounds need no visible label. Anything depicting a person, a place, or a product capability does, and the label goes near the asset rather than buried in a footer. We’ve written about where that line sits in the ethics and disclosure of AI-assisted content.

Training data on brand assets. If you fine-tune a style model on your own photography, document consent from the photographer and any models in the shots. We’ve had one project where that paperwork took longer than the entire production.

A workflow that survives a real marketing team

The structure that has worked across several campaigns:

1. Brief in shot lists, not vibes. Six rows: intent, aspect ratio, duration, motion description, forbidden elements, placement. One row per asset. Vague briefs burn credits.
2. Generate in batches of 8 to 12 per shot. Then stop. More variants produce decision fatigue, not better choices.
3. Sidecar on download. Automated if you’re using an API, a template the designer fills if not.
4. Rubric review, two passes. Rejected clips stay in storage with the rejection reason, because they teach you the prompt pattern that’s failing.
5. Encode on commit. One script, versioned with the project. Nobody hand-exports from a GUI.
6. Publish with poster, reduced-motion fallback and disclosure where required.

Automate steps 3 and 5. Keep humans firmly in 1 and 4. The failure mode we see most often is the reverse: heavy automation of review, manual work on encoding. That gets you a fast pipeline for shipping mediocre assets at optimal file size.