mtv-maker

name	mtv-maker
version	2.1.0
description	Full end-to-end MTV music video creator. From a song concept and optional character reference photo, produces a complete cinematic MTV: (1) writes lyrics and generates music with ACE-Step, (2) generates cinematic scene images, (3) animates with I2V (Grok/Seedance), (4) assembles clips with audio and crossfade transitions, (5) transcribes audio for accurate lyric timing, (6) burns synced lyric subtitles + opening/ending credits with branding. Use when user says "make an MTV", "create a music video", "generate MTV", "/mtv", or describes a song they want turned into a full visual music video.

MTV Maker v2 — I2V Pipeline

End-to-end pipeline: song concept + character photo → music (ACE-Step) → scene images (image_to_image) → I2V video clips (Grok/Seedance) → assembled MTV with synced bilingual lyrics + credits.

Inputs

The user provides:

1. Song concept — genre, mood, style, voice, language, any reference artist/style
2. Character reference photo (optional) — used for image_to_image to generate consistent scene photos. If none provided, generate original scene images with text_to_image.
3. MTV length preference — full song or chorus-only clip (default: 30-45s chorus clip)

Pipeline Mode

ALWAYS use I2V (image-to-video). NEVER use Ken Burns for MTV production — it looks cheap and static. Every scene must have real AI-generated motion.

Mode	Description	Time	Quality
Photo + I2V	Photos animated by Grok or Seedance AI	~45 min	Cinematic

Phase Overview

Phase	What	Tool/Skill	Output
1	Song creation	acestep skill	Full MP3
2	Chorus extraction	ffmpeg	Trimmed MP3 clip
3	Scene image generation	image_to_image or text_to_image	6 cinematic photos (16:9)
4	I2V animation	grok-video-gen or seedance-api	6 video clips (~5-6s each)
5	Video assembly + audio	ffmpeg xfade	Raw MTV with music
6	Lyric transcription	video-processor (Groq Whisper) or manual	SRT with timestamps
7	Caption burn + credits	ffmpeg + ASS subtitles	Final MTV

---

Phase 1: Song Planning & Generation

Use the acestep skill (which in turn uses acestep-songwriting for lyric craft).

Workflow

1. Plan the song — decide genre, BPM, key, duration, vocal style, lyric theme
2. Write complete lyrics — structured with [Intro], [Verse], [Chorus], [Bridge], [Outro] tags
3. Generate with ACE-Step — always --batch 2 for two versions
4. User picks preferred version

Key Parameters

• Full song: 150-200s duration
• Chorus clip: user selects the best ~30-45s segment after listening
• Always generate 2 versions for A/B comparison

Output

• {song_name}_v1.mp3, {song_name}_v2.mp3

---

Phase 2: Chorus Extraction

After user picks a version and specifies the time range:

# CRITICAL: Use -ss BEFORE -i for accurate MP3 seeking
ffmpeg -y -ss {START} -i "{full_song}.mp3" -t {DURATION} \
  -af "afade=t=in:st=0:d=2,afade=t=out:st={DURATION-2}:d=2" \
  -b:a 192k \
  "{song_name}_chorus.mp3"

CRITICAL BUG FIX: When trimming MP3 files, always place -ss BEFORE -i (input seeking). Placing -ss after -i causes silent output due to MP3 frame boundary issues. This was a hard-won lesson.

Verification

# Always verify the trimmed clip is not silent
ffmpeg -i "{song_name}_chorus.mp3" -af "volumedetect" -f null - 2>&1 | grep mean_volume
# mean_volume should be between -30 dB and 0 dB. If < -80 dB, it's silent — redo the trim.

---

Phase 3: Scene Image Generation (6 Images)

Planning the Visual Storyboard

Design 6 scenes that follow an emotional arc matching the lyrics. Each scene should:

• Correspond to a specific lyric section
• Progress the emotional narrative (e.g., intimacy → heartbreak → acceptance)
• Vary in composition: mix close-ups, medium shots, and wide shots
• Build from intimate/close to distant/wide for a "proximity to distance" narrative arc

Storyboard Template

#	Lyric Section	Scene Description	Emotion	Shot Type
1	Opening line	[setting + character + action]	[emotion]	Close-up / Medium
2	Next section	...	...	Medium
3	Building tension	...	...	Medium
4	Emotional peak	...	...	Close-up / Medium
5	Bridge/reflection	...	...	Medium / Wide
6	Final line	...	...	Wide / Silhouette

Image Generation Rules

With character reference photo (typical case):

• Use image_to_image tool with the reference photo
• Prompt must be cinematic — think Wong Kar-wai, Sofia Coppola, Fincher
• 16:9 aspect ratio, 2K resolution minimum
• FACE MUST BE LARGE AND CLEAR — use medium or medium-close shots so face fills at least 15-20% of the frame. If the face is too small (full-body wide shot), I2V will lose facial identity mid-generation, causing cross-eyes, morphing, or identity drift. This is the #1 cause of bad I2V output.
• Face must be front-facing or 3/4 angle — the image will be used as I2V reference, so face stability matters
• NO accessories that clash with the mood — match props to MTV emotion
• Lighting should be dramatic and atmospheric: neon reflections, candlelight, moonlight, golden hour, rain

Without character reference (abstract/scenic MTV):

• Use text_to_image tool
• Same cinematic quality requirements

Image Prompt Structure

Cinematic [shot type] of [character description] in [setting].
[Emotional atmosphere description]. [Lighting description].
[Character pose/action — must be subtle, not posed].
16:9 widescreen composition, shallow depth of field, film grain,
[color palette]. Shot on Arri Alexa Mini, anamorphic lens flare.

DO NOTs for Images

• NO stiff poses — character should look natural, candid
• NO cluttered backgrounds — MTV scenes need clean, atmospheric backdrops
• NO bright/cheerful lighting for sad songs (match mood!)
• NO tiny figures lost in the frame — character should be clearly visible
• Face must be recognizable for I2V consistency

Naming Convention

• mtv_photo_01_{scene_keyword}.jpg through mtv_photo_06_{scene_keyword}.jpg

User Approval Flow

Show images one by one for approval. If user says "do it on your own" or similar, generate all 6 autonomously using the storyboard plan.

---

Phase 4a: Ken Burns Animation (Default — Photo Mode)

Convert each still photo into a ~7s cinematic clip using ffmpeg zoompan. This is fast, reliable, and looks great for emotional/melancholy MTVs.

Ken Burns Presets

Vary the motion per scene for visual interest:

Scene	Motion	ffmpeg filter
1	Slow zoom in	zoompan=z='min(zoom+0.0015,1.5)':d=168:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'
2	Pan left→right	zoompan=z='1.3':d=168:x='iw*0.1+iw*0.4*(on/168)':y='ih/2-(ih/zoom/2)'
3	Slow zoom out	zoompan=z='max(zoom-0.0015,1.0)':d=168:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'
4	Pan right→left	zoompan=z='1.3':d=168:x='iw*0.5-iw*0.4*(on/168)':y='ih/2-(ih/zoom/2)'
5	Zoom in + pan up	zoompan=z='min(zoom+0.001,1.4)':d=168:x='iw/2-(iw/zoom/2)':y='ih*0.6-(ih/zoom/2)'
6	Static + fade	zoompan=z='1.2':d=168:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'

Command Template

# d=168 = 7 seconds at 24fps. Adjust per desired clip length.
ffmpeg -y -loop 1 -i "{photo}.jpg" \
  -vf "zoompan=z='min(zoom+0.0015,1.5)':d=168:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)',scale=1920:1080" \
  -r 24 -t 7 -c:v libx264 -preset fast -crf 20 \
  "mtv_video_01_{keyword}.mp4"

Important: -loop 1 is required for still images. Output must be at least as long as needed for the assembled MTV.

---

Phase 4b: I2V Video Generation (Optional — Animated Mode)

Tool Selection

Primary: grok-video-gen skill (recommended default)

• Fast, free, good motion quality
• Output: ~752×416 (will be upscaled in Phase 5)
• Use --video --aspect 16:9 -r {photo_path} flags

Alternate: seedance-api skill

• Higher resolution, better face consistency
• Output: native 1280×720
• Slower (1-4 min per clip)
• Use --ref-image {photo_path} --duration 5

I2V Prompt Guidelines

Every I2V prompt must carry the emotional DNA of the ENTIRE storyboard, not just describe one scene in isolation. If the MTV is about heartbreak, every prompt should breathe heartbreak — the grief, the stillness, the loss.

Motion must match the MTV mood:

MTV Mood	Motion Style	Example Prompt
Sad/melancholy	Slow, subtle, minimal	"Subtle breathing, slight head tilt, wind gently moves hair, slow blink"
Energetic/upbeat	Dynamic, expressive	"Dancing, spinning, arms moving, hair flowing with movement"
Reflective/nostalgic	Medium, contemplative	"Slowly walking, looking around, hand touches surface gently"
Dramatic/powerful	Bold, sweeping	"Turns toward camera, dramatic wind, fabric flowing, strong pose"

Critical I2V Rules:

• SLOW motion for sad songs — no fast walking, no catwalk strut
• Maintain distance — if camera is too close, face degrades. For walking scenes, camera tracks backward to maintain distance
• Describe motion, not appearance — I2V inherits appearance from the photo. Focus prompt on WHAT MOVES and HOW
• Camera movement matters — specify: "camera slowly tracks backward", "camera static", "slight camera drift to the right"
• Match motion to storyboard context — if the storyboard calls for stillness, prompt for stillness; if it calls for singing, let her sing. The storyboard is the source of truth for what each scene's motion should be

Gacha Management (CRITICAL)

I2V is probabilistic. The same prompt can produce wildly different results. Follow this workflow:

1. Generate raw I2V clip
2. Show raw clip to user for review BEFORE assembly — never assemble blindly
3. Common I2V artifacts to watch for:
   • Character starts talking/moving lips (most common)
   • Eyes cross or drift mid-clip (caused by small face in reference)
   • Purposeful gazing instead of vacant stare
   • Unnatural head movement or body distortion
4. If rejected, regenerate — keep previous take as backup (sometimes the next roll is worse)
5. Only proceed to assembly after all 6 clips are approved
6. Iterate on the weakest link — don't redo everything, identify the one scene breaking coherence and fix just that

Grok I2V Command

SKILL_DIR="/Users/zanearcher/.claude/skills/grok-video-gen"
npx -y bun ${SKILL_DIR}/scripts/main.ts --video \
  -r "{photo_path}" \
  --aspect 16:9 \
  "{motion_prompt}" \
  --output "{output_path}" \
  --timeout 600

Seedance I2V Command

python3 ~/.claude/skills/seedance-api/scripts/seedance_worker.py \
  --prompt "{motion_prompt}" \
  --ref-image "{photo_path}" \
  --duration 5 \
  --output-dir "{session_dir}"

Naming Convention

• mtv_video_01_{scene_keyword}.mp4 through mtv_video_06_{scene_keyword}.mp4

---

Phase 5: Video Assembly

Step 1: Upscale & Slow Down Clips

Grok outputs ~752×416. Upscale to 1920×1080 and apply 0.75x slowdown for cinematic feel:

for i in 01 02 03 04 05 06; do
  ffmpeg -y -i "mtv_video_${i}_*.mp4" \
    -vf "setpts=1.33*PTS,scale=1920:1080:flags=lanczos" \
    -r 24 -c:v libx264 -preset fast -crf 20 \
    "clip_${i}_upscaled.mp4"
done

If using Seedance (1280×720), still upscale to 1920×1080 but skip the slowdown unless desired:

ffmpeg -y -i input.mp4 \
  -vf "scale=1920:1080:flags=lanczos" \
  -r 24 -c:v libx264 -preset fast -crf 20 \
  output_upscaled.mp4

Step 2: Check Clip Durations

for f in clip_*_upscaled.mp4; do
  echo "$f: $(ffprobe -v quiet -show_entries format=duration -of csv=p=0 "$f")s"
done

Step 3: Crossfade Assembly

Calculate total duration: N clips × avg_clip_duration - (N-1) × crossfade_duration

Target: match audio clip duration. Adjust clip speed or add/remove clips as needed.

# 6 clips with 1-second crossfade transitions
ffmpeg -y \
  -i clip_01_upscaled.mp4 \
  -i clip_02_upscaled.mp4 \
  -i clip_03_upscaled.mp4 \
  -i clip_04_upscaled.mp4 \
  -i clip_05_upscaled.mp4 \
  -i clip_06_upscaled.mp4 \
  -filter_complex "
    [0:v][1:v]xfade=transition=fade:duration=1:offset={OFF1}[v1];
    [v1][2:v]xfade=transition=fade:duration=1:offset={OFF2}[v2];
    [v2][3:v]xfade=transition=fade:duration=1:offset={OFF3}[v3];
    [v3][4:v]xfade=transition=fade:duration=1:offset={OFF4}[v4];
    [v4][5:v]xfade=transition=fade:duration=1:offset={OFF5}[v5];
    [v5]fade=t=in:st=0:d=2,fade=t=out:st={TOTAL-2}:d=2[vout]
  " \
  -map "[vout]" -c:v libx264 -preset fast -crf 20 -r 24 \
  "mtv_video_assembled.mp4"

Offset calculation:

• OFF1 = duration_of_clip1 - 1
• OFF2 = OFF1 + duration_of_clip2 - 1
• OFF3 = OFF2 + duration_of_clip3 - 1
• ... and so on

Step 4: Add Audio

ffmpeg -y -i mtv_video_assembled.mp4 -i "{chorus_clip}.mp3" \
  -c:v copy -c:a aac -b:a 192k \
  -shortest \
  "mtv_with_audio.mp4"

This produces the clean base video — never burn captions onto a previously captioned video.

---

Phase 6: Lyric Timing via Transcription

Method 1: Groq Whisper (Recommended)

Use the video-processor skill's transcriber to get accurate timestamps:

GROQ_API_KEY="{key}" python3 ~/.claude/skills/video-processor/scripts/transcriber.py \
  "{chorus_clip}.mp3"

This outputs {name}_original.srt with segment timestamps.

Method 2: Manual Mapping

If transcription is unavailable or inaccurate (common with sung Chinese), manually map known lyrics to approximate timestamps by:

1. Dividing the audio duration by number of lyric lines
2. Adjusting based on musical structure (verses vs chorus timing)

Why ASR Is Mandatory (Never Use Input Lyrics as Captions)

CRITICAL: The lyrics you feed into ACE-Step are a prompt, not a transcript. The AI model interprets them loosely — it may rearrange words, skip lines, change phrasing, add ad-libs, or alter sentence structure. The generated song's actual vocals will often differ from the input lyrics.

Therefore: NEVER use the ACE-Step input lyrics directly as captions. Always run ASR (Groq Whisper) on the final generated audio to transcribe what was actually sung. This gives you:

1. Accurate timestamps — synced to the real audio
2. Actual phrasing — what the model actually sang, not what you asked it to sing

Lyrics Correction After ASR (MANDATORY)

Whisper will misrecognize some sung words (especially names, slang, or stylized delivery). After transcription:

1. Read the Whisper output (words JSON or SRT)
2. Listen to the audio and compare — fix obvious mishears (e.g., "y'all" → "Yao", "undecided" → "sends it")
3. Keep all timestamps intact — Whisper timing is accurate even when text recognition is wrong
4. The corrected words JSON becomes the source of truth for caption generation

Caption Line Breaking (CRITICAL)

Each caption line MUST be one complete, meaningful sentence. Never break mid-sentence.

The caption_video.py script now uses Whisper segment boundaries for line grouping (segments = natural sentence breaks). But when manually building ASS files, follow these rules:

DO	DON'T
"Chalk on her hands and fire in her eyes"	"her eyes Edge of the"
"Edge of the mountain touching the sky"	"the sky First time hanging but"
"Teresa Yao don't mess around"	"around SHE SENDS IT no"

Each lyric line should appear and disappear as a complete unit. Analyze lyrics grammatically — identify sentence boundaries, not arbitrary word counts.

Caption Font & Styling (MTV Standard)

• Font: Impact or Bebas Neue — thick, bold, high-contrast. NOT thin fonts like PingFang SC or Helvetica Neue.
• Size: 50-56pt minimum for 1280x720, 70-80pt for 1920x1080. Captions must be readable at a glance.
• Outline: 3-4px black outline + 1-2px drop shadow for legibility on any background.
• Chorus lines: Use a distinct style (larger font, different color like gold/cyan) to differentiate from verses.

Outro Best Practices

• Pure black background — not a video frame.
• Fade in/out — all text must animate in and out. Never snap on/off.
• Differentiate credits from title text:
  • Title ("SHE SENDS IT"): large, white, bold.
  • Attribution labels ("Credit to", "Powered by"): small, grey, subtle.
  • Names ("Mavis", "EnConvo"): medium, distinct color (gold, cyan), bold — visually separate from the title above.
  • Stagger fade-in timing: title first, credits delayed by 0.5-1s.

Crossfade Transitions

Always use crossfade between intro→main and main→outro. 1s fade duration is standard. Use ffmpeg xfade=transition=fade:duration=1:offset=X.

---

Phase 7: Captions & Credits (ASS Format)

ASS File Structure

[Script Info]
Title: {Song Title} MTV
ScriptType: v4.00+
PlayResX: 1920
PlayResY: 1080
WrapStyle: 0

[V4+ Styles]
Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
Style: Title,PingFang SC,100,&H00FFFFFF,&H000000FF,&H00000000,&H80000000,1,0,0,0,100,100,15,0,1,5,0,5,50,50,80,1
Style: OpenCredit,PingFang SC,48,&H00FFFFFF,&H000000FF,&H00000000,&H80000000,0,0,0,0,100,100,8,0,1,3,0,5,50,50,60,1
Style: LyricCN,Noto Serif TC,56,&H00FFFFFF,&H000000FF,&H60000000,&HA0000000,0,0,0,0,100,100,5,0,1,3,5,2,50,50,100,1
Style: LyricEN,PingFang SC,32,&H80FFFFFF,&H000000FF,&H50000000,&HA0000000,0,1,0,0,100,100,3,0,1,2,3,2,50,50,55,1
Style: EndCredit,PingFang SC,62,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,1,0,0,0,100,100,10,0,1,0,0,5,50,50,80,1
Style: Brand,PingFang SC,42,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,0,0,0,100,100,6,0,1,0,0,5,50,50,30,1

Caption Layout

Bilingual lyrics (default):

• Chinese (LyricCN): 56pt Noto Serif TC, white with dark outline + shadow, bottom-center
• English (LyricEN): 32pt PingFang SC italic, semi-transparent white, below Chinese line
• Both use \fad(800,600) for smooth transitions

Title card (dedicated black frames, first 5-6 seconds):

• Title card MUST be on dedicated black frames, NOT overlaid on video content
• Song title: 100-120pt bold uppercase, centered, \fad(2000,1200)
• Artist/brand credit below: 48pt regular weight, \fad(1500,1000)
• Give the title card breathing room — at least 5-6 seconds
• Format: {Artist Name} on one line, EnConvo on next

End credits (on black frames, last 6+ seconds):

• Main credit: 62pt bold, 词 · 曲 · 演唱 {Artist}
• Brand line: 42pt, Powered by EnConvo
• Both use \fad(1500,3000) — long 3-second fade-out
• Must last at least 6 seconds and extend to the final frame

Timing Rules (CRITICAL)

Element	When	Rule
Title + OpenCredit	0:00 - 0:04.5	Must fade out COMPLETELY before first lyric
LyricCN + LyricEN	Synced to audio	NO credits visible during lyrics
EndCredit + Brand	Black frames only	Start at song_end + 1.5s, end at final frame

Black Frame Ending

Append 8 seconds of black frames + fade audio out 2s before song ends:

SONG_DUR=$(ffprobe -v quiet -show_entries format=duration -of csv=p=0 mtv_with_audio.mp4)
FADE_START=$(echo "$SONG_DUR - 2" | bc)
SOURCE_FPS=24  # MUST match source video fps

# Fade audio
ffmpeg -y -i mtv_with_audio.mp4 \
  -af "afade=t=out:st=${FADE_START}:d=2" \
  -vn -c:a aac -b:a 192k audio_faded.m4a

# Extend video with black frames (SAME FPS as source!)
ffmpeg -y \
  -i mtv_with_audio.mp4 \
  -f lavfi -i "color=black:s=1920x1080:r=${SOURCE_FPS}" \
  -filter_complex "[0:v][1:v]concat=n=2:v=1:a=0[outv]" \
  -map "[outv]" -c:v libx264 -preset fast -crf 20 -r ${SOURCE_FPS} \
  -t $(echo "$SONG_DUR + 8" | bc) video_extended.mp4

# Extend audio with silence
ffmpeg -y \
  -i audio_faded.m4a \
  -f lavfi -i "anullsrc=r=48000:cl=stereo" \
  -filter_complex "[0:a][1:a]concat=n=2:v=0:a=1[outa]" \
  -map "[outa]" -t $(echo "$SONG_DUR + 8" | bc) -c:a aac -b:a 192k audio_full.m4a

# Mux clean (no captions yet)
ffmpeg -y -i video_extended.mp4 -i audio_full.m4a \
  -c:v copy -c:a copy -shortest muxed_clean.mp4

CRITICAL: Black frame fps MUST match source video fps. Mismatch causes wrong duration.

Final Burn (LAST STEP)

ffmpeg -y -i muxed_clean.mp4 \
  -vf "ass=captions.ass" \
  -c:v libx264 -preset fast -crf 18 \
  -c:a copy \
  "{SongTitle}_MTV_final.mp4"

NEVER burn captions onto a previously captioned video — captions stack and become unreadable. Always burn onto the clean source.

---

Common Pitfalls & Lessons Learned

Problem	Cause	Fix
Trimmed MP3 is silent	-ss placed after -i	Put -ss BEFORE -i for input seeking
Face degrades in I2V	Camera too close, or fast motion	Keep distance, slow motion, camera tracks backward
Cross-eyed / identity drift in I2V	Face too small in reference photo	Use medium-close shots, face must fill 15-20% of frame
I2V motion doesn't match vibe	Prompt describes generic motion, not storyboard emotion	Every I2V prompt must carry the emotional DNA of the whole storyboard
Lyrics too bold/heavy	Bold weight + thick outline	Use regular weight (not bold) for elegance, 3px outline max
I2V looks like a fashion show	Prompt describes appearance, not motion	Focus on subtle motion: breathing, wind, slow turn
Catwalk walk in sad MTV	Default I2V motion is energetic	Explicitly prompt: "slow defeated walk", "subtle movement"
Subtitle timing wrong	Whisper misrecognizes sung lyrics	Keep Whisper timestamps, replace text with correct lyrics
Credits overlap lyrics	ASS timing miscalculation	Verify: OpenCredit fades out before first lyric line
Double captions	Burned onto already-burned video	Always use clean (no-caption) source for final burn
Framerate mismatch in concat	Black frames at 30fps, source at 24fps	Check source fps with ffprobe, match exactly
Video too large for sharing	High bitrate + long duration	Use -crf 20 for clips, -crf 18 for final
Ken Burns motion jerky	zoompan fps mismatch	Always set -r 24 on output and d= to match: d=fps×seconds
Ken Burns black borders	Image resolution too small	Use images ≥1920×1080; add scale=1920:1080 before zoompan
Caption breaks mid-sentence	Line grouping by word count, not meaning	Use Whisper segments for line breaks; manually verify each line is a complete phrase
Captions too thin/small	Using PingFang SC or Helvetica at small size	Use Impact or Bebas Neue at 50pt+ with 3-4px outline
Outro text snaps on/off	No fade animation on drawtext	Use ffmpeg alpha expression for fade in/out on every text element
Credits blend with title	Same font/color/size for everything	Differentiate: title=white/large, labels=grey/small, names=gold or cyan/medium
Captions don't match audio	Used ACE-Step input lyrics as captions	ACE-Step interprets lyrics loosely — always ASR the final audio with Groq Whisper, then correct mishears

---

Delivery

1. Save final MTV in session directory
2. Present with Deliverable tool
3. Optionally send via Telegram using im_channels/telegram_actions/reply with files array
4. Caption with song title

---

Quick Reference: Full Command Sequence

1. acestep skill → generate song (2 versions)
2. User picks version + time range
3. ffmpeg -ss BEFORE -i → trim chorus with fade in/out
4. ffmpeg volumedetect → verify not silent
5. image_to_image × 6 → cinematic scene photos (16:9, 2K+)

--- I2V MODE (always) ---
6. grok-video-gen --video × 6 → I2V clips (review each raw clip before proceeding!)
7. ffmpeg upscale + slow → 1920×1080, 0.75x speed

--- ASSEMBLY ---
7. ffmpeg xfade → crossfade assembly
8. ffmpeg mux → add audio → mtv_with_audio.mp4 (CLEAN SOURCE)
9. Groq Whisper → transcribe for timing
10. Correct lyrics in SRT
11. Write ASS file (bilingual lyrics + title card + end credits on black frames)
12. Check source fps: ffprobe → match black frame fps exactly
13. ffmpeg: audio fade-out 2s before end, extend video + audio with 8s black/silence
14. ffmpeg mux → muxed_clean.mp4 (still no captions)
15. ffmpeg ass burn onto muxed_clean.mp4 → {SongTitle}_MTV_final.mp4