cloudhanfeng92-arch/automotive-promo-video-reference

Car Promotional Video Director – No Storyboard Version

This is an independent version, retaining automotive-promo-video and automotive-promo-video-rhythm. The workflow chain is: Script + Text Storyboard → Vehicle Four-Views + On-Demand Interior/Scene Reference Cards → Multi-Shot Reference Image-to-Video Prompts for the Entire Segment → Generation, Editing, and Delivery as Authorized.

Scope & Tools

• No per-shot storyboard images are generated; no per-shot first/last frame is required; no previsualization based on per-shot image generation is made. The text storyboard must still be complete. Independent views from compliant reference boards may be cropped; this is about organizing identity/space references, not generating shot compositions.
• Images use the user-specified Lib Image 2.5 Pro; the user also calls it GPTimage2.5; the execution name follows the current LibTV directory, not an alias to prove the vendor version. Video defaults to Seedance 2.5, multi-image reference mode.
• LibTV operations read the installed libtv-cli Skill and follow the CLI via Execution Rules. Do not silently switch to first-frame mode, other models, or re-add storyboard images.
• When planning, read Rhythm & Segmentation; when in the reference asset phase, read Reference Asset Specification; when writing video prompts, read Fixed Format; when examples are needed, read the 20-second Complete Example. When local editing is required, read ffmpeg-skill; writing Skills/scripts/prompts alone does not start media generation or editing tools.
• When the user asks for high-end car PV, speed & passion, or stronger commercial photography dynamics, read the Car PV Director Branch to reorganize the plan with motion openings, dynamic-static contrasts, and driving climaxes. This branch is not automatically applied for quiet showcases, showroom introductions, etc.
• People, voices, file paths, storylines, and dialogue in user examples are format references only; do not read unrelated files, copy them into the car plan, or treat them as generation authorization.

1. Read Car Images, First Write Script & Text Storyboard

Upon receiving car images, automatically enter this workflow; do not require the user to additionally say "make four-views." First, observe the actual uploaded images and create vehicle-bible.md: source U01 etc., visible appearance, paint/two-tone boundary, grille/light shape/wheels, left/right-hand drive and interior (if visible), and invisible parts. Mark each with observed / user-confirmed / inferred / unknown.

Reference videos guide photography and rhythm; do not assemble vehicle models, wheels, or interiors from different videos into a single car. Parts not visible in a single front-side view (rear, right side, interior) must not be passed off as actual configuration; marked supplementary proposals can be made. Critical configurations require supplementary images or user-approved visual settings.

Use conditions already provided by the user directly; otherwise, suggest 16:9, 20 seconds, ~30fps delivery, no dialogue/voiceover, and clarify these as adjustable defaults. The number of shots in the text storyboard is determined by readability and rhythm; do not adopt the original 26-shot/20-second quota. At 0.5-second granularity, short and long shots must form an informative combination.

Speed and cinematography intensity follow the intention of the film; do not unify "luxury" as slow speed, static shots, or every shot being a slow push. The PV branch can design natural acceleration, subject-camera approaching/receding movements, and close-up fly-bys; maintain tire contact, vehicle weight, and directional coherence; do not claim real performance based solely on advertising shots.

Must first output:

1. Script: Title/a one-line creative idea, audience knowledge progression, setting and time, vehicle state changes, opening hook, and closing; if no dialogue, write a visual script without forcing characters, voices, or narrative voiceover.
2. Rhythm Structure: How long detail short groups, full vehicle stance/cabin breathing room, driving sections, and closing each take, and explain the main reference rules.
3. Text Storyboard Table: Shot numbers for the entire film, assigned CLIP, in/out points for the whole film and within each segment, shot length, shot size/composition, vehicle action, camera action, cut event/transition to next shot, required reference ID.
4. Asset Requirements Table: Vehicle four-views are mandatory; whether interior and main scene cards are needed, why, and what sources are missing. Do not generate useless assets just to fill a template.

When the user only wants script/storyboard/prompts, deliver within that scope. If reference images are missing, deliver a text draft, mark pending bindings in a separate execution note; do not write asset status into model prompts, nor pretend assets have been generated.

2. Automatically Prepare & Generate Reference Assets, No Storyboard Images

"Automatically" means upon receiving images, automatically analyze, plan, organize execution; do not require the user to repeatedly initiate a four-view task. Retain the user's previous requirement to "confirm before outputting images": Display the full batch of image prompts, reference sources, model/parameters, quantity, and known costs or unknowns before generating. If a specific batch or automatic generation authorization already covers content, quantity, parameters, and budget range, proceed directly; only providing a budget cap or uploaded files does not replace confirmation before generation. If the user explicitly authorizes automatic image generation, execute within that authorization scope.

• CAR_EXT_4V: Default one 2×2 four-view image, front, full left side, rear, full right side, same car with same configuration; describe left and right sides separately; do not treat a mirror image as the opposite side. Record inferences for unknown structures and user confirmation.
• CAR_INT_01: Generate only when interior is needed for the script, visibility through windows, or continuous inside/outside action. References must cover relevant cockpit/center console/rear seats/door panels; do not confirm real interior from exterior photos. If clear, trustworthy interior photos already exist, use them directly as references without additional paid generation.
• SCN_01, SCN_02: Generate multi-view scene cards only when the main scene appears repeatedly, multiple camera positions shoot the same location, and entry/exit points/routes/primary light relationships affect continuity. Simple backgrounds can be supported by text description or existing images; explain why no new card is needed. Each card describes only one physical location; do not mix different scenes into the same space.

Verify identity, interior, and space frame by frame according to the Reference Specification, then record acceptance status. Display results for user confirmation; if explicit authorization for self-verification against the archive exists, proceed autonomously. Cropped independent views still retain the original image and source; low-resolution crops do not serve as detailed references.

Subsequent video input can use four-views/realistic reference cards and their clear crops; note each as "same subject, different angle" or "same location, different angle." Scene cards containing top-down floor plans, route arrows, or extensive text must first have a realistic view cropped; schematics are only used for writing spatial descriptions, not as video input. Reference boards must not be treated as layout composites for the frame, nor input through the first-frame port. Do not force the model to sequentially switch between board cells into video just because reference cards exist.

3. Organize Multi-Shot Video Plan for Each Segment Up to 30 Seconds

According to Rhythm & Segmentation, divide the entire film into segments like CAR_CLIP_01. 30 seconds is the upper limit per segment, not a total length limit, nor a requirement to fill exactly 30 seconds.

• If the entire film is ≤30 seconds and the shot/scene relationships are clear and capability permits, prioritize a single complete segment; clearly specify multiple hard cuts within that segment.
• For longer films, split at scene/action or rhythm segment boundaries, preferably avoiding difficult-to-connect short tails. Complex exterior-to-interior action or dense shots may be split into additional segments, but do not revert to "one shot, one image, one video."
• The actual segment length must also satisfy the model's current minimum, maximum, and step size; if the current supported maximum is below 30 seconds, follow the lower maximum. If the input duration is invalid, re-divide or explain material surplus; do not fabricate model parameters.
• Each segment's local time starts from 0.0; indicate its start and end times in the full film, with local shot intervals continuously covering the segment. Default to 0.5 second as the boundary unit; can write 0.0–0.5, 0.5–1.5, 1.5–3.0; not every 0.5 seconds requires a cut. In long shots, action changes can be described at half-second points.
• The timeline is a director's target; do not guarantee frame-by-frame exactness by the model. Upon receiving actual results, check cut points, missing shots, order, and duration drift; if necessary, revise prompts, re-segment, or calibrate with editing within authorization.

Maintain two counts: the number of text storyboard shots and the actual number of CLIP tasks. Example: 20 seconds containing 16 shots, planned as one 20-second video task; do not write as 16 separate 5-second tasks. If splitting into two segments is necessary, clearly state new segment length/reference/budget; do not hide additional generation.

4. Output Submittable Timeline Prompts

For each segment, output a single, copyable, pure video prompt block according to the Fixed Format, in the following order:

Segment Name → Subject Reference → Interior Reference (if needed) → Scene Reference or Setting → Reference Image Content Explanation → Starting State → Each Shot Number/Name/Segment Time/Image/Camera Motion/Transition → Unified Style & Physics Constraints → Ending State.

Only write what the model should present. Model name, mode, resolution, aspect ratio parameters, frame rate, sound on/off, task count, intended/unsubmitted status, pending binding/acceptance notes, local paths, budget, post-production arrangements, and capability checks are all placed in a separate execution-settings.md or structured execution record. Do not include process notes like "generation specs," "this is the director's time target," or "check cut points after generation" in the pure prompts. Timestamps and action windows for each shot must be retained; full-film/segment time conversion and half-second grid explanations go into the storyboard table or execution record. Additionally save a .txt file containing only the model body text; upload only that text.

Use "subject/vehicle reference" as the subject; do not assume "people and voice reference." Only include character basis if the user explicitly requests people. Do not require voice reference files. Real dialogue, environmental sounds, and final content requirements like "no dialogue, no voiceover" can be written into prompts; interface parameters like enableSound=off and post-production music arrangements go into the execution record. Default generation has sound off; configure according to user requirements for environmental sounds; do not automatically request voice.

Execution notes or asset tables can link actual local reference files, marking items pending generation/binding; do not use fake paths. In the pure prompts, only keep reference IDs or tool-supported image references and their visual responsibilities. Before submission, the execution layer binds IDs to actually uploaded resources; local Markdown links are not model-readable upload references; do not mix paths or upload reminders into the prompt body.

Within a segment, shot transitions are allowed; write "hard cut to..." or matching transitions for each shot. Do not write "single continuous shot, no cuts" at the end of a segment, conflicting with a multi-shot timeline. Only prohibit undesigned cuts, vehicle model drift, split screens/grid lines/labels, image card slideshows, extra vehicles, body melting, and incorrect mechanical motion. Describe shot sizes and angles in text; do not require corresponding shot images to be prepared first.

5. Check, Submit as Authorized, and Deliver

First, verify the correspondence between script/storyboard/prompts, that each segment is ≤30 seconds, each segment covers from 0 to its end, the entire film has no gaps, the car/interior/environment is consistent, and each reference role is clear. Use the Timeline Validator to check the JSON Example or project plan:

python3 <skill-dir>/scripts/validate_timeline.py <project timeline.json>

Before generation, add --require-local-assets to check referenced files and asset acceptance status; this does not replace CLI real-time model schema, upload success, and budget authorization checks. Record state as planned → reference_approval_pending → references_ready → references_accepted → prompts_ready → video_authorized → clips_ready → edited → delivered; do not treat state names as user authorization.

If the user only authorized prompts, stop at prompt delivery. If full video is authorized and this batch of video falls within scope/budget, connect references in real image2video mode and submit; no need to ask again; "intended" in prototypes or drafts does not equal actually submitted. If multi-image reference is not supported, explain limitations, retain this version's plan; do not automatically switch to first-frame mode or generate storyboard images.

Record generation task canvas/node/task ID, prompt version, parameters, and final state; for anomalies, first check the original task; do not re-submit. Download the complete raw footage; check each segment's shot order, time targets, vehicle structure, tire contact, reflections, interior/exterior elements, and reference card leakage; then edit, add music, and unify color. Adjust time discrepancies using actual footage; do not pan/zoom reference boards as fake dynamic car shots.

Deliver according to user scope: script, storyboard table, reference assets and their source mappings, each segment's pure timeline prompts, separate execution notes and JSON; when actually performing complete production, also deliver raw videos, actual editing logs, and the accepted MP4. Maintain in the project: script.md, storyboard.md, vehicle-bible.md, references/, reference-manifest.json, timeline.json, clip-prompts.md, pure text .txt, and execution-settings.md; there is no mandatory Sxx_FRAME item.