Instant Video: Segment-Matched Hero Videos as a Service Layer
Fast video generation crossed a threshold in 2026: a 5-second 768p clip now generates in about 3 seconds and costs $0.20-0.40 (MiniMax H3 Max via fal.ai, currently #1 on the image-to-video leaderboard). Video stopped being a production project and became a response object. This page lays out how Kworq and StoryPress can use that: for clients like RoomSketcher and KK&Jay, as a StoryPress platform module, and eventually as the harness that plugs into interface world models like Runway's Solaris.
Audience: Chris (go/no-go), Guy (strategy), Verne (architecture review). Status: concept, nothing client-facing, nothing committed.
The core call: generate fast, serve static
Nothing streams to a visitor as it generates. Generation happens in batches where 3-second-per-clip speed makes iteration nearly free. The visitor always receives a CDN-cached MP4 that starts painting in under a second. Live-as-generated rendering is the Solaris future; it is not buildable or needed today.
The second core call: per-visitor generation is the wrong unit. An anonymous visitor yields thin signals (UTM, referrer, geo, device). Two visitors from the same ad get indistinguishable 'personalization' at $0.40 a pageview. Per-SEGMENT generation is the right unit: a site has maybe 5-12 segments that matter. Pre-generate a small approved pool per segment, serve instantly from cache, and only consider async live generation for identified high-value visitors (and even then, rendered for their next pageview, never blocking one).
“Video is now cheaper than a click. The advantage goes to whoever owns the segmentation, approval, serving, and measurement harness, because the model underneath is a swappable commodity.”
Pipeline architecture (for Verne)
H3 Max i2v
~$8 / pool
nothing unreviewed
ships
H.264 <1.5MB
+ poster JPEG
content-hashed
URLs
UTM/referrer → segment
5 segments = 5 cache keys
poster = LCP
static-image fallback
variant dimension on
pageview + conversion
Hosting: self-host on R2, not YouTube/Vimeo
A 5s 768p clip is 1-2MB: image-sized. Treat it like an image. YouTube/Vimeo embeds pull ~500KB of player JS, break autoplay unpredictably, add third-party branding, cookies and consent baggage, and cannot do segment selection. They solve problems we do not have (long-form, seeking, bandwidth at scale) and create ones we do. Cloudflare R2 has zero egress fees, sits behind the CDN we already run, and the entire player is a native tag: video muted autoplay loop playsinline preload=metadata poster=... .
Selection at the edge
A Cloudflare Worker (or a few lines in the page head) maps UTM/referrer to a segment ID and rewrites the variant URL. Cache stays sane: 5 segments = 5 cache keys, no per-visitor URL explosion. No PII, no cookies needed for v1.
Performance guardrails
The poster JPEG is the LCP element; the video swaps in after first paint, so Core Web Vitals cannot regress. Hard fallback to the current static hero on video failure, save-data, or reduced-motion. Kill criterion: any LCP regression over 200ms.
Measurement
Variant ID pushed as a GA4 event dimension on pageview and conversion events. Every segment runs a 50/50 holdback against the existing static hero. That holdback IS the experiment; without it we learn nothing.
The approval gate is load-bearing
The machine only ever chooses among human-approved variants. Generation is cheap; approval is the deliberate bottleneck. This is what makes generated video compatible with the no-slop rule and safe under a client's brand. Nothing unreviewed ever ships.
Cost reality
| Item | Cost | Notes |
|---|---|---|
| One 5s 768p clip | $0.20-0.40 | ~3s generation, H3 Max via fal.ai |
| Full 5-segment pool (20 clips) | ~$8 | 3-4 candidates per segment, pre-approval |
| Pilot total incl. regens | $10-15 | generation only; build effort is small |
| Serving | pennies/month | R2 storage, zero egress fees |
| Refresh batch | ~$10 | nightly or weekly, approval is the bottleneck |
Application 1: RoomSketcher intent-hero pilot (first, billable)
RoomSketcher already lives in intent-segmented search. Five v1 segments from their existing queries: real estate agents / listings; homeowners / my-apartment / renovation; builders and developers; interior designers; default (organic/unknown). Signal source: UTM on paid, referrer plus landing query where available, else default.
Generation inputs are RoomSketcher's real product screenshots and renders as image references: brand-true image-to-video from actual UI, not text-to-video fantasy. 3-4 candidates per segment, human approval into the live pool, then the A/B against the current static hero on trial-signup / demo-request conversion, 2-4 weeks. Secondary metrics: bounce, LCP unchanged or better.
Honest caveats: video heroes do not automatically convert better and can bury the CTA (hence A/B, not redesign). Paid traffic segments cleanly via UTM; organic mostly lands in default, so the effect shows first on paid landing pages. RoomSketcher must consent to the experiment on their pages and analytics. Pricing is Kworq's call; this is a billable experiment with a case-study clause.
Application 2: KK&Jay catalog product video
Every product page gets a generated 5s spot from existing product photography (image-to-video is H3 Max's strong suit). Dozens of SKUs for roughly $20, and the same clips feed social. Low risk, high volume, same approval gate.
Application 3: StoryPress 'living hero' module (the real prize)
A platform component: the merchant defines segments and brand references, StoryPress batch-generates variants, the component serves by visitor signal and reports conversion per variant. Nobody in the SMB site-builder space ships this. It reuses the A/B and analytics muscles the platform already has. Prototype on SP0002 (Marigold demo site) before it goes near a paying merchant.
Paid channels: Meta is not a no-op, it is inverted
You cannot generate at auction time; Meta renders creative you uploaded. But Advantage+ / dynamic creative exists precisely to pick the best variant per viewer from a pool you supply. So: mass-generate 50-200 approved variants cheaply, upload the pool, and let Meta's optimizer do per-viewer matching inside their auction. Personalization at cohort level via their machinery instead of individually via ours. Same pattern for email: segment-level video thumbnails. Run this only after an owned-channel pilot shows lift.
Where this meets Runway Solaris
Interface world models generate the interface itself, frame by frame, live. Today that is very alpha. But the harness described here (segmentation, brand references, approval, serving, per-variant measurement) is exactly the layer a Solaris-class engine plugs into later: the hero stops being watchable and becomes explorable, and everything around it stays the same. We are building the socket, not betting on one bulb. Our Solaris early-access request (submitted by Guy) tells the same story.
Sequence and asks
Chris: go/no-go on pitching the RoomSketcher pilot; pricing call.
Verne: sanity-check the serving design (R2, edge worker segment mapping, poster-as-LCP, GA4 variant dimension). Anything that fights the platform, say so now.
On go: Carl drafts segment prompts and generates the first candidate pool for internal review only.
StoryPress living-hero prototype on SP0002/Marigold after the pilot ships.
Meta variant-pool test only after an owned-channel result shows lift.
References (check my claims)
MiniMax H3 Max on fal.ai — pricing ($0.04-0.08/s promo/regular at 768p) and the ~2.8s inference example for a 5s clip.
Artificial Analysis video leaderboard — independent quality Elo + generation-time benchmarks; H3 Max ranked #1 on image-to-video at time of writing.
MiniMax official pricing — first-party per-second rates ($0.08/s 768p, $0.13/s 2K).
Runway: Introducing Solaris — the interface-world-model announcement this concept plugs into; see the storefront paragraph.
Cloudflare R2 pricing — zero egress fees, the hosting-economics claim.
Meta Advantage+ creative — per-viewer variant selection from an uploaded pool, the 'Meta is inverted, not a no-op' claim.
A live demo clip belongs in the hero of this page. Generating it requires a fal.ai account and roughly $5-10 of spend, which awaits Chris's approval. That approval is the pilot in miniature.