Remake Video Visuals

Remake Video Visuals

Reconstruct the visual layer of an existing video without disturbing its audio-led timing. Keep creative decisions flexible; make timing, subtitle rendering, assembly, and validation deterministic.

Default decisions

• Preserve the source audio unless the user explicitly requests audio changes.
• Use the source duration as the master duration.
• Prefer static 16:9 frames and clean cuts for stability. Add motion only when requested and preview it before final rendering.
• Keep generated images free of embedded text; render titles and subtitles separately.
• Render white subtitles with a dark outline and no background box. Use standard at 1080p (56px body), projection-large for classroom projection (72px body), or projection-xlarge when maximum readability is requested (88px body). Scale presets with output height.
• Keep subtitles at bottom by default. Use center only when the user explicitly requests centered captions or when the visual design treats subtitles as the primary focal layer.
• Preserve approved subtitle wording. Reflow lines or align timing without replacing the text with speech-recognition output.
• Apply requested color grading to each unique scene image once before video encoding. Keep subtitles outside the grade so white text stays neutral.
• Preserve user-provided files and write outputs to a separate project folder.

Workflow

1. Inspect the source.
   • Run scripts/probe_video.py SOURCE --output analysis.json.
   • Confirm duration, dimensions, frame rate, audio codec, sample rate, and channel count.
2. Establish the text timeline.
   • Reuse an existing SRT when available.
   • If no transcript exists, obtain one with an available transcription tool; do not invent narration.
   • Verify cue timing against the source narration at the opening, middle, and ending. Do not assume an existing SRT is synchronized.
   • When cues lag or lead the narration, run scripts/align_subtitles.py and review its JSON report. Keep text fixed, use speech recognition only for timestamps, and default to a 0.25-second lead-in.
   • Read references/subtitle-workflow.md when aligning, reflowing, or enlarging subtitles.
   • Split scenes at semantic or emotional turns, not uniform intervals.
3. Create the storyboard.
   • Write scene purpose, start/end time, visual metaphor, composition, mood, and image prompt.
   • Maintain continuity in recurring subjects, palette, lighting, period, and camera language.
   • Read references/storyboard-guidelines.md when designing or generating the frames.
4. Prepare review assets.
   • Generate or collect one image per scene.
   • Create a contact sheet with scripts/create_contact_sheet.py.
   • When the user wants iteration, obtain approval before rendering the full video. When they explicitly request end-to-end production, continue without an extra checkpoint.
5. Create the project manifest.
   • Read references/manifest-schema.md and write project.json with absolute source, image, subtitle, and output paths.
   • Ensure scenes are ordered, contiguous, and cover the source duration.
   • Add color_grade only when the storyboard needs deterministic brightness, saturation, gamma, contrast, or warmth adjustments.
   • Set subtitles.preset to projection-large for course-room projection, or projection-xlarge when the user asks for larger text. Set vertical_position, max_lines, max_width_ratio, and auto_wrap explicitly when layout constraints matter.
6. Assemble the video.
   • Run scripts/check_subtitles.py before assembly. Treat overlaps, invalid ranges, and layout overflow as errors; save the layout report and reflowed SRT when changes occur.
   • Run scripts/assemble_video.py project.json.
   • The script pre-grades unique scene images, scales/crops frames, preserves AAC audio by stream copy, and renders subtitle layers without requiring FFmpeg libass.
   • It renders to a same-filesystem staging directory, fully decodes the staged MP4, and atomically moves it to the final path. Existing outputs are protected; pass --force only after intentionally choosing to replace one.
7. Validate before delivery.
   • Run scripts/validate_video.py SOURCE OUTPUT --decode.
   • Run scripts/create_sync_previews.py after the first synchronized render. Review the opening, ending, low-confidence cues, and cues with the largest timing shifts before accepting the full render.
   • Inspect representative frames: opening title, normal subtitle, every scene transition, and final frame.
   • Confirm sync, duration, resolution, audio presence, subtitle safe area, and absence of black boxes or unintended bars.

Deliverables

Return the final MP4 plus the reusable production package:

• scene images;
• SRT subtitle file;
• subtitle alignment and layout reports;
• short synchronization review clips;
• storyboard/contact sheet;
• project.json manifest;
• optional prompt document;
• validation result.

Use descriptive, versioned filenames. Do not overwrite an approved render unless the user requests replacement.

Script requirements

• Require ffmpeg and ffprobe for analysis, assembly, and validation.
• Require Pillow only for subtitle layers and contact sheets. Prefer an existing bundled Python runtime; if Pillow is unavailable, report the missing dependency rather than silently dropping subtitles.
• Use whisper-cli plus a multilingual Whisper.cpp model for automatic narration alignment. Do not bundle large model files in the skill; reuse a local model or obtain approval before downloading one.
• Pass a CJK-capable font with subtitles.font when rendering Chinese, Japanese, or Korean text.