Skip to content

Chopify v1.4.0 - Studio, captions, and any-length sources

Latest

Choose a tag to compare

@mehbul mehbul released this 05 Nov 14:36
· 5 commits to main since this release

Chopify v1.4.0 is the biggest release so far: a complete local Studio, Caption Studio, resumable any-length sources, and fully local audio/visual clip discovery. No paid APIs, uploads, or watermarks.

Install in one command

Windows (PowerShell)

irm https://raw.githubusercontent.com/mehbul/chopify/main/install.ps1 | iex

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/mehbul/chopify/main/install.sh | bash

Highlights

  • Local browser Studio for queueing, review, caption editing, previews, and exports.
  • Caption Studio with 13 presets, full typography control, emphasis, emoji, and a source-synced preview.
  • Any-length sources with resumable downloads, transcription, discovery, and section rendering.
  • Fully local audio and visual evidence for podcasts, streams, silent gameplay, and action-only footage.
  • Resumable projects, batch queue, semantic search, podcast split-screen, SRT editing, and brand presets.
  • One-command installers, a contributor guide, issue templates, and a 1280x640 social preview.

Full changelog

Added

  • Local browser dashboard with project review, progress, logs, previews, caption
    editing, candidate selection and downloadable exports.
  • Caption Studio with 13 presets, full typography/layout controls, word motion,
    highlighting, boxed backgrounds, and a source-synced overlay with 5-second
    preview.
  • Automatic per-cue emphasis for numbers, money, ALL-CAPS and power words, plus
    optional deterministic emoji, capped at two emphasis words and one emoji per cue.
  • Reusable caption presets through brand JSON; project caption overrides persist
    in project.json and layer over preset and brand values.
  • Shared caption CLI flags and --caption-json for chopify.py and
    render_clips.py.
  • Optional hidden and emphasis flags on transcript words; both burnt captions
    and SRT honor them.
  • Direct local video input (MP4, MOV, MKV, WebM, M4V, AVI).
  • Resumable projects with stage invalidation, per-clip checkpoints, and distinct
    output names for render variants; partial failures retain successful clips.
  • Automatic long-source path: sources of 2 hours or more, 8 GB estimated download,
    or unknown duration use an audio-first probe and resumable ingest. Short
    sources keep the original single-pass workflow.
  • Long URLs download audio only, or lowest-resolution video when there is no
    audio; an audio track with no speech also switches to the low-resolution video
    so silent gameplay and streams can score from visual evidence. Local long files
    are read in place.
  • Long-source Studio review: project detail exposes source metadata, progress,
    per-clip render errors, and whether the full source is playable. Watch on a
    long URL candidate queues a 480p preview from its downloaded section and plays
    the result.
  • Resumable 30-minute transcription chunks with 30-second overlap merging and
    global word times (transcript.partNNN.json).
  • Resumable 30-minute discovery blocks with max_len boundary context
    (candidates.partNNN.json).
  • Selected-range section rendering for long URLs: 10-second padding, 30-second
    merge gap, 10-minute cap, source_offset global-to-local time mapping, and
    automatic cleanup after a successful final render.
  • Per-clip section and render errors (render_errors) keep successful outputs
    and cached sections for retry.
  • Exhaustive sentence-seeded discovery with an AI-chosen duration ladder,
    candidates.json / selection.json diagnostics, and paginated Studio review.
  • Fully local audio-evidence timeline with cached 1 Hz features, content
    profiles, adaptive evidence weights, and a 7.0 text floor for speech-sparse
    windows.
  • Action-only and speech-sparse clip discovery using audio event seeds, grid
    seeds for varying tracks without peaks, and a clear failure for flat sources.
  • Fully local visual evidence from cached 2 fps analysis: motion spikes,
    scene/UI changes, facecam presence, action arcs, and boundary quiet, plus a
    sparse whole-source baseline.
  • Silent and no-commentary gameplay, PC-game, and stream support: flat audio is
    treated as unavailable so visual action can score on its own.
  • Visual shortlisting before frame decoding, visual-only seeds for silent
    footage, a 6.0 cap for flat action windows, and evidence-checked boundary
    refinement.
  • Visual diagnostics in candidates.json and selection.json:
    visual_baseline, visual_ranges, visual_sections, and warnings.
  • evidence fields in candidates.json and segments.json, plus
    evidence_available and profile metadata in selection.json.
  • Guided CLI prompts for minimum score and preferred length, plus --no-input
    and --explain for scripts and review.
  • Serial batch queue with isolated folders and aggregate failure reporting.
  • Word/timestamp editing and clip-relative SRT, including retimed tightened cuts.
  • Validated reusable JSON brand presets for fonts, colors, size, margins and logo.
    Brand files now accept every Caption Studio field.
  • Semantic clip search with optional local Ollama embeddings and document caching.
  • Vertical podcast split-screen with two adjustable fixed crop centers.
  • Workflow, troubleshooting, local API, and machine-readable product documentation.
  • Static documentation website with canonical/social metadata, JSON-LD, sitemap
    and a GitHub Pages deployment workflow.
  • One-command installers for Windows and macOS/Linux, a contributor guide,
    issue templates, and a 1280x640 social preview for the documentation site.

Changed

  • Default minimum clip score is now 8.0 (was 6.5 heuristic / 8.0 LLM).
  • Long sources select the audio-first, chunked, block-scanned, section-rendered
    path automatically; there is no duration ceiling or mode toggle.
  • Render cache keys include the section offset and section identity, so long
    project retries reuse finished clips without re-downloading their sections.
  • Every clip at or above the threshold is returned by default; the ten-clip cap
    is now opt-in through --max-clips.
  • Clip length is selected from a duration ladder; --preferred-len is a soft
    preference rather than a fixed target.
  • Default clip range is 15-300 seconds (was 20-75).
  • Audio evidence adjusts the calibrated text score through the source profile
    and evidence weight; content profiles never change the user-facing threshold.
  • Combined audio and visual evidence uses 60/40 weighting; one modality is
    shrunk toward neutral, and text remains the final fallback.
  • Discovery cache keys include the visual feature version, so an analysis
    upgrade invalidates only affected work.
  • Discovery and selection are cached separately, so changing the score
    re-filters without rescanning the transcript.
  • Matching successful stages resume by default; --no-resume forces processing.
  • Caption style resolution is named preset, then brand preset, then per-project
    override; aspect defaults fill unset font size and margin.
  • SRT cue grouping follows the caption style's words-per-cue value on project
    renders.
  • Docs clarify setup downloads, fixed framing, language limits and score meaning.

Fixed

  • Overlapping candidate windows are no longer blocked by marking covered
    sentences as taken.
  • Zero passing clips now writes an empty segment list instead of a fallback clip.
  • LLM reranking now adjusts the evidence-aware final score instead of the raw
    text score, and no longer compounds an earlier adjustment on rerun.
  • Split-screen draft panels use even YUV420 dimensions to avoid crashes in older
    ffmpeg/libass builds; padding preserves the intended preview resolution.
  • Changed URL downloads cannot reuse an older source file.
  • LLM prompt definitions are initialized before standalone scorer execution.
  • Render failures return nonzero instead of reporting success.
  • Relative output paths resolve before ffmpeg changes its working directory.
  • Identical clip titles cannot overwrite each other's project exports.
  • A failed long-source section records a per-clip error instead of discarding
    successful renders; remaining sections still render and retry later.
  • Long-source preview playback no longer draws the source caption overlay over
    the already-rendered draft.
  • A stalled long-source section download falls back to a full-source download
    and continues rendering with source_offset=0.