whiteboard-video

name whiteboard-video description Generate smooth hand-drawn whiteboard and story videos directly inside Codex from text scripts, GPT Image 2 color storyboards, scene plans, SVGs, line art, or local images. Supports 30 built-in visual styles, automatic recommendations, custom media recipes, natural-block coarse/detail drawing with block-local or whole-scene color reveal, local image-to-line-art extraction, stroke/path drawing, hand/pen-tip following, narration/TTS including Doubao Voice 2, and FFmpeg composition. Whiteboard Video Use this skill to create local-first whiteboard videos without Canva. This skill is an adapter for the separately installed whiteboard-video-engine Python package. Uploaded photos and dense illustrations are converted to line art locally; do not use GPT Image 2 to generate final line art. For script-driven story videos, keep one registered asset pair per scene: GPT Image 2 creates the color storyboard , then a local neural extractor derives line art from that exact color file. The preferred stack is: gpt-image-2 for full-color storyboard frames only, never production line art or embedded captions. Informative Drawings local model, preferably anime_style , for registered line-art extraction. Anime2Sketch local model for illustration/anime-like sources or as the neural fallback. Optional vtracer SVG vectorization when installed. Optional Doubao Voice 2 ( seed-tts-2.0 ) or Edge TTS for narration. Keep the default sidecar SRT for editing, or explicitly use --burn-subtitles for a ready-to-publish subtitled MP4. There is no edge-detection fallback. If neither neural model is installed, extract-lineart and render-photo must fail instead of silently producing a weak outline. Quick Start Install the engine before using this skill: python3 -m pip install "git+https://github.com/gnipbao/whiteboard-video-engine.git" For local engine development: python3 -m pip install -e /path/to/whiteboard-video-engine When the engine is installed in a virtual environment, invoke the wrapper with that environment's Python, or set WHITEBOARD_ENGINE_PYTHON=/absolute/path/to/venv/bin/python . The wrapper will re-exec only the explicitly configured interpreter. Run commands from the project root that contains tools/lineart , but always call the installed Skill wrapper by absolute path. The wrapper delegates to the installed engine package while local model wrappers are auto-discovered from the current working directory. Never call a project-local whiteboard-video/scripts/whiteboard_cli.py . Old project copies may prepend a bundled src directory and silently shadow the installed engine, causing stale defaults such as the procedural hand cursor to reappear. MOCK=1 python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" run examples/ten-second-demo.md -o /tmp/whiteboard-demo.mp4 --scenes 2 --fps 30 --width 640 --height 360 --scene-assets direct-lineart --animation-preset block-speedpaint python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" extract-lineart photo.png -o lineart.png --provider auto python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" render-photo photo.png -o /tmp/photo-whiteboard.mp4 --duration 15 --fps 30 --lineart-provider auto --stroke-detail rich --hand asian python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" render-image lineart.png --source-image photo.png --source-fit exact --size-from-image --color-fill contour-wipe -o /tmp/color-fill-whiteboard.mp4 --duration 15 --fps 30 --tail-color 4.5 --hand asian python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" render-image examples/apple.svg -o /tmp/apple-whiteboard.mp4 --duration 2 --fps 24 --width 640 --height 360 --hand asian python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" render-image examples/apple.svg -o /tmp/multiline-text.mp4 --duration 6 --fps 24 --width 720 --height 960 --draw-text-file caption.txt --draw-text-position top --draw-text-align left --draw-text-reveal line-wipe --draw-text-order before --hand none python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" list-styles python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" recommend-styles story.md -- limit 5 python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" list-hands python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" doctor Visual Style Selection The engine provides 30 versioned, media-named visual recipes. The stable default is warm-crayon-storybook , also configurable with WHITEBOARD_STYLE . Style selection belongs to plan-script and run : planning uses the semantic visual guidance, while run also inherits the recipe's maintained renderer defaults: block_fill_style , color_fill_scope , stroke_detail , line_thickness , line-art snap and its threshold, preferred/capped block counts, overlap, and block order. The resolved recipe and optional --theme are stored in project.json ; their semantic fields participate in planning and resume fingerprints. Do not pass style-selection flags to render-photo or render-image ; those single-image commands do not resolve a recipe. They use explicit command defaults, including --line-thickness 0 , --stroke-detail rich , --block-fill-style crayon , --color-fill-scope block , and --block-overlap 0.08 , until their own flags override them. At the start of a new script-to-story job, run recommend-styles to inspect the best local candidates. This is decision support, not a mandatory approval gate: Honor an explicitly named style without asking again. When the user requests automatic selection, pass --style auto ; it chooses the first deterministic content-aware recommendation without a model or network call. When no visual preference is given, do not block on a 30-option question. Proceed with the stable default, or with one clearly better native recommendation when the user has delegated art direction; briefly report the choice. Prefer native for unattended final production. Use a short preview before an adaptive style, and explicitly flag experimental styles as source-dependent. Compatibility groups: native (15): warm-crayon-storybook , colored-pencil-diary , clean-whiteboard , minimal-line-explainer , marker-whiteboard , rough-diagram , pressure-ink-notes , semantic-ink , anime-graphite , bean-doodle-infographic , organic-contour-doodle , naive-marker-notes , notebook-pencil-doodle , inked-storybook , blueprint-pencil . adaptive (9): kid-crayon , raw-kid-crayon , emotional-watercolor-sketch , ink-wash-minimal , retro-gouache-concept , nordic-gouache-storybook , sunlit-storybook , editorial-portrait , real-crayon-paper . experimental (6): ballpoint-scribble , warm-flat-storybook , zine-riso-collage , manga-screentone , linocut-editorial , ms-paint-doodle . List or recommend recipes: CLI= " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" python3 " $CLI " list-styles python3 " $CLI " list-styles --compatibility native python3 " $CLI " list-styles --json python3 " $CLI " recommend-styles story.md -- limit 5 --json Choose a built-in recipe, use automatic selection, provide a bounded inline description, or load a repeatable UTF-8 text/JSON recipe: python3 " $CLI " run story.md -o /tmp/story.mp4 --style colored-pencil-diary python3 " $CLI " run story.md -o /tmp/story.mp4 --style auto python3 " $CLI " run story.md -o /tmp/story.mp4 \ --custom-style "Loose blue-pencil travel sketch, one warm-orange accent, broad white space" python3 " $CLI " run story.md -o /tmp/story.mp4 \ --custom-style-file /absolute/path/to/style.json \ --theme "Quiet early morning with restrained optimism" --style , --custom-style , and --custom-style-file are mutually exclusive. --theme is additive story-specific art direction and cannot replace the production contract: complete subjects, extractable outlines, natural object grouping, useful negative space, and no generated captions or pseudo-writing. Custom JSON should normally use extends plus bounded render overrides; see references/pipeline.md for the schema. An explicit custom JSON id must begin with custom- . provenance is not an accepted author field; the engine marks loaded recipes as user-authored itself. Visual recipes use generic media and production-method names, not artist names, and include no third-party sample images or brush assets. Style 9, anime-graphite , defaults to color_fill_scope=scene . Its people and props still draw as complete natural line-art blocks, but the registered color frame—including snow, walls, streets, sky, and paper-toned washes—is revealed only once across the full canvas. For this and similar full-bleed styles, make the storyboard prompt describe one continuous, low-detail environmental backdrop and reject panels, frames, rectangular scenic cutouts, or disconnected background islands. Script to Story Video The production scene pipeline is: script -> scene plan -> GPT Image 2 color frame -> local neural line art -> coarse natural blocks -> local details -> block-local color OR one whole-scene registered color/background pass -> optional sparse handwritten annotations -> silent or narrated MP4 -> editable sidecar SRT -> optional final subtitle burn-in block-speedpaint infers spatial drawing blocks for coarse contours and local details. With --color-fill-scope block , each block then receives its own local color beat. With --color-fill-scope scene , the same natural line blocks remain intact while the complete registered color frame and continuous background arrive in one full-width pass. Scene scope uses roughly the first 72% of the drawing interval for block lines; the global color pass starts near 68%, overlaps the last details by about 4%, and finishes at the end of the interval. A narrated run can use phrase timing cues to pace this same drawing clock: active speech advances the picture and meaningful pauses briefly hold it. The pipeline always writes an editable SRT beside the MP4 and leaves the picture clean by default; pass --burn-subtitles only when the requested deliverable needs narration subtitles baked into the final MP4. Optional annotations are short, positioned labels that never enter the GPT storyboard, extracted line art, or object grouping; the renderer intentionally adds them as late scene-video pixels, independently of the final SRT subtitle layer. Configure credentials through the shell or a secret manager, never in scripts, prompts, project.json , or committed .env files: Images and automatic scene planning: OPENAI_API_KEY ; optional OPENAI_BASE_URL must point to the intended trusted OpenAI-compatible endpoint. It is not required when both --scene-plan and --storyboard-dir are supplied. Silent master: use --tts-provider none ; no speech credential is initialized or required. Doubao new console: DOUBAO_TTS_API_KEY (aliases: DOUBAO_API_KEY , MODEL_SPEECH_API_KEY ). Doubao legacy console: set both DOUBAO_TTS_APP_ID and DOUBAO_TTS_ACCESS_KEY instead of the API key. Doubao defaults: DOUBAO_TTS_RESOURCE_ID=seed-tts-2.0 , DOUBAO_TTS_ENDPOINT=https://openspeech.bytedance.com/api/v3/tts/unidirectional/sse , DOUBAO_TTS_VOICE=zh_female_vv_uranus_bigtts , MP3 at 24 kHz. The provider requests Seed-TTS 2.0 subtitle timing and uses its sentence/word timestamps for animation and SRT timing. Use only a speaker ID enabled for the user's account and do not redirect credentials to another endpoint. Run the real color-storyboard pipeline: python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" run story.md \ -o /tmp/story.mp4 \ --scenes 6 --fps 30 --width 1920 --height 1080 \ --style warm-crayon-storybook \ --image-model gpt-image-2 --image-quality low \ --scene-assets color-to-lineart --lineart-provider auto \ --animation-preset block-speedpaint \ --tts-provider none Use --image-quality low for drafts and medium for a final render when the visual gain justifies the extra cost. The full run command defaults to 30 fps, gpt-image-2 , no burned narration, and block-speedpaint . Its fill, fill scope, line, snap, and natural-block settings are not separate fixed CLI defaults: when omitted, they inherit the resolved style snapshot. The stable warm-crayon-storybook recipe currently resolves to automatic line width, rich stroke detail, crayon fill with color_fill_scope=block , at most four preferred natural blocks, and 0.16 overlap; selecting another style may change any of these values. Style 9 resolves to color_fill_scope=scene . Connected objects are never split merely to reach a count; use --draw-blocks 0 for uncapped automatic grouping up to the resolved max_draw_blocks . Use --tts-provider none for a silent edit master, or select Doubao/Edge and pass --voice <speaker-id> . Both narrated and silent runs write <video-name>.srt . Add --burn-subtitles to render that SRT into the final -o MP4 while retaining the sidecar; style it with --subtitle-font , --subtitle-font-size , --subtitle-margin-v , and --subtitle-outline . The legacy --captions and --no-captions flags are deprecated no-ops and are mutually exclusive with --burn-subtitles . Scene annotations remain independent. Only run treats the following as overrides of the selected style. Omit them to inherit the recipe: --block-fill-style crayon|clean|soft-wash|dry-brush --color-fill-scope block|scene --stroke-detail balanced|rich|max --line-thickness 0..16 ( 0 requests automatic source-aware sizing) --line-art-snap / --no-line-art-snap and --line-art-snap-threshold 1..254 --max-draw-blocks N , --draw-blocks N or 0 for automatic grouping, --block-overlap 0..0.65 , and --block-order reading|source --block-sequence 1,0,... for an inspected explicit inferred-block order; this is a run-time ordering instruction rather than a recipe default For a 16:9 Doubao Voice 2 delivery with burned narration subtitles: python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" run story.md \ -o /tmp/story-subtitled.mp4 \ --scenes 6 --fps 30 --width 1920 --height 1080 \ --image-model gpt-image-2 --image-quality low \ --scene-assets color-to-lineart --lineart-provider auto \ --animation-preset block-speedpaint --draw-blocks 4 --block-overlap 0.16 \ --tts-provider doubao --voice zh_female_vv_uranus_bigtts \ --burn-subtitles --subtitle-font "sans-serif" \ --subtitle-font-size 16 --subtitle-margin-v 22 --subtitle-outline 1.6 FFmpeg/libass burns subtitles only after the clean scene composition and SRT have succeeded. The engine renders to a temporary sibling file and atomically replaces the requested MP4, so a subtitle-filter failure leaves the clean MP4 and SRT available for recovery. With Doubao Voice 2, use valid provider subtitle timing as the preferred narration clock. Group returned word timestamps into short readable phrase beats; drawing advances at a steady rate while a phrase is spoken and briefly holds across real punctuation pauses. The measured narration duration, not the estimated scene-plan duration, controls a narrated scene; duration_sec remains the silent-render fallback. Persist provider timing beside cached audio as scene_NN.alignment.json , and clamp every SRT cue to the real audio endpoint so no subtitle leaks into the visual tail hold. If Doubao returns no valid timing, retain valid authored timing_cues ; only when neither exists should the SRT use deterministic estimated phrases while drawing keeps its continuous clock. Narrated composition must have one audio file for every scene; never compact a partial audio list because that shifts later voices into earlier scenes. If Codex or another tool already generated the color storyboards, preserve their scene order and name them scene_01.png , scene_02.png , and so on ( .webp , .jpg , and .jpeg also work). Save the approved scene list as JSON with id , narration , image_prompt , and positive duration_sec fields. An optional annotations list may contain at most two objects with text , x , and y ; coordinates are normalized top-left positions. Keep each label to 2–5 Chinese characters when possible, leave most scenes empty, and keep y <= 0.72 so the lower area remains available for post-production subtitles. An optional ordered timing_cues list accepts text , local start_sec , local end_sec , and optionally cumulative draw_to progress ending at 1.0 ; use it for a pre-aligned silent or imported-audio workflow. In a Doubao run, valid official timestamps replace authored cues; authored cues remain the fallback if the provider returns no usable timing. Then supply both assets to bypass OpenAI scene planning and image generation entirely: python3 " ${CODEX_HOME:- $HOME /.codex} /skills/whiteboard-video/scripts/whiteboard_cli.py" run story.md \ -o /tmp/story-from-frames.mp4 \ --scene-plan /absolute/path/to/scene-plan.json --fps 30 \