Csv landscape video pipeline

CSV Landscape Video Pipeline (16:9)

When to Use

Use when users want to convert a CSV file of practice/exam questions (typically 50-100 questions) into a single landscape 16:9 educational video with:

• NVIDIA green/black branding (#76B900 on #050805)
• Professional question presentation (question card → options → answer reveal → explanation)
• TTS voiceover (Sarvam AI English voice or fallback TTS)
• Single long-form video (~10-20 minutes for 60 questions)
• Production-quality output ready for YouTube/education platforms

Prerequisites

• Python 3.10+
• Manim Community Edition v0.19+
• LaTeX (TeX Live / MiKTeX / MacTeX)
• ffmpeg
• Python packages: pandas, edge-tts (fallback), requests, tqdm
• Optional: Sarvam AI API key for premium TTS

NVIDIA Brand Colors (Official)

Role	Hex	Usage
Background	#050805	Scene background (near-black)
Primary Green	#76B900	NVIDIA brand green - accents, highlights, correct answers
Panel Background	#101510	Question/option panel backgrounds
Text Primary	#F5F7F5	Primary text (questions, options)
Text Muted	#B8C2B8	Secondary text (explanations, metadata)
Correct Green	#76B900	Correct answer highlight
Incorrect Red	#E84D3D	Incorrect answer highlight
Accent Blue	#76B900	Progress bars, accents

Video Specifications (Landscape 16:9)

Parameter	Value
Resolution	1920×1080 (Full HD)
Aspect Ratio	16:9
Frame Rate	30 FPS (production), 15 FPS (draft)
Quality	-qh (1080p60) production, -ql (480p15) draft
Duration	~10-15 seconds per question → ~10-15 min for 60 Q
Audio	48kHz stereo, TTS voiceover

CSV Expected Format

Column	Required	Description
Question	Yes	Question text
Question Type	Yes	multiple-choice or multi-select
Answer Option 1..6	Yes	Option text (up to 6)
Explanation 1..6	Optional	Per-option explanation
Correct Answers	Yes	Comma-separated indices (1-based)
Overall Explanation	Optional	Full explanation
Domain	Optional	Category/domain tag

Pipeline Stages

CSV INPUT → PARSE → PLAN → RENDER SCENES → STITCH → ADD AUDIO → FINAL MP4

Stage 1: Parse CSV (parse_csv.py)

• Read CSV with pandas
• Validate required columns
• Clean text (escape LaTeX special chars)
• Output: questions.json + plan.md

Stage 2: Plan Scenes (plan.md)

Generate a scene plan with:

• Scene sequence: Intro → Q1 → Q2 → ... → Q60 → Outro
• Per-question timing: ~10-15s each
• Color palette, typography, layout specs
• TTS script per question

Stage 3: Generate Manim Scenes (generate_scenes.py)

One scene class per question + intro/outro:

• IntroScene - Title card with exam name, question count
• QuestionScene_N - Question card → Options → Reveal → Explanation
• OutroScene - Summary, score, call-to-action

Stage 4: Render (render.sh)

manim -qh --format=mp4 script.py IntroScene QuestionScene_1 QuestionScene_2 ... OutroScene

Stage 5: Stitch (stitch.sh)

ffmpeg -f concat -safe 0 -i concat.txt -c copy stitched.mp4

Stage 6: Add TTS Audio (add_audio.py)

• Generate TTS for each question's narration script
• Concatenate audio files
• Mix with video via ffmpeg

Stage 7: Final Output

final_output.mp4 - Ready for upload

Scene Design (Landscape 16:9)

Layout (1920×1080)

┌─────────────────────────────────────────────────────────────┐
│  [NVIDIA Logo]  NCP-AAI Practice Test 1              Q 1/60 │  ← Top bar (100px)
├─────────────────────────────────────────────────────────────┤
│                                                             │
│   ┌─────────────────────────────────────────────────────┐  │
│   │  Question text (wrapped, centered, large font)      │  │  ← Question card (center)
│   │                                                     │  │
│   └─────────────────────────────────────────────────────┘  │
│                                                             │
│   ┌─────────────┐  ┌─────────────┐  ┌─────────────┐        │  ← Options grid (2x2 or 3x2)
│   │  A. Opt 1   │  │  B. Opt 2   │  │  C. Opt 3   │        │
│   └─────────────┘  └─────────────┘  └─────────────┘        │
│   ┌─────────────┐  ┌─────────────┐  ┌─────────────┐        │
│   │  D. Opt 4   │  │  E. Opt 5   │  │  F. Opt 6   │        │
│   └─────────────┘  └─────────────┘  └─────────────┘        │
│                                                             │
│   ┌─────────────────────────────────────────────────────┐  │  ← Answer reveal panel (bottom)
│   │  ✓ Correct Answer: B                                │  │
│   │  Explanation: [Overall Explanation text...]         │  │
│   └─────────────────────────────────────────────────────┘  │
│                                                             │
│                    [Progress: ████████░░ 1/60]              │  ← Progress bar
└─────────────────────────────────────────────────────────────┘

Animation Sequence Per Question (~12-15s)

Time	Animation	Duration
0.0-1.5s	Fade in question card + top bar	1.5s
1.5-3.0s	Stagger fade-in option cards (A→B→C→D...)	1.5s
3.0-3.5s	Pause for reading	0.5s
3.5-5.0s	Highlight correct answer (green glow), dim others	1.5s
5.0-6.5s	Slide up answer panel with explanation	1.5s
6.5-12s	Hold for TTS narration (explanation)	~5-6s
12-13s	Fade out all, progress bar advance	1s

TTS Integration

Primary: Sarvam AI English Voice

• API: https://api.sarvam.ai/text-to-speech
• Voice: en-IN-neel (English male) or en-IN-ananya (English female)
• Requires: SARVAM_API_KEY environment variable

Fallback: edge-tts (Microsoft Edge TTS)

• Voice: en-US-GuyNeural (male) or en-US-AriaNeural (female)
• No API key needed, works offline

TTS Script Per Question

"Question {n}: {question_text}. 
 Option A: {opt_a}. Option B: {opt_b}. Option C: {opt_c}. Option D: {opt_d}.
 The correct answer is {correct_letter}. {overall_explanation}."

Usage

# 1. Parse CSV and generate plan
python parse_csv.py "NCP-AAI_Practice_Test_1.csv"

# 2. Review plan.md, adjust if needed

# 3. Generate Manim scenes
python generate_scenes.py

# 4. Render draft (fast)
manim -ql script.py IntroScene QuestionScene_1 QuestionScene_2 ... OutroScene

# 5. Render production (slow, high quality)
manim -qh script.py IntroScene QuestionScene_1 QuestionScene_2 ... OutroScene

# 6. Stitch scenes
bash stitch.sh

# 7. Generate TTS audio
python add_audio.py --tts sarvam  # or --tts edge

# 8. Final output: final_output.mp4

Project Structure

csv-video-pipeline/
├── input/
│   └── NCP-AAI_Practice_Test_1.csv
├── parse_csv.py
├── generate_scenes.py
├── add_audio.py
├── script.py              # Generated Manim script
├── plan.md                # Generated scene plan
├── questions.json         # Parsed questions
├── concat.txt             # ffmpeg concat list
├── render_draft.sh
├── render_prod.sh
├── stitch.sh
├── tts_scripts/           # Generated TTS text files
├── audio/                 # Generated WAV/MP3 files
├── media/                 # Manim output
│   └── videos/script/1080p60/
└── final_output.mp4       # Final deliverable

Configuration (config.yaml)

video:
  resolution: "1920x1080"
  fps: 30
  quality: "high"  # high|medium|low
  question_duration: 14  # seconds per question

branding:
  background: "#050805"
  primary: "#76B900"
  panel_bg: "#101510"
  text_primary: "#F5F7F5"
  text_muted: "#B8C2B8"
  correct: "#76B900"
  incorrect: "#E84D3D"

tts:
  provider: "sarvam"  # sarvam|edge
  sarvam_voice: "en-IN-neel"
  edge_voice: "en-US-GuyNeural"
  sample_rate: 48000

intro:
  title: "NCP-AAI Practice Test 1"
  subtitle: "60 Questions • Agent Architecture & Development"
  duration: 5

outro:
  title: "Practice Test Complete"
  subtitle: "Review your answers and improve!"
  duration: 5

Verification Checklist

Before final delivery, verify:

• Video resolution: 1920×1080, 30fps
• Duration: ~12-15 minutes for 60 questions
• Audio: Clear TTS, synced with visuals
• No text overflow/cutoff in any scene
• Correct answer highlighting works
• Progress bar increments correctly
• NVIDIA branding consistent throughout
• Final file size reasonable (<2GB for 1080p30)

Troubleshooting

\n\n| Issue | Fix |\n|-------|-----|\n| LaTeX errors in text | Escape _, %, $, #, &, {, }, ~, ^, \\ in parse_csv.py |\n| Text overflow | Reduce font size, increase panel height, or split long questions |\n| TTS too fast/slow | Adjust question_duration in config, or add SSML <break> tags |\n| Render too slow | Use -ql for draft, -qm for medium, only -qh for final |\n| Audio sync drift | Ensure TTS sample rate matches video (48kHz), use ffmpeg -af aresample=48000 |\n| IndentationError: unexpected indent in generate_scenes.py | The skill file's helper functions section (lines 100-175) must use 4-space indentation, not 8. Lines 101-174 inside generate_script() should be indented by 4 spaces relative to the function. |\n| SyntaxError in generated script.py (unterminated string) | The escape_latex function in generate_scenes.py writes LaTeX escape dict to the generated script. Use exactly 4 backslashes in the skill source: \"'\\\\\\\\\\\\\\\\': r'\\\\\\\\textbackslash{}'\" (produces '\\\\\\\\': r'\\\\textbackslash{}' in script.py). 8 backslashes in source = double-escaped output = broken syntax. |\n| CRLF line endings causing parse errors | Normalize to LF: dos2unix generate_scenes.py or sed -i 's/\\r//g' generate_scenes.py before running. |\n\n## References

Issue	Fix
LaTeX errors in text	Escape _, %, $, #, &, {, }, ~, ^, \ in parse_csv.py
Text overflow	Reduce font size, increase panel height, or split long questions
TTS too fast/slow	Adjust question_duration in config, or add SSML <break> tags
Render too slow	Use -ql for draft, -qm for medium, only -qh for final
Audio sync drift	Ensure TTS sample rate matches video (48kHz), use ffmpeg -af aresample=48000

Related Skills

• csv-quiz-video — Alternative approach: renders one clip per question (resumable), beat-driven audio sync via add_sound, preview mode, per-domain voiceover. Better for long-form quiz videos where resumability and voice variety matter.
• manim-video — Core Manim animation patterns for educational content.
• manim-reels — Vertical 9:16 reel patterns (adaptable to landscape).

Architecture Comparison

Aspect	csv-quiz-video	csv-landscape-video-pipeline (this skill)
Scene structure	One clip per question (resumable)	One giant script with all scenes
Audio sync	Beat-driven per clip (add_sound + wait)	Post-render TTS concat + ffmpeg mix
Resumability	✅ Skip existing clips	❌ Full re-render on change
Preview mode	✅ --preview (first N questions)	Manual scene selection
Parallelization	Possible (independent clips)	Sequential
Complexity	Higher (orchestrator + env vars)	Lower (single script)