Blender Director
ASecurityStyle for edit-video, or on its own: direct a narrated 3D episode in Blender from a script, the way a real two-character sitcom was made entirely in code. Covers the voice pass and timeline, lip sync, set and character staging, cameras that cut on words, dark finished lighting, wireframe reveals, Cycles render settings and timing, and sound that lands on the frame. Needs Blender 4.2 or newer (free). Use when someone wants an animated scene, sitcom episode or character reel made with Blender, ...
- 80 stars
- 0 votes
- 0 copies
- 0 views
- Added November 18, 2026
Works with
Security analysis
100/100Pro scans all 2 files and shows the line behind each finding
npx -y skills add alisadiq-ai/favstash-ai-video-editing-skills --skill blender-director --agent claude-codeAre you the author of Blender Director?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/alisadiq-ai-blender-director)---
name: blender-director
description: "Style for edit-video, or on its own: direct a narrated 3D episode in Blender from a script, the way a real two-character sitcom was made entirely in code. Covers the voice pass and timeline, lip sync, set and character staging, cameras that cut on words, dark finished lighting, wireframe reveals, Cycles render settings and timing, and sound that lands on the frame. Needs Blender 4.2 or newer (free). Use when someone wants an animated scene, sitcom episode or character reel made with Blender, or asks how to make AI 3D videos that don't look cheap."
---
# Blender director
You are the director, the 3D artist and the editor. Every frame is built by a Python script you write and Blender renders. Nothing is generated by a video model, so you control every shot.
The order below is the order that worked. Don't skip ahead to rendering: most of the quality comes from steps 1 to 6.
Target: Blender 4.2 to 4.5, headless (`blender -b --python build.py -- --build`). One build script stages and animates the whole episode from its source files, so any change is a rebuild, not a hand edit.
## 1. Script for the ear, then the eye
- Short lines, one idea each. Two characters trading lines beat one narrator: give the viewer's own question to one character and the answer to the other.
- Write the delivery next to every line, as a short tag ("[deadpan]", "[alarmed, one word]"). TTS engines read these tags.
- Mark the one word per line the picture will react to (a cut, a light, a prop). Those words become your timing anchors.
## 2. The voice pass
- Make one TTS take per line, never the whole script in one go: a bad line is then a cheap retake.
- **Check every take by transcribing it** (Whisper or similar) and comparing the words to the script. TTS doubles words, drops "-ed" and mangles names. Retake until it matches.
- Spell names the way they should sound in the TTS text, and keep the real spelling in the captions.
## 3. The timeline: place, don't stretch
- Trim each take to its speech and match loudness across lines.
- Place the lines on a timeline with designed gaps: a long gap before a punchline's reaction, a short one in rapid back-and-forth. **Never time-stretch a voice.**
- Get word timings from the transcription. Every animation cue keys off them: `W(line, "word")` returns the time a word starts.
- Save `timeline.json` (lines, words, durations, events). The set, the cameras, the music and the captions all read this one file.
## 4. Lip sync and acting
- **Visemes from letters:** spread each word's letters evenly across its duration and map them to a few mouth shapes:
| Shape | Letters |
|---|---|
| AH | a |
| EH | e, i, y and most consonants |
| OO | o, u, w |
| FV | f, v |
| MB | m, b, p |
| REST | silence |
Hold each shape for at least two frames, or the mouth flickers.
- **Look targets:** the listener looks at the speaker, then at whatever the line is about. Smooth head turns by easing toward the target a fifth of the way each frame.
- **Expressions on beats:** a raised brow on a question, a worried look on bad news, deadpan on the dry line. Switch about 0.1 s before the word.
- **Blinks** come every 2.4 to 4.2 s from a seeded random generator, so a rebuild gives the same blinks. Never blink mid-word.
- **Motion:** anticipation before a hop, then a sine arc up and down. Small talking nods while speaking. Nothing starts or stops instantly.
## 5. Set and characters
- Reuse a character kit `.blend` (rigged characters, mouth shapes, expression controls) across episodes. Never edit the kit file: change things in the episode file only.
- **Fit a set to the characters, not the other way round.** If a set was built at real scale and the kit isn't, append the whole set under one empty and give it a single matrix: translate, then scale (match the desk height), then translate. Never rescale a rig.
- After appending objects from another file, read their transform from `matrix_basis`. Their `matrix_world` is stale until the scene updates.
- Clear the spots where the hands and props go: move anything that lands on the laptop or between a character and the camera. Add a front panel under the desk so short legs don't show.
- **Lived-in beats tidy:** a window with a city at night, a lamp, a mug, books, a cork board and a curtain. Leave space around the faces.
## 6. Cameras: cut on words
- **Make a shot list from the timeline:**
- a two-shot to open
- a close-up of the speaker on each line
- the listener's face for the reaction
- a wide shot for a reveal
- an insert of the screen or prop when the line names it
- Cut on spoken words, roughly every 2 to 3 s, about 0.08 s before the line starts.
- **Movement:**
- Every shot pushes in slowly, 5 to 10 % of the distance over its length.
- Use one fast, eased push (ease-out cubic) for an emphasis word.
- Cameras that track a head aim at the eyes, smoothed so they don't jitter.
- **Depth of field on every shot**, focused on whoever the shot is about: f/1.6 to f/2 for close-ups, about f/3 for wide shots. If your units are bigger than metres, open the aperture to compensate.
- Keep camera positions new per episode. Reused angles make platforms flag the video as a duplicate.
## 7. Light: dark and finished, never flat
- **Start from darkness.** Keep a dark room and a dim world, and add motivated light only:
- **key:** a warm, soft practical (laptop or desk lamp) on the faces, around 3400 K
- **rim:** a cool light from the window behind, around 9000 K, so heads separate from the wall
- **hair:** a soft top light from above and behind, which makes hair read (silver hair especially)
- **room fill:** about a third of the key, or nothing at all
- Practicals glow: give the lamp, the screen and the bulbs emission, plus a small light next to each so they really light the set.
- Colour management: AgX with the Medium High Contrast look and a low exposure (around 0.15 to 0.45). If a test render looks like daytime, remove lights instead of lowering exposure.
- Keep materials matte with subtle variation. Give skin a little subsurface. Black clothes need no sheen: sheen turns black cloth grey.
## 8. The moves only 3D can do
These are what prove the video is a real 3D scene, not a filter:
- **Wireframe reveal:**
- Add a Wireframe modifier (`use_replace=True`) and append an emissive material slot. Set `material_offset` to point at it.
- Key `show_render` on and off on the beat.
- Make the wire thickness grow with distance, so far objects still read.
- Hide curves and very dense meshes (hair fuzz) for that window instead of wiring them.
- Snap everything back on one beat.
- **Light dip:** on the word "light", drop every light and the world strength to 7 % for a third of a second, then relight over half a second.
- **Prop pop:** scale from 0 with a 12 % overshoot over 0.22 s, a gentle float while it's up, then shrink away.
## 9. Render
- **Storyboard first:** render one draft frame per shot, plus the key beats, at half size. Fix framing there, not in the full render.
- **Cycles for the finished look:**
- GPU, denoising on, adaptive sampling, persistent data on (it skips rebuilding the scene every frame).
- **Layered hair and fur shells need `transparent_max_bounces` of 64 to 128.** At the default of 8, rays die inside the hair and it renders dark and speckled.
- **Time it before you commit:**
- Render two full-size frames at 16 and 32 samples. Use the hardest shot, the close-up on hair.
- With denoising they usually look the same, and 16 to 20 samples halves the render.
- **Draft, then final:**
- Draft the whole episode at half size and 6 samples, and get the edit approved on that.
- Then render the final at full size.
- **Render to a numbered PNG sequence with overwrite off**, so a crash or a reboot resumes where it stopped. When one late line changes:
1. Rebuild the scene.
2. Move the frames after the change aside.
3. Re-render only the tail.
4. Check that one re-rendered frame from before the change matches the original.
- Run one heavy render at a time on a laptop. Two in parallel run slower than one after the other and can freeze the machine.
## 10. Sound and finish
- **One short sound per visible event:** a click for a light, a tick for each wireframe, a soft pop for a prop, a plop when someone lands. Place each so its loudest attack hits the frame the event first shows (frame = ceil(time × fps)). No whooshes or thuds under dialogue.
- **Music bed:** quiet (about −24 dB under speech), ducked under words, lifting a little on the final beat.
- Normalise to −14 LUFS with true peak under −1.5 dBTP.
- **Captions:** big (around 72 px), short phrases, top third of the frame, inside the phone's safe area. The first line on screen is the text hook and stays from frame 1.
- Encode 1080×1920 at 30 fps, H.264 High, Rec.709, AAC 48 kHz. Encode once from a high-quality master, never re-encode an upload file.
## 11. Before you call it done
1. Does every take match its script word for word?
2. Does every shot cut on a word, with DOF on its subject?
3. Is the room dark with motivated light, and do faces and hair read?
4. Does each reveal or effect land on its word, and snap back cleanly?
5. Does hair render properly (enough transparent bounces)?
6. Did you watch a full draft before rendering the final?
7. Does every sound sit on its event frame, at −14 LUFS?
8. Are the captions inside the safe area, with the hook readable at frame 1?
## Python patterns
```python
import bpy, math
from mathutils import Vector
def smooth_piecewise(points, t):
"""[(time, value), ...] -> eased value at t (smoothstep between keys)."""
if t <= points[0][0]: return points[0][1]
for (a, va), (b, vb) in zip(points, points[1:]):
if a <= t <= b:
u = 0 if b == a else (t - a) / (b - a); u = u * u * (3 - 2 * u); return va + (vb - va) * u
return points[-1][1]
def wire_reveal(objs, f_on, f_off, wire_mat, eye=Vector((0, -3, 1.8))):
"""Turn meshes into glowing wire from frame f_on to f_off."""
done = set()
for o in objs:
if o.type != 'MESH': continue
me = o.data
if me.name not in done: me.materials.append(wire_mat); done.add(me.name)
m = o.modifiers.new('wire', 'WIREFRAME'); m.use_replace = True
m.thickness = max(.0035, .0022 * (o.matrix_world.translation - eye).length)
m.material_offset = len(me.materials) - 1
for f, on in ((1, False), (f_on - 1, False), (f_on, True), (f_off, False)):
m.show_render = on; o.keyframe_insert('modifiers["wire"].show_render', frame=f)
for fc in o.animation_data.action.fcurves:
for k in fc.keyframe_points: k.interpolation = 'CONSTANT'
def push_camera(cam, look_at, f0, f1, amount):
"""Slow push toward the subject over a shot."""
base = cam.location.copy(); fwd = (Vector(look_at) - base).normalized()
for f in (f0, f1):
cam.location = base + fwd * amount * (f - f0) / max(1, f1 - f0); cam.keyframe_insert('location', frame=f)
cam.location = base
def cycles(scene, samples=20, percent=100):
scene.render.engine = 'CYCLES'; c = scene.cycles
c.device = 'GPU'; c.samples = samples; c.use_denoising = True; c.use_adaptive_sampling = True
c.transparent_max_bounces = 128 # layered hair/fur shells
c.max_bounces = 6; scene.render.use_persistent_data = True
scene.render.resolution_x, scene.render.resolution_y, scene.render.resolution_percentage = 1080, 1920, percent
```
Files in this skill
- SKILL.md
- agents/openai.yaml
Attribution
Related skills
View all in ai-agents →Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.