Removing the account sounds like it removes work. It actually moves the work into three places.

1. Identity without sign-up

Every visitor still needs an identity, because a paid video takes minutes to render and the buyer has to be able to come back to it. So the browser gets an anonymous identity on first visit, and jobs and credits attach to it. Signing in, with Google or an emailed code, is optional and only exists to carry that history to another device.

2. Recovery

Anonymous identities get lost: a cleared browser, a new phone, a purchase made on a laptop and checked on mobile. The fix is to let a buyer reclaim purchases using things they already have, namely the email they paid with and the receipt number from the payment provider's email. Signing in with the paying email does the same automatically.

3. Honest failure handling

AI renders fail sometimes. In a pay-per-video model, a failure that costs the buyer money is the fastest way to lose trust. The rules are simple to state:

• If a photo cannot be used, nothing is charged.
• If a render fails on our side, what was paid comes back as credits, so the video can be made again.
• Unused credits can be refunded within 14 days.

Credits also make the credit packs simple: 52, 140 or 300 credits, every current template costs 16, and pack credits never expire.

What it costs

The checkout has more edge cases than a subscription, and analytics are lumpier. But the conversion argument is strong: the urge to make a trend video is short, and every screen between "I want this" and "here are my photos" loses people.

AIRap is at airap.tv; pricing details are here. It is an independent project and not affiliated with the artists or the original videos.

The problem with one-shot generation

If you hand a video model a photo and a description of the trend, it will produce something. But trends are specific: a ramp at night, a hook-shaped launch, a flip, a landing in smoke. One-shot generation drifts on exactly those details, and it drifts differently for a hatchback, a motorcycle and a dog.

Step 1: plan

A language model with vision (Claude, in our case) looks at the uploaded photo and writes a plan for that specific input: what the subject is, what must be preserved (colour, dents, roof rack and stickers on a car; markings on a pet), and how the fixed scene should adapt. For a two-person format, it also decides who goes where.

Step 2: render

An image or video model takes the plan plus the photo and renders the result. Because the plan has already resolved the ambiguity, the render instructions can be strict about the parts that make the trend recognisable.

What the split buys you

• Robustness to odd inputs. The Flying Car template accepts cars, bikes, pets and people; the plan step is what makes "anything goes up the ramp" work.
• Honest failure handling. Credits are reserved when a job starts and returned automatically if no result comes back, so nobody pays for a failure.
• Privacy by default. Photos and results are private to the account and purged after 30 days by a scheduled job.

What it costs

Latency. Planning adds a little time, and video renders take minutes. For a 15-second clip that has to match a trend beat for beat, a few minutes of waiting is the price of a result that actually looks like the trend.

The other templates are at yayclip.com.

When we built root pages for Wordiyo we tried a different shape: group the family by what comes before the root.

Branch by prefix

For gen that gives branches like these:

• in-: ingenious, ingenuity, ingenuous, disingenuous
• con-: congenital, congenial
• de-: degenerate, degeneration
• re-: regenerate, regeneration
• mal- / ben-: malignant, benign
• pre-: pregnant, pregnancy

Each branch is a small, learnable unit. De- plus "birth, kind" gives degenerate, falling away from your kind. Re- gives regenerate, being produced again. The prefix carries a direction, and the learner gets to rebuild the meaning instead of reading it.

Show variants, or the tree breaks

The same root hides under different spellings. Gen also appears as gene and, squeezed, as bare gn: pregnant, benign, malignant, cognate. If a page only matches the string "gen", half the family goes missing and the learner concludes the root idea is unreliable. Listing variants explicitly at the top of the page fixes that.

Put the traps on the page

A tree also makes false friends visible. Ingenious (clever) and ingenuous (naive, frank) sit on the same branch and come from the same "inborn" idea, yet they drifted to almost opposite reputations. Placing them side by side, with a one-line note, does more than any mnemonic.

The same goes for words that only look like members. On our port page, opportunity is flagged as coming from the neighbouring Latin word portus (harbour), not from portare (to carry): ob portum, a wind blowing you toward the harbour.

Free layer, deep layer

A tree is useful even at its shallowest: the word, its pronunciation, a definition and the parts it is made of. That layer is free on every word. The deeper layer (what each part means, the story of the root, memory tips) sits on top for learners who want it.

The takeaway

If you are building anything that teaches words, group by structure, not by alphabet. The structure is the lesson.

You can see the full gen tree here: https://wordiyo.com/root/gen

Disclosure: I'm part of the team behind Wordiyo.