Releases · kleZ799/clipmint

Repository navigation

Release list

ClipMint v1.26.1 Latest

Latest

github-actions released this 05 Oct 11:55

v1.26.1

3dd8ca9

Two fixes.

The window no longer opens unstyled

Now and then ClipMint opened as plain black text on white, with the menu
lying on top of the page and nothing laid out. The page had arrived but its
look hadn't, and the window had no way to fetch it again. The look now comes
inside the page itself, so it can't go missing on its own. If part of the app
ever fails to load, the window reloads itself.

If it happened to you, closing and reopening the app fixed it. After this
update it shouldn't happen at all.

The prompt box reads "hooks" correctly

Typing "hooks that ask a question" switched on Start each clip with its own
loudest moment, because the word "hooks" was taken as that setting. It now
means what you meant: the kind of moment to look for. That setting is decided
only by its checkbox. The reply under the box also says the right number of
clips when you don't name one: "I'll find 10 clips with hooks that ask a
question", not "a clip".

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

ClipMint-linux-x86_64: about 280 MB, x86-64, self-contained, glibc 2.35 or
newer. chmod +x it and run it. Still untested on a desktop distribution.
Every published build is started by the release workflow and asked for its
interface before the release exists, but nobody has yet made a Short with it
on Fedora or Ubuntu. Take ClipMint-linux-README.txt too.

ClipMint-macOS-arm64.zip: about 167 MB, Apple Silicon only. Still a beta,
and still has never been run on a Mac. I don't own one, so it is built and
checked by GitHub's macOS runners and published without a person opening it.
Uploading to YouTube has only been tested on Windows. If it doesn't start, or
signing in doesn't behave as described, that is a bug here:
~/Movies/ClipMint/app.log and
an issue are how it gets
fixed.

Also published, for installs from before the rename only: StreamToShorts.exe,
StreamToShorts-linux-x86_64 and StreamToShorts-macOS-arm64.zip, the same
three files under their old names. A fresh download wants the ClipMint ones.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.26.0

github-actions released this 04 Oct 10:36

v1.26.0

ea93097

Type what you want in your own words, and ClipMint understands it.

Ask for it like a message

The What to make box used to understand a fixed list of phrases. Now your AI
model reads the whole thing, the way you'd text a friend who edits for you:

give me 2 funny ones and 2 where I ask chat a question, around 30 sec each, 720p is fine

Spelling mistakes, Hinglish, several requests in one sentence: all fine. It
works out every setting your words imply, including how many clips, how long,
what kind of video, quality, framing, captions and the edit. Then it answers
right under the box:

✦ I'll make 4 clips: 2 funny ones and 2 where you ask chat a question, about 30 seconds each, in 720p.

Mixed requests really are mixed. Ask for 2 funny and 2 that ask a
question, and you get 2 of each. Before, it took the four best moments,
which could all be funny. If the video doesn't have enough of one kind, the
best of the rest fill in, so you never get fewer clips than you asked for.
Quality from your words. "720p is fine" or "best quality" sets
Source quality under Render, tagged set by your words.
Anything the Edit box can do, you can say too: "yellow captions at the
top in bebas, no emoji".
Short phrases still work instantly, even with no AI key. The model
reads your message about a second after you stop typing. If it's slow, the
preview says it's still reading and catches up.

Your data

What you type in the box is sent to your AI provider (Gemini, Groq or
OpenAI) with your own key, so it can be understood. It was already used to
pick moments; the privacy page now says so.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.25.0

github-actions released this 04 Oct 09:40

v1.25.0

cc1a611

Captions are now yours to style: ten fonts, any colour, and top, middle or
bottom.

Font, colour and placement

Under Render → Edit, below the caption style, there are three new rows.
The style you picked still decides the rest: capitals or sentence case, how
many words at a time, the pop, the box.

Font. Keep the style's own, or pick one of ten: Montserrat, Anton, Bebas
Neue, Poppins, Archivo Black, Lilita One, Luckiest Guy, Bangers, Permanent
Marker and Bungee. Seven of them are new in this release, and like the others
they ship with the app, so they look the same on every PC. Each one is sized
to fill the line the style was made for, so a narrow font doesn't come out
tiny and a wide one doesn't run off the screen.
Colour. One row for the word being said and one for the rest of the
line. Pick a swatch, or the rainbow one for any colour you like. Dark text
gets a white outline (and a light box on Clean), so it stays readable.
Placement. Auto, Top, Middle or Bottom. Auto keeps the old behaviour: on
the seam between webcam and gameplay, otherwise low, above the apps'
buttons. Moving the captions on a single clip in the Frame panel still
works, and wins for that clip.

The live preview shows all of it before you render, and it now draws the
caption at its real size. Before, it showed captions half again bigger than
they came out.

Your choices are remembered on this PC. Clips you already made keep their
look; a trim or a caption fix on an old clip renders it the way it was.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.24.2

github-actions released this 03 Oct 06:28

v1.24.2

ab9b608

Transcribing works again.

The fix

v1.24.1 stopped every job at Transcribing with "Failed: open() got an
unexpected keyword argument 'metadata_errors'". Nothing in your video or
settings caused it.

A library ClipMint uses to read audio put out a new version three days before
v1.24.1 was built, and it removed something the transcriber still asks for.
The build picked up the new version, so every transcription failed. This
release goes back to the version that works, and the build won't pick up the
broken one again.

Jobs that failed can be finished with Try again. The download is
kept, so nothing is fetched twice.
Nothing else changes. Hinglish captions from v1.24.1 are still here.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.24.1

github-actions released this 02 Oct 08:34

v1.24.1

8d4af7a

Hindi captions now come out in Hinglish.

Hindi, written the way it's read

When a clip was spoken in Hindi, its captions came out in Devanagari, and so
did every English word said along the way: "भाई ये गेम क्या है यार". That's
an accurate transcript and the wrong caption. Hindi Shorts are read in Roman
script, the way people type in chats and comments.

Now the same clip is captioned "bhai ye game kya hai yaar".

Your AI provider does the spelling when one is set up in Settings
(Gemini, Groq or OpenAI). It spells Hindi the casual way and knows that
गेम is the English word game, not gem.
Without one, ClipMint spells it by rule. Common words come out right
(kya, nahi, yaar, accha, karna, samajhna). English said in Hindi comes out
only roughly, apart from the stream words it already knows (video, game,
bro, stream, subscribe).
The timing doesn't move. Each word is rewritten in place, so the
highlighted word, pause cuts and punch-ins land exactly where they did.
A word it spells wrong can be fixed in the player like any misheard
word.
Clips you already made switch over the next time they render again: a
trim, a reframe or a caption fix.

English, and every other language, is captioned exactly as before. Only words
in Devanagari are touched.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.24.0

github-actions released this 29 Sep 05:36

v1.24.0

00d69ed

ClipMint now reads how the Shorts you posted actually did, tells you what's
working on your channel, and learns from it, carefully.

Until now the ranking was a guess that was never checked. Now it gets checked.

How your Shorts did

Connect your channel in Settings, and the library opens with a new panel:

Every clip on your channel is found, including ones you uploaded
yourself, which ClipMint matches by title. Each card shows its views, and
Most views sorts a run by them.
It says what's working, in plain words. For example: "Clips with a
title the AI actually wrote got a median of 42.5 views; clips without got
3.5 (10x more)."
Every pattern is tested against chance. With a few dozen Shorts most
differences are noise, so a finding is marked Looks real only when it
clearly isn't. Everything else is listed as no clear link yet, rather than
dressed up as advice.
Only clips at least two days old are compared. Yesterday's Short hasn't
had its views yet.

The ranking learns

When one of the measured signals (loudness, reaction words, motion, build-up or
chat) really does go with views on your channel, it counts for more in your
next run. It only moves after a dozen clips, never by more than 40%, and only on
a finding that held up. Until then the ranking stays exactly as it was, and the
panel says so.

On the channel ClipMint was built with, that's what happened: none of the
ranking scores tracked views yet, so nothing moved. The one clear finding was
the titles, and it was 10×. If an AI provider runs out of quota before titles
are written, a free Groq key in Settings is the fix.

Your data

It uses the same YouTube connection as uploading. There's no new permission to
grant. ClipMint checks by itself at most every six hours, or when you press
Check now. The numbers stay on your PC, are deleted after 30 days without a
refresh (YouTube's rule), and go the moment you disconnect. The
privacy policy has the
details.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.23.0

github-actions released this 29 Sep 04:57

v1.23.0

1e398a7

When ClipMint frames a clip wrong, you can now fix it yourself: drag the frame
to where it should be, and move the captions up or down.

Automatic framing is right most of the time, and every clipping tool's reviews
say the same thing about the rest: the crop locks onto the wrong person, cuts
the action in half, or shows the empty side of the game, and there's no way to
fix it by hand. Now there is.

Frame it yourself

Open a clip and press Frame in the player (or R).

You see the whole picture. The full source frame opens with the part the
Short shows marked on it, and everything it leaves out dimmed. A slider picks
which moment of the clip to look at.
Drag the box where it should be. It renders again from the original
video in a few seconds, and holds that framing for the whole clip.
On webcam-over-gameplay clips, your webcam stays found for you. It's
shown shaded, and the box is the game underneath it, for when the action is
off to one side.
It sticks. A trim or a caption fix afterwards keeps the framing you
chose. Back to automatic hands it back to the app.

Move the captions

In the same panel, the Captions slider moves them up or down, for when they
cover a face or the thing you're pointing at. A line on the clip shows where
they'll land before anything renders. They stay clear of the top of the frame,
where the clip's hook line goes.

Moving only the captions leaves the framing automatic, so a face-following clip
keeps following.

Needs the original video

Both re-render from the video ClipMint downloaded, so it has to still be on
this PC. If you've cleaned up downloads, the panel says so rather than failing
quietly.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.22.0

github-actions released this 28 Sep 16:26

v1.22.0

863c396

On a podcast or interview with two people in the shot, ClipMint now frames
whoever is talking.

A vertical Short only fits one of two people sitting side by side. ClipMint
used to pick the biggest face, which is whoever sits nearer the camera, and
stay on them. So the other person's best line went out over a shot of someone
listening.

It follows the conversation

It watches mouths, not sizes. Each person's mouth is compared frame to
frame, with head movement taken out first, so a nod or a laugh isn't
mistaken for talking.
It cuts like an editor. When the other person takes over, the frame
cuts to them rather than panning across the table. It cuts as they start
talking, not a second late, because the whole clip is planned before it's
rendered. No shot is shorter than two seconds.
Tested on real episodes. On two episodes of a two-person podcast, cut
into 30-second clips the way ClipMint cuts them, it framed the speaker 86%
and 96% of the time. Following the biggest face managed 73% and 10%, and
caught none of the second person's lines.

What doesn't change

One person on camera: framed exactly as before.
Shows that cut between a camera per person: already show the speaker,
and are left alone.
Webcam-over-gameplay streams: use their own layout and aren't affected.
A face in a video playing on stream, or a poster behind you: too small
to take the shot away from you.

It adds about a third of a second to each 30-second clip.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.21.0

github-actions released this 28 Sep 15:00

v1.21.0

02fcd99

ClipMint now reads a stream's chat, and ranks on what the live audience
actually reacted to.

Every other thing the ranking knows about a moment is a guess. The AI guesses
from the words, the loudness check guesses from the volume, the vision check
guesses from four frames. A live chat is not a guess. It is the people who were
watching, typing "WHAT" at the same second.

It reads the chat

Fetched while the video transcribes. On a YouTube stream that kept its
chat replay, ClipMint downloads it alongside the transcription, so it costs
no extra waiting on most runs. It's kept beside the video as a tiny summary,
so a second run over the same stream doesn't fetch it again.
It finds the moments the words miss. The biggest bursts of chat are
marked in the transcript the AI ranks from, so it now finds moments nobody
said anything about: a clutch played in silence, a jump scare answered with
a gasp, something on screen chat saw before you did.
Every candidate is scored on how hard chat reacted. Each spike is
measured against how busy chat usually is at that point in the stream, so a
packed finale doesn't drown out a big moment early on. Hellos, goodbyes and
"just subbed" floods don't count.
It knows when you asked for it. When you say "everybody type W" or
"type 1 if you agree", the flood that follows is your request, not a
moment. On the stream this was tested on, half of the biggest bursts were
exactly that, so ClipMint listens for you asking and sets those aside.
A new Chat bar on the clip card, next to Hook, Moment, Look, Energy and
Pace, and a note, "Chat went off", when the live audience reacted.

When there's no chat

Nothing changes. A video that isn't a stream, a stream whose channel turned
the chat replay off, or a chat too quiet to mean anything (under about 150
messages, or 3 a minute) is ranked exactly as before, and nothing already
ranked is thrown away. Small streams usually fall here: a "spike" in a chat of
a dozen people is three of them saying hello at once, and ClipMint would
rather not pretend.

Only YouTube chat replays are read for now. A Twitch or Kick VOD is ranked
without its chat.

Downloading a VOD yourself? Save its chat next to it with
yt-dlp --write-subs --sub-langs live_chat <url> and ClipMint will read the
.live_chat.json file beside the video.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

ClipMint v1.20.0

github-actions released this 26 Sep 09:19

v1.20.0

c2009dc

ClipMint now picks the moments a stranger would watch, looks at them before
it cuts them, and puts each clip's hook on screen.

This release came from a real channel's numbers. Shorts where you could see
what happened (a clutch, a physics mess, a famous story twist, a game's own
joke with the streamer's comeback) reached about a thousand viewers each.
Shorts of the streamer reacting to something the clip never showed, raging at
nothing visible, or talking to chat averaged eight. The old ranking had those
the wrong way round.

It picks what a stranger would watch

A new Short is shown first to a handful of people who have never heard of you.
Every moment is now ranked on whether one of them would stay:

On a stream, the payoff has to be on screen. A visible payoff with your
reaction ranks first. Reactions to something off-screen, swearing with no
visible cause and chat talk are pushed down.
Your brief narrows it, it doesn't lower the bar. "Only the rage
moments" now means the rage moments a stranger could follow, not every time
you swear.

It looks before it cuts

Every candidate's opening is checked for black screens, dark rooms and
frozen frames, straight from the video, and how much the picture moves is
weighed in. It takes about half a second a clip.
Then the best ~20 are shown to the AI as a stranger would meet them:
the first instant, a second and a half in, the payoff and the end. Menus,
loading screens and stream dashboards drop to the bottom before anything is
rendered. It's careful, not trigger-happy: it only marks a clip down hard for
something the frames prove, so a quiet story beat that lands isn't thrown
out.
Every clip card has a new Look bar, next to Hook, Moment, Energy and
Pace, showing how the frames came across.
No more near-duplicates. Two clips that share more than a quarter of
their footage are one clip. Before, two Shorts cut from the same 14 seconds
could both come back.

The hook line is on screen

Every clip's on-screen hook, the short line written for it next to its title,
now goes across the top for its first couple of seconds, in your caption
style. When the clip opens on a replay of its loudest moment, the line is on
that too. Turning captions off turns it off.

Titles the AI didn't write don't go to YouTube

When the AI can't write a clip's title, it falls back to the first thing said
in the clip, which on a stream can be "Holy shit | Dying Light".

It asks again first, in smaller batches, and it's no longer thrown by an
AI that numbers its answers oddly.
Those clips can't be uploaded as they are. In Upload all they start
unticked with a note, and neither upload button sends one until you use
Rewrite titles or type your own.
Out of quota? It tells you. Titles are the last thing a run asks for, so
on a long stream they're the step most likely to hit a free tier's daily
limit. ClipMint now stops asking and says so. A second provider's key in
Settings (Groq is free) covers it.

ClipMint.exe: about 224 MB, Windows, self-contained. Unsigned, so
SmartScreen will warn: More info → Run anyway.

On v1.5.0 or later on Windows, or on any Linux build? The app will offer this
to you.

MIT licensed. Fork it, ship it, build on it. The licence asks one thing: keep
the copyright notice.

Assets 9

Releases: kleZ799/clipmint

Release list

ClipMint v1.26.1

The window no longer opens unstyled

The prompt box reads "hooks" correctly

Uh oh!

ClipMint v1.26.0

Ask for it like a message

Your data

Uh oh!

ClipMint v1.25.0

Font, colour and placement

Uh oh!

ClipMint v1.24.2

The fix

Uh oh!

ClipMint v1.24.1

Hindi, written the way it's read

Uh oh!

ClipMint v1.24.0

How your Shorts did

The ranking learns

Your data

Uh oh!

ClipMint v1.23.0

Frame it yourself

Move the captions

Needs the original video

Uh oh!

ClipMint v1.22.0

It follows the conversation

What doesn't change

Uh oh!

ClipMint v1.21.0

It reads the chat

When there's no chat

Uh oh!

ClipMint v1.20.0

It picks what a stranger would watch

It looks before it cuts

The hook line is on screen

Titles the AI didn't write don't go to YouTube

Uh oh!