Changelog

The Free plan now makes one test video of up to 30 seconds.

A new account is given the few credits that video costs, where it was given 60. A video that fails before it is made, or that you stop, does not use the test up; one that is made does, even if it does not pass our quality checks. An account on the Free plan that has already made a video needs a plan to make another, and the credits it holds stay its own. An account whose paid plan has ended, or to which we have added credits by hand, goes on spending the credits it holds. Credits earned by inviting people are spent on a paid plan.

The voices Adam and Lewis are offered again, so nine voices are offered now.

A series set to Adam makes its next videos in Adam again, rather than in Fenrir, and one set to Lewis in Lewis rather than George, and Start again and Make another like this open the create form with that voice chosen. The Stoic wisdom and Motivation niches fill in Adam again, and Greek mythology Lewis. Videos made in Fenrir or George in the meantime keep that voice: a line you change when regenerating a scene is read in the same voice as the rest of that video.

Drawn pictures are now checked before the video is made, and one found to show nudity, blood or writing is drawn again once.

A picture-review model is asked about each picture; it can miss things, so this is a check and not a promise. A picture still found to show nudity or blood after its redraw is blurred, and that version of the video fails its quality check, so it is not published and its credits are returned. Something from the wrong time for the story is noted under Quality control rather than drawn again. When the check cannot run, the video is made as before.

Each scene’s line of narration is listened back to, and a line heard saying a different figure is recorded again.

The line is recorded once more, a little slower, and the new take is kept only if every figure in it is heard right and it gets no name wrong that the first take got right. A figure or a name still heard differently is noted under Quality control, naming the scene. When the check cannot run, the video is made as before.

Narration is now read up to 15% faster, at a pace set by the video’s tone and voice, and the voices Adam and Lewis are no longer offered.

Energetic is the quickest tone and calm the slowest, but some voices are read at much the same pace in every tone. A script the language model writes is sized for the pace it will be read at. Of the voices offered, Adam and Lewis had the two lowest grades from the makers of the voice model, so seven voices are offered now. A series set to Adam makes its next videos in Fenrir, and one set to Lewis in George, and Start again and Make another like this open the create form with that voice chosen. Videos you already have keep their voice and pace: a line you change when regenerating a scene is read in the same voice, and at the same pace, as the rest of that video.

Videos now carry quiet sound effects: a hit on the opening and where the ending begins, a whoosh on scene changes, and a rising sound into up to two turns in the story.

They are placed by rule, at most four in any ten seconds and none within half a second of another, so some scene changes go without one. They sit below the voice. In a calm or inspirational video they play at half their usual level or lower, and a calm video gets no rising sounds. When there is music, it dips in the short pause before the ending. Regenerating a scene of a video made before this change gives it only the opening hit and the whooshes.

Pictures drawn in a named art style now hold together better, and every video is encoded in the colour standard phones and browsers expect.

Reds and yellows no longer play back slightly hotter than they were drawn. In any art style other than Match the format, a narrow plain border the generator draws around a picture is cut off where one is found, though not every border is; a script you paste with its own Style: line is left as drawn. In Comic ink, Watercolour, Flat vector and Retro print, when a video’s pictures come out in many unlike colours, they are pulled part of the way toward two or three colours the video shares. Retro print now asks for navy blue, burnt orange and cream, and Watercolour for sage green, ochre and slate blue. When a sentence of the narration gives the story’s year with its season or time of day, such as ‘on a winter morning in 1919’, the pictures are asked for that light and, for most places, that season. Regenerating a scene of a video made before this update does not trim its pictures or pull their colours together.

The language model’s scripts now pick their opening line for what gives a viewer a reason to stay, and the opening card shows a short headline of it where one fits.

The first line is weighed against the other openings the model offers and, when it passes the checks, against one-sentence lines from the first part of the video. A figure other than a year, a name or a contradiction counts for a line; opening on a date or on who reported it counts against. A clearly stronger line can be moved up to open the video, with its picture. A script you write yourself keeps its first line. The opening card now shows up to six words of the opening line in capitals where such a cut still says something on its own and fits at full size, and otherwise the sentence, or its first part, as before. It stays until the line has been said, for 2.5 to 4.5 seconds, and the thumbnail, taken while it is up, shows the headline too.

A News explainer now opens with an AI-GENERATED label, and the caption suggested for posting it ends by saying it was made with AI.

The label is at the top of the picture for the first two seconds. The suggested caption ends ‘Written and voiced with AI.’, or ‘Voiced with AI.’ when the language model did not write the words; the line is already in the caption box when you publish, and what you leave in the box is what gets posted. A series that posts by itself includes the line. Regenerating a scene of a News explainer adds the label to the new version, even for a video made before this change. Videos in other formats get neither.

The same video can no longer be posted to two of your accounts on one platform.

In the publish panel you tick one TikTok, YouTube or Instagram account for a video; ticking a second account of the same platform unticks the first. Once the version you are posting has gone to one account of a platform, is on its way there, or may be there because the platform never said what became of it, the panel names that account and does not offer the video to another account of that platform. If that post fails or is cancelled, the video is free to go to another account. A series with two accounts of one platform switched on now posts only to the first of them in its list, and the series form and page say which. Until now one video could go to several accounts of a platform at once. Give each account videos, and a series, of its own.

Videos posted to TikTok now leave out your brand kit’s logo and watermark, unless you turn on Show on TikTok for the kit.

TikTok asks apps not to add branding to the videos they share there. Each kit on the Branding page has its own Show on TikTok switch, and it starts off. While it is off, a new video made with the kit is also made without its logo and watermark, and that copy is the one posted to TikTok; YouTube, Instagram and your download get the video with the marks, as before. While the switch is off, a video that carries the marks and has no such copy, such as one made before this change, is not posted to TikTok, and the publish panel says why: turn the switch on, or change a scene so the video is made again with the copy.

A series that posts automatically now sends one account at most two videos a day, and holds back a video too much like an earlier one.

The limit counts automatic posts from all your series together over any 24 hours; posts you publish yourself are not counted. A third waits until the older of the two is a day old, and the series page says when it will go. While an account already has two videos from your running series being made or waiting to be posted to it automatically, a scheduled slot that would post to it is skipped before its video is made, so no credits are used. Each new series video is compared with up to ten of the series’ earlier videos, and it is not posted automatically if its narration shares too many five-word phrases with one of theirs, if it opens with the same first eight words as one of them, or if more than 30% of its pictures look the same as one of theirs. The series page says why, and the video stays on its project page for you to publish yourself.

Videos on a paid plan now get a head start in the queue over Free plan videos, and a waiting video tells you when renders are running behind.

A new video, a scene change or a series Run now on a paid plan goes ahead of Free plan videos queued up to about half an hour before it, and every video moves up the longer it waits, so none waits for ever. When a video someone asked for has waited more than 20 minutes to start, the page of a video waiting in the queue says that renders are running slower than usual and that nothing is lost. The create page says the same, unless it is already saying the picture generator is offline.

Posts to Instagram now carry at most five hashtags, counting any in the caption.

Instagram has announced a limit of five hashtags for a reel. The publish panel says so under the hashtags, and asks you to remove any past five before you post. A post scheduled earlier with more goes out with five: hashtags in the caption past the fifth keep their word without the #, and the extra ones from the hashtag list are left off.

You can now choose a length of up to 2 minutes.

The length slider on the create and series forms stops at 2 minutes, and a script written for 2 minutes is cut to fit them. A script you paste in has to read in 2 minutes or less, counting the short pauses between scenes. The finished video is as long as its recorded voice, so it can still come out a few seconds over. A series set to a longer length before today keeps running: each new video is written for 2 minutes, and says so under Quality control on its page. Videos you already have are not changed.

A page or link opened over plain http now redirects to the secure https address.

The create form and the series form say which brand kit a video gets, and let you pick another or none.

A video used to take the workspace’s default kit without a word on the form. Start again and Make another like this keep the kit the video was made with.

You can withdraw from a plan within 14 days of buying it, from the Billing page.

Before checkout you tick a box asking for the plan to start as soon as you pay. If you then withdraw, you are refunded what you paid less the plan’s own credits you have used; once every credit the plan includes has been used, it has been provided in full and the right ends. A purchase made without that request is refunded in full.

Your credit activity is on the Billing page: every hold, charge and return, newest first and in words.

Checkout asks for your billing address and, for a business, a VAT number.

Every page of the app says when your workspace’s billing needs attention, such as a payment that did not go through.

A finished video offers Make another like this, which opens the create form with the same settings.

You get a notification in the app when a video fails, a post fails, a series pauses or a connection expires.

Help and Contact are in the app, and a failed video’s support reference goes into the contact form for you.

When ClipSpell posts to TikTok directly, it asks you who can see the post, whether comments, duets and stitches are allowed, and whether it promotes a brand.

The answers are yours to give for every post, as TikTok asks, and a series keeps them per account.

Every upload to YouTube is marked as altered or synthetic content.

The pictures and the narration voices ClipSpell makes are generated, and YouTube asks for that to be labelled.

TikTok and Reels covers are the video’s thumbnail frame rather than its first frame.

The picture generator is now asked to draw people clothed and to leave an injury to the narration rather than draw it.

The project list pages past the first fifty videos, and can be searched and filtered.

The dashboard showed and counted only the newest fifty, so a busy series lost its older videos from view within weeks.

The project page updates by itself while a post is on its way out.

Scripts written by the language model now come out close to the length you chose.

They often fell well short of it, at a half to three quarters of the words asked for.

Scenes next to each other vary their pictures instead of repeating one.

One test video drew the same storage tank in ten of its thirteen scenes.

A single video no longer ends by promising a next part, or with a moral added on at the end.

The opening card keeps whole sentences, and a caption still being spoken when it leaves is shown rather than skipped.

Download MP4 and Download captions keep working however long the page has been open.

Both links used to expire an hour after the page loaded.

Music tracks and logos larger than 1 MB upload now, up to 25 MB.

A data-deletion request from Facebook now erases the connected Instagram account’s id, name and access, as the data-deletion page says it does.

A factual video’s script is written against the Wikipedia article on its topic, when one is found, and the project page credits the article.

A sentence of narration whose figure or name the article does not support is dropped, never reworded. The project page names the article and links the exact revision it was written from, as its licence asks. That is one check against one source, not a fact check.

The Word pop and Highlight pill captions are bigger, by about a fifth and a sixth.

The end card waits for the last caption to finish instead of covering it.

Connect more than one account on a platform, and post one video to several of them at once.

Each account has its own row on the connections page, and the publish panel lists every one with a tick box. How many accounts you can connect depends on your plan.

You can leave a review of ClipSpell from a project page or from Settings.

A review appears on the home page only with your consent and after a person has read it, and none are shown until there are three.

Signed in, the home page is your own Home, with your balance, instead of sending you to Projects.

A video is never finished with plain backgrounds where its pictures should be.

If the picture generator is away, the video waits in the queue and says so, with its credits returned until it runs, rather than going out without its pictures.

A script that says the same thing twice is caught and rewritten before the video is made.

Captions are laid out better: lines are balanced, kept inside each platform’s safe area, and numbers are shown as digits.

If YouTube’s daily upload limit is reached, a post waits and goes out later instead of failing.

Hear a mood before you choose it: every mood has its own music and a play button, and you can pick one of your own tracks.

The create form shows a picture for every format and art style, and plays every voice before you choose.

A failed or stopped video can be started again with its settings, renamed or deleted.

Captions download as SRT as well as VTT, and the project list shows each video’s poster.

Videos play and seek on an iPhone.

New pages for pricing, features and common questions, and a site menu behind the logo.

Every figure on them is read from the product itself — the credits a new account gets, what a video of a given length holds, how many voices and styles there are, and where a post can go today — so the pages cannot promise something the product has stopped doing. A new account now opens on a first-video guide instead of an empty list.

Invite people and earn credits. Your link is in Settings.

Anyone who signs up through your link starts with extra credits, and once they finish their first video you get credits too. Paid in credits rather than money, so there is no threshold to reach and nothing to invoice — and it is deliberately the finished video rather than the signup that pays you, which is what keeps the programme from being gamed.

We were telling you your scripts came from a template when a language model was writing them.

The product checks whether a model is configured before saying so, and it was only looking for one kind. On an installation running a local model it concluded there was none and showed a notice saying scripts came from the built-in template writer. The scripts were model-written the whole time. We promise never to present template output as model output; this was the same promise broken the other way round, and the terms and privacy pages now name which writer produced your script.

Asking for a password reset on an installation that cannot send email now tells you who to contact.

It used to say what an administrator would need to configure, which is no use to somebody locked out of their account and reading it. The same wording had spread to four other places; all five now say what you can actually do.

A render that stops for good no longer shows as still working.

If the machine was restarted at the wrong moment, a video could be left reading as in progress for ever, with nothing coming. Those now end properly and say so, and no credits are spent on them.

You can type your topic on the front page and it follows you through signing up.

The site is faster. Every page now loads about a sixth of the code it used to.

It had been served by the development server, which sends a large amount of tooling no visitor needs. Nothing about the product changed — it simply arrives quicker.

A video that fails our own quality checks is no longer charged for. Its credits come back in full.

The promise was already on the pricing page and the code did not keep it: a render that finished and then failed the checks was billed like a successful one, while everything else about it said it had failed. Re-cutting a single scene was affected the same way. If it happened to you, get in touch — the ledger records every charge, so it can be checked.

A script you write yourself is now priced against what you actually supplied, not against the length you asked for.

The estimate could be far too low for a long authored script, which meant a hold that did not cover the work and a surprise at the end.

A render that resumed after an interruption could place a second credit hold. It no longer can.

The site now tells you what each platform can actually do from here, rather than stating it in advance.

It had said publishing was unavailable everywhere while two platforms were connected and working. Publishing reach is now read from the live configuration, in one place, so the landing page, the privacy notice and our own checks cannot disagree.

The terms page named the wrong image model and contradicted itself about closing your account.

Both were true when written and had quietly stopped being true. The model is read from the running configuration now, and states no licence at all rather than a confident wrong one. Closing your account is a button in Settings, which is what the page says.

A failed connection to a social platform could record one of our own credentials in our internal logs.

No customer data was involved and nothing left this installation. Provider errors are now filtered before they are written down.