Creative Space
Turn a Script into a Blender Previsualization

เนื้อหา

Provide: The script or scene text of a scene (dialogue, action, scene); the target duration, aspect ratio; the problem you want to solve (pacing, blocking, a certain transition). If you have reference pictures, reference films, or storyboards, please provide them together.

Receive: A dynamic storyboard video (scenes and characters with gray and white blocks, real camera position and camera movement, duration of dialogue and action stuck, onscreen shot numbers; dialogue remains in the shot sheet and optional temporary voice track); shot table; 3×3 key shot overview; overhead blocking diagram; .blend project that can continue to be modified.

What dynamic storyboards need to answer is "how long will this scene take to be shot, whether the rhythm is right, and whether the shots can be connected." It is not whether the picture is good or not. Therefore, gray and white bodies were used throughout the whole process, without adding textures or physical performances, and spent time focusing on camera position, duration and editing points.

Three usages

UsageWhat to do
The script becomes a dynamic storyboard (default)Take down the shot → Blender sets up the scene to set the camera position → Render it into a preview video
shot timeline and 3×3 storyboardOnly text shot table and 9 key shot still frames, the entire animation will not be rendered
Change only one sectionIf you have already rehearsed, you can only change the scene, duration, camera movement of a certain shot, or re-schedule the rhythm of a certain section, and leave the rest unchanged

Tools and responsibilities

  • Agent: Unpack shots and write shot tables; write Blender Python scripts for each step, and use the command line to let Blender execute in the background. Main force.
  • Blender: Blender is installed on the user's local machine and runs in the background through the command line. No plug-ins or interface are required. Required. For usage, see .
  • FFmpeg: Combine the rendering results of each shot into a video, mix it into a temporary voiceover, and create a 3×3 overview. Required.
  • storyboard planning: Organize the target storyboard, duration and elements, check the space and action continuity of adjacent shots; the general storyboard method follows it, this inspiration only supplements the requirements of Blender preview.
  • Flova (optional): Temporary voiceover, the duration is based on the actual length of the voiceover; after the preview is confirmed, the blockout model screen can be used as a composition reference to generate keyframe or finished videoshot.
  • Captions (optional): If requested, offer a suitable HyperFrames, Remotion or FFmpeg post-production workflow.

Method index

  • : Find Blender, run in the background, store all states in a .blend, volume characters, multiple camera positions, rendering and burning information, and common pitfalls. Steps 3–6 Read.

Boundaries

  • Do not write scripts for users: If there is no script or scene text, the work will not start; if the script is not clearly written (who stands where, where to enter), list it and ask the user, or mark "my hypothesis".
  • Only run locally: Requires agents that can run Blender and FFmpeg locally. The cloud host can only do the text part of the "shot timeline and 3×3 storyboard".
  • Blocks are not performance: Characters only change position, direction and blocking. Record body performance and expressions in the shot sheet.
  • Scripts have full permissions: Blender scripts written by Agent can read and write local files, only in the project directory; .blend from unknown sources will not be opened.

Workflow

Confirm the script and goals (confirmation point) → Take apart the shot (confirmation point) → Set up the scene and create a top-down blocking diagram (confirmation point) → Set the camera position and do camera movement → Duration → Rendering and splicing (confirmation point) → Modify → Deliver

1. Confirm the script and goals (confirmation point)

  • Read the script and ask: target duration, aspect ratio, when to use it (self-reflection, showing to the team, proposal to the client), and the problem you most want to solve.
  • Find out what is not explained in the script but must be known in the preview: the size of the space, the location of doors and windows, who enters from where, and where the props are. Listed for user confirmation, or marked as hypothesis.
  • Check this machine: whether Blender and FFmpeg can be run from the command line, and the Blender version. If Blender is missing, it means that this inspiration cannot run. You can do the text part of the "shot timeline and 3×3 storyboard" first.

2. Remove shot (confirmation point)

Use the storyboard planning to organize the shot table docs/shots.md, and continue to use its number, script association and continuity check; this time, complete the scene type, camera position and focal length, main camera movement, character start and stop position, estimated seconds and corresponding frame interval. The camera position first writes the relative position and height, and the specific coordinates are written into the Blender script. First confirm which shots the emotional focus is on, and then confirm the entire list for the user. As long as the shot sheet and 3×3 shot plan are delivered with the text version, the storyboard is not considered delivered when the picture has not yet been rendered.

3. Set up the scene and create an overhead blocking picture (confirmation point)

Follow : Use basic bodies to build a space (floor, walls, doors, main furniture and props). For each character, place a body block with an orientation mark, a different color, and a name on the top of the head at the position of the first shot. Render a top view, mark the starting and ending positions of the character and each camera position, and let the user confirm the space and blocking. If the space is not right, everything will have to be reworked later.

4. Set the camera position and do camera movement

There is one camera for each shot, and the focus and position are set according to the shot table; push, pull, shake, and follow use keyframe, and follow the shot with the "align target" constraint. Keyframe is also used for character movement. Put all shots on the same timeline, use markers to switch cameras, and the frame interval is the shot duration.

5. Card duration

  • Follow the storyboard duration method: read the dialogue, estimate actions and pauses, convert the frame interval according to a fixed frame rate, and indicate that it is still an estimate.
  • When the user wants to listen to the rhythm, use Flova to generate a temporary voiceover, and adjust the shot frame interval according to the actual length of each voiceover.
  • When there is too much difference between the total duration and the target, first tell the user which shots the difference is, and do not compress evenly.

6. Rendering stitching (confirmation point)

Low-definition rendering is done shot by shot, and the shot numbers, shot sizes and timecodes are burned into the image; they are pieced together using FFmpeg and mixed with temporary voiceover. In the 3×3 overview, select the keyframes that create space, emotional focus, transitions and endings in chronological order. Each frame is marked with the shot number and duration. If there are less than 9 shots, different stages of long shots can be used. The same shot is marked, and no additional shots are added to complete the frame. Show the user the entire preview to confirm the rhythm and editing points.

7. Modification

Only change the involved shots according to the user's opinions: change the shot table → change the camera, frame range or position of those shots → only re-render these shots → re-stitch. For each modification, record a line in docs/changes.md: which shot was changed, what was changed, and why.

When the shot duration changes, the subsequent camera cut marks, camera and character keyframes and sounds are moved synchronously; only the connection between adjacent shots is checked when the scene is changed, and the entire scene is not redone.

8. Delivery

Deliver preview video, shot sheet, 3×3 overview, overhead blocking diagram, .blend project and temporary voiceover. When users want to continue making finished videos: they can use the blockout model screen of each shot as a composition reference, press the Flova Skill to generate a keyframe, and then generate a video shot.

Project directory

previs-<片名>/
├── docs/     brief.md、shots.md(镜头表)、changes.md
├── scripts/  每一步的 Blender 脚本,按步骤编号
├── scene/    previs.blend
├── audio/    临时配音
├── render/   各镜头渲染帧、俯视图
└── output/   previs.mp4、contact-3x3.png、plan-top.png

Blender command line

No plug-ins, no MCP, no interface: Agent writes a Python script for each step and lets Blender execute it in the background.

Find Blender

First, search on the user's local machine: can blender be run directly from the command line, is there a common installation location, and what is the version (blender --version). If you can't find it, please tell the user the path or install it first. Different versions of Python have different interfaces. Please confirm the version before writing a script. If you use an uncertain interface, print it in the script and check it first.

Run in the background

blender -b --factory-startup --disable-autoexec scene/previs.blend --python scripts/04_cameras.py
  • -b runs in the background; --factory-startup does not load the user's own settings and plug-ins; --disable-autoexec does not execute the scripts included in .blend.
  • In the first step, if there is no .blend, do not write the file name. Create a new scene in the script and save it.
  • When a script error occurs, Blender does not necessarily return failure. Add --python-exit-code 1 and print "Complete" at the end of the script as a basis for judgment.

A .blend to store all states

Each run is a completely new Blender process, with no memory of what it did last time. So each script: Open scene/previs.blend → Change → Save with bpy.ops.wm.save_as_mainfile. The script is stored in scripts/ according to the step number, and can be rerun when an error occurs. Find objects by name, not in order; object names are prefixed: SET_ scene, CHR_ character, CAM_S01 camera.

Scenarios and Characters

  • Unit: 1 = 1 meter. The person is about 1.7 meters high, the door is about 2.1 meters high, and the table is about 0.75 meters high.
  • Scenario: Use cubes, planes, and columns to build the ground, walls, doors, windows, and furniture in a uniform light gray color; the doors and windows are hollowed out or marked with dark colors.
  • Character: A capsule (cylinder plus two hemispheres) plus a small cone pointing straight ahead, indicating the direction; each character has a color, and the name text is placed on the top of the head. By default, the text lies flat on the ground, with only a line left when viewed from the side: add an "align to target" constraint to it to face the current camera, or stand it up and face the main camera position.
  • Movement: Keyframe the character's position and orientation without skeletal animation.

Multiple camera positions

  • One camera for each shot CAM_S01, CAM_S02..., the focal length is set according to the shot table.
  • Put all shots on the same timeline: Each shot has a frame interval. Place a timeline mark at the starting point of the interval and bind the camera of this shot. The camera will be cut to wherever the playback ends.
  • Add a "target" constraint to the camera in follow-up shooting. The target is a character or an empty object; push, pull, and traverse the relative camera position to set the keyframe.
  • The frame rate is fixed (default 24), and the duration is rounded when converted into frames and recorded in the shot table.

Rendering

  • Engine: Use Workbench (BLENDER_WORKBENCH), which is the fastest and does not require lights and materials; the coloring is set to by object color, and the character color is written on the object color.
  • Resolution: Preview using half the target aspect ratio (e.g. 960×540 or 540×960).
  • Shot-by-shot rendering: Each shot is rendered separately into the PNG sequence render/S01/, and only those shots are re-rendered when a section is changed.
  • Burning information: Use the metadata that comes with Blender to burn, open the notes, write the shot number and shot size into the notes; keep dialogue in the shot sheet and optional voice track, and update each shot before rendering. By default, the file name, date, and rendering time will also be burned. Turn it off, leaving only notes, frame numbers, time codes, and camera names; increase the font size to a size that can be seen clearly at the preview size.
  • Overhead view: Add an orthogonal camera to take a picture from directly above, render one, and mark the starting and ending of the character and each phase camera position (you can place a small cone at the phase camera position and then render).

Splicing

  • The PNG sequence of each shot is compiled into finished video segments using FFmpeg, and then spliced in order; mixed in when there is temporary voiceover.
  • 3×3 Overview: Take one frame from each of the 9 shots and use FFmpeg’s tile=3x3 to create one.
  • Blender supplies the shot labels. FFmpeg handles encoding and assembly.

Common pitfalls

  • There is no interface when running in the background. Operations that rely on the interface context (selection in the view, certain bpy.ops) will fail. Instead, use the writing method of directly modifying the data.
  • Camera through wall: When the camera position is placed outside the room, the wall should be hidden and not rendered. Please indicate this in the notes of the shot sheet.
  • The keyframe defaults to ease in and out, and uniform motion (tracking, traversing) needs to be changed to linear interpolation.
  • After each step is completed, open the top view or render a frame for self-examination before proceeding to the next step.

Storyboard and asset handoff

When new shot planning is needed, the host uses this workflow directly; reuse a clear existing shot list or edit brief. Record each S01 segment’s script link, visual/information goal, duration/time, subject/action, camera or element motion, exact text/sound, transition and production method. Link E01 target elements to user material and actual resources; keep REF-E01 source identities separate and mark unavailable assets as pending. Check identity, space, action and sound continuity; retime to actual speech. For marketing, label applicable hook/turn/proof/cta roles and the viewer gain, using grounded demonstrations; teaching and narrative work retain their own structure. For an unclear expression, find a real reference and record what to borrow/replace. No separately named planning skill or file is needed.

ขั้นตอนสร้างสรรค์

  1. Flova

    Flova

    Generate temporary dialogue

    Create voiceovers from the script to plan action and shot duration.

  2. Blender

    Blender

    Create an animatic

    Use a simplified scene to demonstrate character blocking and camera movement.

  3. FFmpeg

    FFmpeg

    Assemble the previs

    Join shots in order, mix temporary voice and retain shot numbers.