Skip to content

Releases

v0.9.3

Choose a tag to compare

@github-actions github-actions released this 21 Nov 14:30

Highlights: presenters for the case explainer (ten drawn hosts with their own voices, your own built in the app, or an optional real face that talks); expressive narration through a local voice server, in each section's mood and optionally in your own voice; a Voice-over tab; multi-speaker dubbing; a choice of transitions; larger sequence diagrams; and hooks that are never lost.

  • Case explainer: narrate in your own voice. With the expressive narrator, give a 10-30 second recording (--case voice_sample=PATH, or upload or record it on the Case card and tick that it is your voice or the person agreed): the local voice server copies it for the whole narration, each section still in its mood. It stays on this PC. A redo with another recording speaks the narration again.
  • Case explainer: expressive narration (optional). With --case narrator=voice_server ("Expressive" on the Case card) the narration is spoken by the local Chatterbox voice server set up on the Voice-over card, if it is running: each section in its mood (the script now gives every section one: tense, excited, serious, warm or calm), in the presenter's voice, with captions timed from the audio by Whisper on the CPU. Everything stays on this PC; it uses the GPU while it speaks. Without a running server the standard voice is used, and a server failure mid-way switches the rest to the standard voice.
  • Case explainer: choose the transition between sections (--case transition=number|fade|slide|zoom|wipe|cut, a dropdown on the Case card). To compare, redo an earlier run as a quick draft with another transition: the script and voice are reused, so it takes about a minute.
  • Case explainer: the narration sounds less read-out. The script is asked to be spoken like a presenter talking (short sentences, contractions, "you", a question now and then, terms explained, linked as a story); the free Edge voice reads it at a normal pace with the more expressive voice, without the shorts-style "..." pause after every sentence; Edge gets one retry before another voice takes over.
  • Voice-over in your own voice (--mode VOICEOVER, the "Voice-over in your voice" card in the app): a script (text or a .txt file) is read in a clone of your voice from a 10-30 s recording, with a tone preset (neutral, warm, excited, calm, serious, dramatic, sad) that always changes the pace and sends hints the voice server may use. It needs a local voice server that clones from a sample (VoiceStudio, F5-TTS or XTTS, set as VOICE_STUDIO_URL), so the script and the sample stay on this PC; there is no fallback to a stock voice. The voice server can be your own or an optional add-on installed only on request ("Install the voice server" on the card, or python -m engine.voice_server install; about 5.5 GB, local only). Options: --voiceover KEY=VALUE. Only your own voice, or someone who agreed.
  • Case explainer: the hook is never lost silently. A hook that is missing, too long or holds something the sources do not say is logged with the reason, asked for once more from the checked narration, and otherwise replaced by the narration's first sentence when that is short enough.
  • Dub: several people talking (--dub_speakers off|auto|2-4, the "Several people talking" choice on the Dub tab; off by default). The lines are grouped into up to four voices by pitch, on this PC, and each voice gets its own male or female voice in the target language (a second speaker of the same kind gets the same voice pitched differently); lines of different speakers are never merged. Pitch grouping is a heuristic: check that the voices match the people.
  • Presenters have their own voice: each ready-made one speaks with a voice that fits it (Maya, Priya, Nora, Luna and Miso with female voices; Alex, Sam, Leo, Biscuit and Bolt with male ones), and the builder has a voice picker with a Listen button. The Preview now talks with sound, and the Case card plays the chosen presenter's voice. With ElevenLabs credits a premade voice of the same kind is used; a voice chosen with --voice still wins. Before, the free Edge fallback used a female voice for everyone.
  • Case explainer: sequence diagrams are larger: the participants and the last three messages fill the panel and the view slides down as each message is told (the names travel along); in 9:16 long messages wrap. Measured on a real five-message diagram: the type went from about 28 px to about 44 px in 16:9.
  • Case explainer: number cards only for real figures. A section becomes number cards only when at least two labels hold a figure (money, a percentage, a decimal, a value of 10 or more, or a number with a unit), and never for a numbered list ("1 Law", "2 Algorithms"), which becomes a checklist.
  • Presenters, optional add-on: a real face that talks. Save a photo on the Presenters card (your own face, or someone who agreed) and a full render makes it talk with the narration, on this PC with SadTalker (python -m engine.talking_head install, about 5 GB, NVIDIA GPU; or the Install button in the app). A draft shows the photo still, and so does a run without the add-on. The photo and the voice stay on the PC.
  • Layout: scenes lay out in the room they really have (Flow and Timeline), AutoFit measures content wider than its box, and captions shrink their type to stay within two lines.
  • Case explainer: presenters. A drawn host in the corner who talks with the narration: the mouth follows the words, it blinks, lifts its brows on long words and nods. Ready-made people of different looks (Alex, Maya, Priya, Sam, Leo, Nora), two cats (Miso, Luna), a dog (Biscuit) and a robot (Bolt), or your own: the new Presenters card in the app builds one from parts (skin, hair, eyes, beard, glasses, lashes, colours; fur and pattern for animals) with a talking preview, and can suggest the parts from a photo (a local vision model by default, so the photo stays on the PC; Gemini only if you choose it, which sends the photo to Google). --case avatar=NAME; saved ones live in presenters/ in the workspace. The scenes and the cover keep out of the card in both shapes.

v0.9.2

Choose a tag to compare

@github-actions github-actions released this 20 Nov 22:15
  • MCP server: four new tools. create_explainer (a case explainer from a topic and sources, or one continuous animation), dub_video, list_videos and publish_video (previews by default; it posts only when dry_run is false, to your Upload-Post account or your own YouTube channel). The MCP server could only clip and make idea shorts before.
  • Browser app: a Publish videos card (pick any finished video of any run, filter by tall, wide or square, preview it, edit title, description and tags, publish through Upload-Post or direct to your own YouTube channel, or use a file from this PC), a Settings card (keys, upload defaults, YouTube connect) instead of the accordion, and Publish later: queue a video for a time you type; the app sends it while open.
  • Packaging: the Mermaid checker (remotion-video/scripts/check_mermaid.mjs) is now included in the wheel and the source archive, so a pip install can check diagrams before drawing them.
  • README and demo page: the Highlights video (a long video cut to its best moments in story order) is listed; MCP Docker check expects the 9 tools.
  • README: leads with everything ShortsFlow makes (clips, case explainers, animated explainers, idea-to-video, recaps, dubs, publishing), not only clipping; new case-explainer section.
  • Browser app, Settings: Connect YouTube directly. Choose the Google OAuth client file, press Connect YouTube (the Google sign-in opens in the browser on this PC, once), and publish to your own channel as private, unlisted or public, from the Publish box or automatically after a run; several channel slots; no third party. Real uploads have not been tried with a live Google account, only with fakes.
  • Case explainer: a news template (the headline, what happened in order, who says what side by side, what to watch next; every claim attributed, no side taken); use it with --case theme=newsroom for a news look.
  • Sources: a page that shows no text until its JavaScript runs (news and patch-note sites such as dota2.com) is now opened in the headless Chrome the video renderer already uses, on this PC, and its rendered text is read; this only happens for a page that would otherwise be skipped as empty. Page text is now read as statements and bullets with menus, language pickers, version lists and footers dropped, so the first part of a long page is no longer navigation.
  • Mermaid: a diagram the model wraps in JSON is unwrapped (it was missed before, and the section silently fell back to Manim); a fallback now says why in the log; sequence diagrams reveal one message at a time with the participants always visible, and their type size is set at render time.
  • Case explainer: Mermaid diagrams. The model can draw a section as any Mermaid diagram type; it is parsed by Mermaid itself (in Node, no browser) with a 6-node limit and held to the sources, a problem goes back to the model once, and a diagram that still fails falls back to the plain flow. Mermaid places everything, so nothing can overlap; the video draws it in the glass panel with large type, steps revealed in order, the current one glowing, and a camera that follows chains longer than four steps. The "how it happened" sections of the case, postmortem and explainer templates are diagrams by default. Manim is unchanged. --mode CASE review text has a DIAGRAM: block.
  • Case video layout: every scene is measured and fitted into the room between its heading and the caption bar (AutoFit), and the caption zone is one shared definition (captionZone), so a long caption or long labels can no longer overlap a scene whatever the topic; timeline labels grow away from the line; landscape animations use a wider panel and a two-row chain; a hook line on the cover. scripts/layout_stress.py writes a worst-case script to check a layout change against.

v0.9.1

Choose a tag to compare

@github-actions github-actions released this 20 Nov 06:57

Highlights: a Settings page in the browser app (keys, publishing), the animated explainer mode --mode ANIMATE, a recipe template, research papers and PDFs read properly, and a local-model fallback when the cloud text model is rate-limited.

  • Browser app: a Settings section. Add, replace or remove API keys (stored in .env, only the last four characters shown, switched off when the app is open to other devices); choose upload platforms (through Upload-Post); publish a finished case video or clip set now, with a preview that sends nothing; optional automatic publishing after a run (off by default, never twice to the same platform). New shortsflow/settings.py.

  • Case explainer: new recipe template (the dish, the steps, the numbers, the tips). Counters keep a unit with its number (130°C, 20 g, 3 cm) and show a range whole (5–7); at most four counters and three checklist rows per section, so they clear the caption bar; four counters in landscape are drawn smaller, in one row. A PDF link without .pdf (arXiv) is read as a PDF; a decimal a PDF splits (2 .14) is joined; 3 x 10^9 in a label reads 3 billion; results sections ask for outcomes, not setup; the prompt bans author lists. Manim: the overlap check also finds a line, arrow or dot drawn through a text; retries after known crashes are told the fix; colour aliases in the kit.

  • Case and animate: when the cloud text model is rate-limited or down (Gemini 429), the script and the animation code are asked of the local Ollama model (LOCAL_LLM_MODEL) with a long timeout, and the cloud is skipped for five minutes; --case llm=auto|local|cloud (local keeps everything on this PC). New engine/case_llm.py.

  • New mode --mode ANIMATE: a topic (and optional --sources) in, ONE continuous Manim animation under the whole narration out, in 16:9 and 9:16, for equations, processes and diagrams. It is the case pipeline with the new animation template (one section, a 10 to 14 sentence narration, the animation fills the frame with no heading), so the narrator, captions, music, nine looks, --case options, plan_only, resume and draft all work; --mode EXPLAINER is unchanged. The animation is built from the Manim kit and checked for overlaps before it is rendered.

  • Animations: each failed try is logged with its reason; a scene that fails only the picture checks (overlap, cut off) on the last try is rendered anyway with a warning, a crash never is; a run whose animation section ends with no animation stops with a clear message (the script and narration are kept: --case resume=FOLDER) instead of making a title card; the kit helpers also work as scene methods (self.play_chain(group)); the kit refuses a Manim object as a label; the prompt for a whole-frame animation asks for real Manim shapes for geometric or numeric ideas; the script step asks the model once more when its answer is not a usable script.

  • Browser app: a Layouts control (auto or template) and the animation choice in the Case tab.

  • A test keeps AGENTS.md true: every file it names exists, its Python version matches pyproject.toml, its LLM fallback order matches the code, and a new CLI mode cannot ship undocumented (24 older modes are pinned as known documentation debt that can only shrink).

  • --mode CASE layouts: two new scenes, compare (two columns, left | right rows under a heading row) and chart (bars drawn to scale from Name: number labels, numbers counting up), and the model now picks each section's layout itself (--case layouts=auto, the default; layouts=template keeps the template's). It chooses from timeline, flow, counters, checklist, compare, chart and quote and never repeats one back to back; terminal and formula stay with templates that ask for them. The SOURCE chip no longer sits on the title in portrait.

v0.9.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 17:34

New: the case explainer (--mode CASE). A topic and up to three sources in, a finished sectioned explainer video out, in 16:9 and 9:16. It fits any subject that can be told in sections (an event, a failure, a product, a scam, a historic moment).

  • Grounded. The script is written only from the sources. A sentence holding a figure or name no source contains is dropped; with no source the video says UNVERIFIED. Every section shows a small "SOURCE host" chip and the end card says how many sentences were left out.
  • Sources can be web pages, PDFs (link or file, needs pypdf: extra sources), YouTube links (captions; with none, the audio is listened to), audio or video files (listened to by the local speech model, first 40 minutes), caption files (.srt, .vtt), text files or pasted notes.
  • Made for you: a narrator voice (ElevenLabs when its key is set, else Edge; after one failure the rest of the video uses Edge), captions, synthesised light music and sound effects (nothing downloaded, nothing licensed), drawn scenes that light up as each item is spoken (timeline, flow, counters, checklist, terminal, quote, formula), optional stock video or photos behind the scenes (PEXELS_API_KEY) or your own image or video, and a Manim animation of the mechanism, floating over the picture.
  • Pacing: each section is cut into 3.5 to 6.5 second beats at the narrator's sentence ends; every beat has its own picture, a punch-in and a flash.
  • Nine looks (theme): dark, clean, warm, neon, newsroom, blueprint, terminal, editorial, midnight. Each sets colours, the animated background, card style and type (system fonts only).
  • Options in one schema (engine/case_options.py, built on the reusable engine/mode_options.py): --case KEY=VALUE (repeatable; --case help lists them) or --case_file FILE.json. Keys: template, theme, formats, background, manim, draft, plan_only, resume. The house rule for other modes is in AGENTS.md.
  • Review and iterate: plan_only writes just the script so you can edit it; resume=FOLDER redoes the render with the saved script, narration and animations (new look, background or shape, no new voice); draft renders a half-size draft in about a minute.
  • Animations are checked before they are rendered: built from a tested kit (engine/manim_kit.py: chain, branch, cycle, before and after, bar), run once at a tiny size to find texts that overlap or cross a box edge, and checked on sample frames (cut off, nearly empty, cramped); a failed check goes back to the model with the reason.
  • Browser app: a Case explainer card (topic, sources, sections, look, shapes, background, quick draft, earlier runs to redo) with a review box to edit the script as plain text before rendering.
  • See docs/FEATURES.md ("Case explainers").

Clip tool

  • Review step: below the chosen moments, up to three runner-ups the tool ranked next are shown unticked; tick one to render it too (plan.json marks them "spare": true).
  • A clip run saves every moment that was in the running, best first (candidates.json). With --gold, scripts/clip_benchmark.py adds recall@5 and recall@8.
  • scripts/picker_bench.py: an offline benchmark of the moment picker on the transcripts of the four public test videos (seconds, no video, no AI); --ablate shows what each picking rule is worth.
  • Experimental, off by default: SHORTSFLOW_LOCAL_RERANK=1 has a model running on this PC (Ollama, LOCAL_LLM_MODEL) rank the best 15 candidates; there is no cloud fallback for the text. SHORTSFLOW_SENTENCE_WINDOWS=1 picks no-key shorts as sentence windows; it does not beat the current picker on the benchmark yet.

Other

  • shortsflow doctor --report prints a block to paste into a bug report (versions and the doctor checks, no keys, no folder or account names).
  • Python 3.10 is tested in CI; requirements-full.txt allows Gradio 6 like pyproject.toml; runs started by the MCP server and the API open no console window on Windows.
  • engine/modes/clip_extract.py: run() is a short orchestrator calling named steps. No behaviour change.

v0.8.1

Choose a tag to compare

@github-actions github-actions released this 16 Nov 17:35
  • Recap made like a reference (--recap_like, still experimental): copies a reference's colour look, widescreen shape and landscape clips, the film-style look (subtitle text, hook, end title), its shot length and number of places, and a headline held over the whole recap when the reference does that. No black dip on fast cuts. Scenery ratings are remembered by picture, so the same footage is never rated twice.
  • Recap picking: itinerary and held-headline recaps rank down frames with lens flare or a blown-out patch and windows where the camera is turning or shaking.
  • Recap: itinerary labels drop a leading "Day N:" and administrative-area names; the local rotation check no longer turns a clip upside down because of a pale strip.

v0.8.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 21:13

Experimental: make a recap like a reference reel. --recap_like REFERENCE (a link such as YouTube, or a video file) and the "Make the whole recap like this reel" box on the Recap screen read a reference reel's pace, picture shape, text layout (a "DAY N" label over each shot, a headline opening, the ending), label colour, size and lettering, and set the recap's options for you (anything you set yourself wins). It has been tried on one reference, so expect to adjust. To read the text it needs a picture reader: Gemini with your key (the reel's frames are sent to Google) or a local Ollama vision model; without one only pace, shape and length are copied. It approximates the look with the fonts that ship with ShortsFlow (it cannot copy an exact font) and never reuses the reference's footage, music or graphics. New options it uses: --recap_itinerary (an itinerary reel: "DAY 2" above the place name, straight cuts, a tiered opening headline, a soft dark band behind labels), --recap_text_style (colour, size, bold, glow), --recap_font display (a heavy display serif, Abril Fatface, SIL Open Font Licence) and --layout crop for recaps.

  • Recap: "Let me review the shots first" (--recap_plan_only, then --recap_moments): the shot picked for each day is shown with a preview in the final shape; swap it for an alternative, trim it or drop it before anything is rendered. An itinerary reel picks each day's most scenic shot.

  • Style cloning (--clone_style, "Copy the style of a reel"): the reference can be a link as well as a file.

  • Recap and Clip tool Auto-edit lists include the templates saved or cloned since the app started; saved templates keep their text boxes; the browser app can join the shorts into one video (--join_reel).

  • Failed runs now say the likely cause in plain words (FFmpeg or Node missing, graphics card busy, speech model download, full disk, blocked link).

  • Facts videos no longer repeat: the last 40 hooks and facts go into the prompt as "do not repeat" (history kept between daily runs by the workflow cache).

  • Recap review: rendering the reviewed shots no longer picks shots again (it was slow and used the vision model), and the reviewed files match whatever the slash style. FFmpeg and the other helper programs the recap starts no longer flash a console window on Windows.

  • The transcript cache keeps the newest 200 transcripts (older cache files are removed). Fixed: Stop then Continue re-cut every piece; a reviewed plan lost its titles and warnings at render; the continue-a-run list only accepts a run's own command.json inside sessions/; the cut-piece cache key now covers grade, HQ, superres, filler file and crop points, and finished clips have a key too.

  • Style cloning: --clone_style REFERENCE --save_as NAME (and "Copy the style of a reel" on Start here) measures a reference reel on this PC (shot length from the cuts, colour, loudness) and saves a template; pauses, filler cuts, punch-in zooms and cutaway pictures are inferred from the pace and the summary says which parts were measured and which were guessed. Captions, graphics, music and the reel's footage are not copied. Use it with --auto_edit NAME.

  • Browser app: "Save these settings as my template" on Start here keeps the chosen Auto-edit style, caption style, extras, colour grade, visual template and covers under a name (a JSON file in ~/.shortsflow/templates/); it shows up in the Auto-edit list straight away and as --auto_edit NAME on the command line.

  • Seven new caption styles (--caption_style): MARKER (highlighter sweep), SOFT (clean words fading in), WORDCUT (words cut in), NEGATIVE (keyword in a muted tone), HANDWRITTEN (script with an underline), SIMPLE (tall plain capitals) and TYPEWRITER (typed letter by letter). Each has a preview picture in the browser app's picker. Fonts Caveat and JetBrains Mono (SIL OFL) are bundled.

  • Known issue, not fixed: the bundled caption fonts (Anton, Noto Sans for Cyrillic, Devanagari and Arabic) are never loaded, because remotion-video/package.json tells the bundler that only CSS files have side effects and it drops import "./fonts". Captions use the machine's own fonts. Turning the import on stalled a browser tab in the E2E render on GitHub's runner, so it is left off until it can be tested there. Caveat and JetBrains Mono (the two new styles) are loaded only when their style is used.

  • Fixed: the letterbox check could report an impossible crop window (for example negative sizes on a video that starts black), which made FFmpeg fail on every clip that used the plain face crop: those clips were dropped silently. Such a window is now ignored, a clip cut that fails on the GPU encoder is retried once on the CPU, and any cut that still fails is reported.

  • Browser app redesign: one "What do you want to make?" chooser with icon cards (Shorts, Highlights, Dub, Recap, Photo reel, Idea, Clip tool) instead of tabs, a video-kind picker that sets the moment picking, edit style, framing and captions in one choice, a plain "Picture in the tall frame" choice (automatic, crop, fit), everything else under one Advanced section, a progress view (stages, bar, listening and rendering percent, elapsed time), "Your projects" tiles with thumbnails of earlier runs, and a primary button that stays in view. Checked at phone width, in light and dark mode, with keyboard-only use and for text contrast.

  • Review the moments first: --plan_only writes the chosen moments to plan.json without rendering, and the browser app shows them with a playable preview in the shape of the final short, a timeline, nudges, your own extra moments and a Render button. --only_between START-END limits the search to a part of a long video.

  • --auto_edit auto picks the edit style from the file name, --user_context and the video length, and prints why and the runner-up.

  • Videos and photos already on this PC are used in place: the link boxes accept a file path, Recap accepts folder paths (each day folder stays a day), Photo reel accepts a photo folder, and Browse buttons open the system file or folder dialog. When the app is opened from another device (--host 0.0.0.0 or --share), paths and Browse dialogs are switched off for visitors; uploads and web links still work.

  • Recaps: --recap_font handwritten, and a visual template with intro and outro covers laid over every finished recap (portrait and landscape); the template text boxes are worded for a trip.

  • Game footage: on Automatic the review shows the picture the tool will use and the render uses the same decision.

  • Fixed: a cut piece (temp_segments/seg_N.mp4) was reused by its number, so a rerun with other moments or another framing could use old footage (wrong clip, old bars). A piece is now reused only when its source, moment and settings match; a finished clip only when it is newer than its piece.

  • Fixed: ticking "game footage" forced the fit layout, so "Fill the screen (crop)" was ignored.

  • The transcript of a video is kept once on this PC (~/.shortsflow/cache/transcripts, SHORTSFLOW_TRANSCRIPT_CACHE=off turns it off), so the same video is not transcribed again in every new run.

  • Fixed: a Gemini model that did not answer in time was retried four times on each of four models (up to 300 s per try), which could hold a run for tens of minutes. A timeout now moves on to the next model, and the optional video outline waits at most 40 s per try.

  • Game footage with --visual_context: a moment that looks like real filmed footage (a just-chatting part of a stream recording, for example) is flagged in the plan and on the review card instead of passing as a story moment. It is only flagged, never removed.

  • --join_reel (shorts): after the clips are made, they are also joined into one video, reel_joined.mp4 in the session folder, with crossfades or the look chosen with --transitions. The shorts stay as they are. This is the vertical counterpart of the highlights video, which was landscape only.

  • Browser app: the Stop button only shows while a run is going, and light mode has a warm page behind white cards so the sections stand out (dark mode unchanged).

  • A project logo (one wide bar over three tall bars: a long video becoming shorts) in the browser app header and tab icon, the README, the website and a social preview card (assets/branding/).

ShortsFlow 0.7.8

Choose a tag to compare

@ghost1412 ghost1412 released this 13 Nov 13:00
  • Scheduled runs (the daily channel workflow) pick their mode by a weighted rotation, not by dice: every 4-hour slot gets the next mode, the weights are met exactly and the same mode never comes twice in a row. Random draws had given streaks (three "spot the lie" runs in a row). SCHEDULED_MODE still sets a fixed mode or your own weights.
  • Clip benchmark: recall and precision against a gold set now match a clip and a moment when they overlap by half of the shorter of the two. The old rule asked a 30 s clip to cover half of a moment that can be 60 s long, so the numbers were capped by the moment lengths (a pick starting at the same second as a moment scored 0).
  • Groq: a "429 Too Many Requests" (the per-minute token allowance of a free key) no longer switches the AI picking off for the whole run. The call now waits for the allowance (Retry-After, at most 65 s, four tries) and asks again; a request above a model's limit moves to the next model instead of waiting.
  • Clips can now be 1:1, 4:5 or 16:9 as well as 9:16 (--aspect_ratio, the Output shape in the browser app): the crop, layouts, captions, templates and covers follow the frame. It used to print "not implemented" and make 9:16.
  • Browser app: a "Look: template and covers" section on Start here and on the Clip tab (visual template with a preview picture, intro and outro covers, text for the template). It works for shorts and for the highlights video.
  • Highlights video (long, landscape): --visual_template, --intro_animation and --outro_animation now work on it. Short transparent pieces for the start and end are rendered with the animated renderer and laid over the video with FFmpeg, so a long video is never rendered frame by frame in the browser.
  • --segment_transition fade|slide|wipe|flip|clock-wipe|iris joins the background clips or photos of photo reels and idea-to-video modes with Remotion's @remotion/transitions (pinned 4.0.509, like the other Remotion packages); the default stays the crossfade.
  • Highlights video: eight more FFmpeg transitions between moments (--transitions iris|blinds|panels|wipe|radial|slice|pixelize|squeeze); docs/TEMPLATES.md now lists what the highlights video supports.
  • Intro and outro transitions: --intro_animation / --outro_animation (tiles that break apart and join, panels, iris, blinds, fade) with --animation_color; also settable per clip template. Previews in docs/templates/.
  • Visual templates: --visual_template (10 designs: podcast-card, neon-frame, news-banner, minimal-line, title-bars, retro-vhs, film-bars, polaroid, episode-badge, story-progress) draws a designed skin over the picture, with optional text slots (--template_text KEY=VALUE) and an accent colour. Re-renders keep it. Previews in docs/templates/.
  • Clip templates: six new built-in --auto_edit styles (minimal, news, storytime, motivation, tutorial, cinematic) and your own as JSON files in templates/ or ~/.shortsflow/templates/ (a template can start from another with based_on; a bad file is skipped with a warning). They show up in the command line and in the browser app's Auto-edit list. See docs/TEMPLATES.md.
  • Removed the hosted-dashboard job reporting (Supabase job status, per-user YouTube credentials, --user_id, decrypt_secret, the cryptography dependency). --video_id stays as an optional session-folder name. Supabase storage is still used only for Instagram Reel hosting (falls back to file.io).
  • The hook-cue bonus needs a short number, price or percent (a year no longer counts); "No. 5" is not a sentence end; a closing curly quote after "..." is recognised; a sliver at the end of a pause no longer cuts a quiet stretch short.
  • Browser app: a clip picked by the failsafe (analysis failed, fixed window) shows a warning on its card; fallback: true is written to the results file.

v0.7.7

Choose a tag to compare

@ghost1412 ghost1412 released this 12 Nov 19:52
  • Without an AI key, windows full of greetings, thanks and sign-offs ("welcome to", "thanks for joining", "subscribe", "voice check") rank lower, so intro and outro talk is picked less often as a clip.
  • A weaker overlapping clip restarts at a sentence start or is dropped; it never opens mid-sentence. "Plan A." / "vitamin A." no longer count as trailing off.
  • The loud-pause check covers the tail of a pause and the log counts seconds kept.
  • Sentence splitting without word timings handles closing quotes; the hook cue stops at the first real sentence (not after "Dr." / "No.").
  • Opening rules treat a window as English only when it is mostly English (a Hindi window with a few English words no longer gets the English bonus).
  • Clip benchmark: warns when the gold file has no entry for a run's video; the workflow fails when the gold file is missing; model-labelled gold sets for the four public videos are in docs/gold/.
  • scripts/gold_sheet.py: turns a session's transcript into a time-stamped labelling sheet and your hand-picked moments (12:30-13:05) into the gold file the clip benchmark reads (--gold, recall and precision); the benchmark workflow takes an optional gold file.
  • A clip is no longer shortened because its last sentence is complete: only Whisper's "..." after a word that needs a continuation, or an ending on "a / the / and / because", counts as trailing off ("Look at this." and "...what I am most proud of." stay).
  • --tighten keeps laughter and applause inside a long pause: the pause is checked in 0.1 s windows and only the quiet stretches are cut (it used to average the whole gap). Without the audio file it says so and cuts by timing.
  • When two clips overlap and the weaker one is too short to give way, it is trimmed or dropped, never the better clip.
  • Japanese and Hindi text without word timings splits into sentences at 。!?।.
  • The hook cue is stricter: the first sentence only, and "you" / "what" no longer count. self-cont no longer flags "This is ..." and now catches "He's / It's ..."; clean end uses the same evidence rule.
  • README and browser app: a key also lets the AI pick and score the moments (the transcript goes to the provider), and the footer lists the OpenStreetMap lookup. Removed the leftover dashboard setup guide, database schema and webhook default.
  • docs/BENCHMARKS.md: the clip quality results on four public videos and a highlights video, measured on GitHub's runners with no AI key, with the weak spots listed (hook cue, low-resolution sources, slides, no comparison with Opus Clip yet).
  • Without an AI key, the picker prefers a window that opens with a hook (a question, a number, or a word that speaks to the viewer: "you", "never", "why" ...) over an equally strong one that does not, so more clips start on a reason to keep watching.
  • The repository is now ghost1412/shortsflow (was shorts-generator). GitHub redirects the old address, including git clone, the raw install one-liner and release downloads; the demo page moved to https://ghost1412.github.io/shortsflow/ (the old page address is gone). The package, the installer and the commands are unchanged.

v0.7.6

Choose a tag to compare

@ghost1412 ghost1412 released this 12 Nov 12:39
  • Without an AI key, a highlights video (--extract_mode long) fills the asked length again: when no provider answered, one failed call returned a single fixed 30 s slot (a 33 s reel for a 180 s target); now the signal picks fill it (176 s for 180 s on the NASA sample).
  • Clips and highlights skip a weak opening: a mic check ("testing, one two three") or a sentence that leans on the one before it ("And, and it just goes ...", "Uh, ...") is skipped at the start (up to 20 s, never below the minimum length), and the no-AI picker ranks such windows lower.
  • Without an AI key, a title never ends on a trailing-off fragment ("... places that.", "... data about."); the next complete sentence is used.
  • --layout auto fits the whole picture (centred over a blurred copy) when no face shows on screen in a clip, as in slides or a screen recording, instead of cropping the content off the sides. Explicit layouts are unchanged.
  • Review fixes: "So I quit my job ..." / "That is the biggest mistake ..." are no longer treated as weak openings (only connectives like "And", "But", "Because", filler sounds, and "So,"/"Well," with a comma are); the opening skip no longer adds up when a clip is snapped twice and never gives up more than 40% of a clip; when two clips overlap the lower-scoring one gives way (the better earlier clip kept losing its payoff sentence); wide shots of people are not mistaken for slides (a smaller face counts in the source check); any unexpected error in the moment picking now falls back to the signal picks instead of one fixed 30 s clip; Hindi/Japanese/other non-English text is no longer penalised by the English opener rules, and Japanese/Hindi sentence ends (。!?।) count.
  • --tighten keeps loud pauses (laughter, applause, a reaction after a punchline): a gap between words is cut only when it is clearly quieter than the speech; quiet gaps and gaps under steady background noise are still cut.
  • Clip benchmark: len/target (share of clips within 25% of the asked length; independent of the sentence rules).
  • Without an AI key, highlights-video chapter titles no longer come out as "I'm, Back, You're": contractions and weak words are skipped, so a chapter gets its content words ("Bone, Density, Loss"). The clip benchmark no longer grades the chapter list (timestamps) of a description as claims.
  • A clip no longer ends on a sentence that trails off on a word needing a continuation ("... the places that."): it ends on the sentence before, if that costs under 30% of the clip.
  • Clip benchmark: cue (a question, number or direct address in the first 3 s), self-cont (the first word is not an unresolved "It/This/They ...") and clean end (the last sentence does not trail off).
  • README: a short overview around the podcast use case (install, what you get, how it is measured, privacy, comparison); the full reference of every mode, option, API and recipe moved unchanged to docs/FEATURES.md. The README recipes no longer use flags that are not implemented.
  • Repository clean-up: the old hosted dashboard (web/, server.py and its test, the seed script), a stray notes file and an unrelated demo clip were removed; mcp_server.py (the MCP server) is unchanged.
  • Privacy wording: --visual_context says the frames go to Google when a Gemini key is set; the browser app footer and README list what leaves the PC when you switch it on; the auto layout description mentions the fit case.
  • Slides and screen recordings that take the fit layout under --layout auto also zoom into the part of the picture that changes (browser tabs and empty margins stay out; at most about 1.7x, same 16:9 window, so the captions stay below the picture). Conservative: only when a content area of at least a quarter of the frame keeps changing; a cursor, sidebar, webcam inset or a screen where everything changes is left as a plain fit.
  • Clip benchmark: hook ok (does a clip open on a line that stands alone), reel placeholder titles ("Clip #N") are not counted as ungrounded, and the workflow uploads one sample frame per clip.
  • Browser app: the timeline strip in "Edit this short" has drag handles. Drag the blue start or end edge; letting go fills the Start/End box (1 s minimum between them) and redraws the strip.
  • Without an AI key: picked shorts are at least 30 s apart (the top-scoring windows could be neighbours, one scene cut in two).
  • --tighten works for every short again: the shorts are cut in parallel and shared one filter-script file, so ffmpeg read two plans mixed together and most shorts were silently left untightened (4 of 5 in a podcast run). Each short now has its own.
  • --tighten never cuts a spoken word: on stage recordings (a keynote with applause) the loudness rule took 6-17 s of speech out of every clip, mid-sentence. Now every transcribed word is kept and only what lies between words is cut (speech lost on that keynote: 0).
  • --tighten also cuts pauses between spoken words (0.8-4 s, a margin kept), so pauses under steady background noise are removed too; loudness alone left them in. On the NASA sample: silent share 4% -> 0%, a 25 s clip lost 3.7 s of pauses. Not applied to game footage; longer gaps (music, action) are kept.
  • Clip benchmark: gaps s (pauses between words in the picked span) and cut s (what tightening removed, from the finished files).