One-Take White-Box Previsualization Skill

One-liner: User provides an image or a description, this skill outputs a white-box video; after the white-box is finalized, it writes prompts, and the white-box video is used as a motion reference alongside the prompts for the final video model. Long takes don't rely on mystical prompts; they rely first on precomputing the camera path and having the user visually confirm it.

0. Execution Method (We Only Talk About These Three Steps Externally)

User provides a reference image, or a description ("10 seconds, orbit from behind me to the front-side, objects on the desk float up first").
The skill generates a 3D white-box camera preview video; both parties iterate together until the user says "This is the feeling."
After finalization, deliver two items: White-box preview video + detailed video prompts, both fed to Seedance 2.5 for final rendering.

What happens internally (scene, coordinates, validation) is the skill's business; don't bother the user with technical details. Full intake rules, required questions, and plain-language translation table are in references/00-flow-intake-iteration.md, which is the master flow document for this skill. Read it first on every trigger.

1. Six-Stage Main Process (Order is Fixed)

Stage	User Action	Your Delivery	Pass Condition
① Intake	Provide reference image / description	3-sentence restatement of "the scene I understand," including subject, camera movement direction, whether floating/weightless	User says "yes" or changes one sentence; confirm aspect ratio and duration
② Env Check	Binary choice (install / don't install)	`check_blender.py` result + one-line route explanation	Route is clear, proceed to A or B
③ First White-Box Video	Watch the video	Playable mp4/webm + 3–4 keyframes + "Tell me what's wrong directly"	Video is delivered and viewable. Providing only coordinate table = breach of contract
④ Conversational Iteration	Give feedback in plain language	`v2/v3...` new versions of the video + 1-line change description	Each round changes only one thing; user explicitly approves
⑤ Finalization + Prompt Writing	Confirm the feeling is right	Four validations `RESULT: PASS` + six-section prompts	Do not write prompts if validations have not PASSED
⑥ Delivery	User takes assets for final rendering	White-box video + prompts + upload description (video1 = motion reference, image1 = visual reference)	All prompt coordinates come from measured JSON

Stages ③ and ④ can loop multiple times. This is the core value of the skill: fix camera issues in 3D, not by rolling dice with video model prompts.

2. Six Ironclad Rules

Order rule: Build scene → White-box video → User confirmation → Write prompts. Writing prompts before the user has seen the video is a failed draft.
Prompts use the user's reference image as the sole visual basis. The white-box is not a visual reference; it is only for locking motion, timing, and coordinates; quality, shaping, lighting, and color must all revert to the reference image.
Camera movement must be unidirectional and continuous. Horizontal angle must be monotonically increasing or decreasing; reversing left/right is a failure; readjust the curve.
Subject must never leave the frame. Validate frame-by-frame projection; the subject's NDC must stay within a safe margin; when the subject rises, the camera must tilt up simultaneously.
Zero interpenetration. If the distance between any two moving objects in any frame is below a threshold, it is a failure; solve by prioritizing "clear space before rising" and "light objects fly first, then heavy objects," not by cramming density.
User only speaks plain language. User says "higher" or "the chair hits my leg"; you translate that into parameters and change only that one thing; don't explain polar angles, NDC, or AABB to the user.

3. Environment Detection and Route Selection (Tip 1: Detect and Build)

check_blender.py
  ├─ Installed → Branch A (Blender, highest precision, can export frame-by-frame measured data)
  ├─ Not installed
  │    ├─ User agrees to install → Guide to portable version (05-blender-setup.md) → Branch A
  │    ├─ User has other version/path → Fill in path → Branch A
  │    └─ User refuses installation (or no permissions, time constraints) → Branch B (Three.js web white-box)
  └─ Branch B also goes through the full six stages, just uses "motion reference"; validation and prompt writing standards are identical

Installation guide phrasing — give one decision point at a time, no long paragraphs:

Currently, there's no Blender installed. Two options:

I install the portable version (about 300MB, installs on a non-system drive, no registry edits, doesn't use C drive, can be deleted entirely) — highest precision, can export frame-by-frame measured coordinates;

Don't install, I'll generate a white-box preview video using Three.js in the browser — you'll see the camera path in a few minutes, slightly lower coordinate precision. Which do you choose?

User chooses 1 → follow references/05-blender-setup.md; chooses 2 → follow references/06-threejs-fallback.md.

4. Stage Details

① Intake

Read references/00-flow-intake-iteration.md §§1–3: Determine if input is an image, a description, or both. Only three mandatory questions (duration/aspect ratio, subject orientation, camera intention); infer if possible, otherwise use template values from knowledge/kb-levitation-office.md as fallbacks. If user says "whatever you think," use defaults for the first version and let them watch to point out issues.

③ First White-Box Video (Tip 3: Output White-Box Preview First)

First build a static white-box and calibrate world coordinates (1 unit = 1 meter, origin at ground level) → parameterize camera trajectory → 480P low-sample render → output video + keyframes. Do not write prompts at this stage. Acceptance criteria: playable, no black frames, subject stays in frame, room walls/floor are static.

④ Conversational Iteration

Use the translation table from 00 §4 to translate user's speech into parameters, change one thing at a time, output a new version with a 1-line change description (e.g., v2: radius 3.45→2.95, subject fuller in final frames). Version naming follows 00 §5.

⑤ Finalization, Validation, and Prompt Writing (Tip 2 Auto-Check + Tip 4 Write Prompts Against White-Box)

All four validations (polar angle monotonicity / subject out-of-frame / cross-group penetration / ceiling collision) must be RESULT: PASS, then export the frame-by-frame measured JSON, then write prompts according to the six-section structure. Every coordinate in the prompts must come from this JSON; manual estimation is forbidden. Writing standards are in references/04-prompt-format.md.

⑥ Delivery (Integration with Seedance 2.5)

Follow the delivery format from 00 §6: tag the white-box video as motion/camera/timing reference, tag the reference image as the sole visual reference, start the prompts with an emphasis on "strictly replicate the camera motion and takeoff sequence of the reference video," align duration with the white-box, and break camera movement into 2–3 second segments.

5. Branch A: Blender Execution Manual

Scene and coordinate standards: references/01-scene-spec.md
Camera motion design standards: references/02-camera-motion.md
Layered levitation and anti-penetration timing: references/03-levitation-timing.md
Installation and connection: references/05-blender-setup.md
Scripts:
- scripts/build_scene_template.py — Creates empty white-box scene skeleton (room, desk, dual monitors, chair, human block)
- scripts/render_previs.py — Low-sample 480P preview rendering (EEVEE, no materials)
- scripts/export_cam_data.py — Exports frame-by-frame cam/lookat/elevation/object z to JSON
- scripts/validate_trajectory.py — Four validations (monotonicity/out-of-frame/penetration/ceiling collision)

Command-line sequence (tested locally; $BLENDER is absolute path to blender.exe, replace paths per project):

# 1 Probe
python scripts/check_blender.py --json

# 2 Build scene + write orbit trajectory and layered levitation animation (--orbit 1), and export coordinate list
"$BLENDER" -b --python scripts/build_scene_template.py -- \
  --orbit 1 --center 0.0 0.15 --dur 10 --fps 24 \
  --a0 186 --a1 300 --r0 3.45 --r1 2.45 --z0 0.78 --z1 1.20 --l0 1.05 --l1 1.45 \
  --out "Z:/project/scene.blend" --csv "Z:/project/coords.csv"

# 3 Four validations (must run before finalization), must be RESULT: PASS
"$BLENDER" -b "Z:/project/scene.blend" --python scripts/validate_trajectory.py -- \
  --center 0.0 0.15 --ceiling 2.75 \
  --keys Person_Head,Person_Torso,Desk_Top,Monitor_H_Body \
  --moving Minion_,Monitor,PC_Case,Keyboard,Mouse,Person_,Desk_,Chair_

# 4 Export frame-by-frame measured data (only use this data for writing prompts)
"$BLENDER" -b "Z:/project/scene.blend" --python scripts/export_cam_data.py -- \
  --out "Z:/project/cam.json" --center 0.0 0.15 --keys Person_Head,Desk_Top,Chair_Seat

# 5 Preview rendering: first a 16-frame sample to verify pipeline, then full sequence
"$BLENDER" -b "Z:/project/scene.blend" --python scripts/render_previs.py -- \
  --out "Z:/project/previs.mp4" --res 960x540 --samples 4 --start 1 --end 240

For iterative changes, just re-run step 2 (modify params) → step 5 (render); add steps 3 and 4 after finalization. --keys only includes subject key points (head, torso, desktop, main monitor); adding accessories below 1m from the subject (chair seat, etc.) will inevitably fail out-of-frame validation — this is a physical result of the lens. --moving only includes prefixes of objects that will move; if not set, no penetration detection is performed; objects are grouped by the first part of their name; contact within the same group is not counted.

6. Branch B: Three.js Web White-Box

Plan and limitations: references/06-threejs-fallback.md
Runnable directly: scripts/threejs_previs.html (single file, CDN-imported three.js, includes trajectory playback, perspective/top view toggle, JSON export, webm recording)
Limitations: Human modeled as block approximations, all white materials; can only verify motion direction, composition relationship, and timing; cannot verify lighting or texture.
Output: webm + keyframe screenshots; coordinates exported from the same set of parameters in the page.

7. Prompt Delivery Standards (Tip 4)

Fixed six sections, all required: 【Generation Unit】 【Reference】 【Continuity】 【Narration and Audio】 【Timeline with Coordinates】 【Negative Prompts】.

Three hard requirements:

Write against the white-box, don't guess: Every camera movement sentence must refer back to the white-box timecode, e.g., "0–2.5s Starts at the subject's lower left back, camera height 0.78m, radius 3.45m, polar angle 186°, elevation angle 3.7°."
Distinguish reference roles clearly: The white-box video locks only motion, the reference image locks only visuals; both must be written explicitly.
Write anti-penetration logic as causal order sentences, not "be careful not to penetrate."

Writing techniques, phrasing, negative examples, and common failure causes are in references/04-prompt-format.md; project-measured parameters are in knowledge/kb-levitation-office.md (can be directly reused as template values).

8. Common Fault Diagnosis

Symptom	Root Cause	Resolution
Camera oscillates left-right erratically	Polar angle curve is non-monotonic / interpolation bounces	Monotonic parameterization + first-order difference sign check + LINEAR interpolation
Subject leaves frame mid-shot	Subject rises but camera doesn't follow	Keyframe camera z and lookat z synchronously with a follow coefficient of 0.8–0.95
Desk/chair penetrates person	Person and chair take off simultaneously	Person flies first; chair is delayed by 0.4–0.7s, and clears horizontally before rising
Equipment collides with desk	Equipment and desk take off simultaneously	Equipment rises about 1.3m and hovers; desk is delayed by 2–3s
Room drifts along	Lookat point or parent relationship is wrong	Only apply z-displacement to the subject group; walls, floor, windows are static
Final video character turns around, face is broken	Orientation not locked	Write in 【Reference】 to lock back/side-back orientation; add turning around, looking back, face close-up to negative prompts
Final video looks like 3D gray model	White-box was treated as visual reference	Specify clearly in 【Reference】 that "white-box references motion only"; lock visual style to realistic photography
Model completely ignores the reference video	Opening sentence didn't emphasize replication	Write the first sentence as "Strictly replicate the camera motion and takeoff sequence of the reference video"; break camera movement into 2–3s segments
Prompt coordinates don't match the visual	Used estimated coordinates	Go back to stage ⑤, get values from the measured JSON; manual estimation is forbidden

9. High-Quality Delivery Checklist

10. Script Parameter Quick Reference

Script	Key Parameters	Description
`check_blender.py`	`--json`	Read-only probe; outputs found/path/source/version
`build_scene_template.py`	`--orbit 1` `--center cx cy` `--dur` `--fps` `--a0 --a1` `--r0 --r1` `--z0 --z1` `--l0 --l1` `--out` `--csv`	Without `--orbit`, only outputs static scene; with it, writes frame-by-frame camera and levitation keyframes (LINEAR)
`validate_trajectory.py`	`--center` `--ceiling` `--keys` `--moving` `--margin` `--tol`	Exit code 0=PASS 1=FAIL; four checks
`export_cam_data.py`	`--out` `--center` `--keys`	Outputs `[{f,cam,rad,ang,pitch,对象:[x,y,z]}]`
`render_previs.py`	`--out` `--res` `--samples` `--fps` `--start` `--end` `--engine`	First a 16-frame sample, then full sequence
`install_blender.ps1`	`-TargetDir` `-Version` `-Mirror auto	official

Parameters are always passed after -- to the Python script.

11. External Talking Points (When Asked "What Can This Skill Do?")

Give it a reference image, or a description. It first builds your camera movement as a 3D white-box and runs it for you to watch. If something's wrong, modify it. Once satisfied, it writes the prompts. The white-box video and prompts are then fed together to Seedance 2.5 for the final rendering. Environment detection and modeling, four automatic validations, white-box preview, and writing prompts against the camera path — all included. Whether it's an orbit, a rise, or weightlessness, the visuals are smooth without sudden penetrations or deformations. With this one-take Skill, long takes are terrifyingly stable — do you still need to spend a huge amount of time learning those mystical prompts for AI video generation?

12. Skill Self-Check Status

Step	Result
Detect Blender	Passed (4.5.15 LTS portable, D drive)
Build scene + trajectory + levitation animation	Passed (240 frames @24fps, coordinate CSV saved)
Four validations	PASS: Out-of-frame 0 (960 NDC checks), new penetration 0, ceiling collision 0, all three monotonicity checks True
Export frame-by-frame data	Passed (240 frames JSON)
Preview rendering	Passed (EEVEE_NEXT, 16-frame sample mp4 normal)
Installation script	Passed (`-DryRun` tested: official 403 → Aliyun 4.5.16, ~380MB)

Template default parameters are in knowledge/kb-levitation-office.md and can be used directly as a starting point for new projects.

Daham0/blender-previs-one-take

blender-previs-one-take 是什么？

在你喜欢的 AI 中提问

文档