Formula Cloning

Formula Cloning lets you reverse-engineer any video’s style — visuals, audio, script structure, and effects — into a reusable formula, then generate new videos on different topics that match the same style exactly.

Feed in a reference video (or two) and Fabric extracts everything: color palette, camera work, hook type, narration pacing, music mood, subtitle style, and color grading. The formula maps directly to the AI Shorts pipeline inputs, so any extracted formula can drive full video generation.

How It Works

                    Reference Video(s)
                          │
                    ┌─────┴─────┐
                    │  Download  │  (yt-dlp: YouTube, TikTok, Instagram)
                    │  + Probe   │  (ffprobe: resolution, fps, duration)
                    └─────┬─────┘
                          │
                    ┌─────┴─────┐
                    │ Transcribe │  (Whisper large-v3, word-level)
                    │  Extract   │  (ffmpeg scene detection + sampling)
                    │ Keyframes  │  (capped at 30 frames)
                    └─────┬─────┘
                          │
              ┌───────────┼───────────┬──────────┐
              │           │           │          │
        analyze_      analyze_    analyze_   analyze_
        visuals       audio       script     effects
       (Gemini      (Gemini)    (Gemini)   (Gemini
        Vision)                             Vision)
              │           │           │          │
              └───────────┼───────────┴──────────┘
                          │
                  synthesize_formula
                  (consensus if 2+ videos)
                          │
                    save_formula
                  (~/.fabric/formulas/)

The four analysis branches run in parallel via Gemini, extracting:

Visuals

Color palette, color temperature, film stock, lighting style, camera focal length, movement, cut pacing, overall aesthetic

Audio

Voice gender/tone/pace, music genre/tempo/mood, volume level, sound effects style

Script

Hook type, emotional trigger, tone, segment structure, narration density, word choice patterns, CTA style, hook template

Effects

Subtitle style/colors/position, hook overlay, transitions, color grade, post-processing effects

Quick Start

Extract a Formula

CLI

# From a YouTube video
fab-workflow formula-extract \
  --input video_paths='["https://www.youtube.com/watch?v=VIDEO_ID"]' \
  --input formula_name="viral-style"

# From a local file
fab-workflow formula-extract \
  --input video_paths='["/path/to/reference.mp4"]' \
  --input formula_name="my-style"

# From multiple videos (extracts what's consistent)
fab-workflow formula-extract \
  --input video_paths='["https://youtube.com/watch?v=abc", "https://youtube.com/watch?v=xyz"]' \
  --input formula_name="brand-style"

Python SDK

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("global/formula-extract", input={
    "video_paths": ["https://www.youtube.com/watch?v=VIDEO_ID"],
    "formula_name": "viral-style",
})
result = fabric.wait_for_run(run["id"])
print(result["output"]["formula_path"])

Generate From a Formula

CLI

fab-workflow formula-shorts \
  --input formula="viral-style" \
  --input topic="Why cold showers boost productivity" \
  -o ./output

# With quality override
fab-workflow formula-shorts \
  --input formula="viral-style" \
  --input topic="5 habits that changed my life" \
  --quality premium \
  -o ./output

Python SDK

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("global/formula-shorts", input={
    "formula": "viral-style",
    "topic": "Why cold showers boost productivity",
})
result = fabric.wait_for_run(run["id"])
print(result["output"]["output_path"])

Tip

Pass -o ./output (a directory, not a file) to get both the video file (ai-short.mp4) and metadata (output.json) saved to that directory.

Extraction Input Parameters

Parameter	Type	Default	Description
video_paths	string[]	required	List of video file paths or URLs (YouTube, TikTok, Instagram)
formula_name	string	"untitled"	Name for the saved formula
formula_path	string	~/.fabric/formulas/{name}.json	Custom save path for the formula
whisper_model	string	"large-v3"	Whisper model for transcription

Generation Input Parameters

All AI Shorts parameters are supported. The formula provides defaults that can be overridden:

Parameter	Type	Default	Description
formula	string	required	Formula name or path to formula JSON
topic	string	required	The subject of the new video
hook	string	""	Override the formula’s hook pattern
mood	string	from formula	Override the formula’s tone
visual_style	string	from formula	Override the formula’s visual aesthetic
quality	string	""	Quality preset (formula handles style, quality handles models)
duration_secs	int	from formula	Override the formula’s pacing

Formula Schema

Formulas are saved as JSON at ~/.fabric/formulas/{name}.json. Every field is optional — missing fields fall back to AI Shorts defaults.

Visuals

{
  "visuals": {
    "color_palette": {
      "dominant_colors": ["#ED4A2C", "#9000FF", "#1E1E1E"],
      "color_temperature": "warm",
      "saturation_level": "vibrant",
      "contrast": "high"
    },
    "film_stock_match": "ARRI ALEXA Mini LF",
    "lighting_style": {
      "source": "mixed",
      "behavior": "soft diffused",
      "color_tone": "contrasting warm red and cool purple/blue tones"
    },
    "camera_work": {
      "dominant_focal_length": "35mm",
      "movement_style": "static",
      "pacing": "moderate"
    },
    "cut_rhythm": {
      "avg_segment_duration_secs": 4.0,
      "pattern": "long takes intercut with graphic overlays"
    },
    "overall_aesthetic": "Modern high-contrast broadcast style with vibrant neon lighting"
  }
}

Audio

{
  "audio": {
    "voice": {
      "gender": "male",
      "tone": "energetic",
      "pace": "fast",
      "style_description": "Fast, energetic commentary with personal interjections"
    },
    "music": {
      "genre": "electronic",
      "tempo": "upbeat",
      "mood_prompt": "Upbeat suspenseful electronic instrumental with driving rhythm",
      "volume_level": "subtle_background"
    },
    "sound_design": {
      "has_sfx": false,
      "sfx_style": "minimal"
    }
  }
}

Script

{
  "script": {
    "hook_type": "curiosity_gap",
    "emotional_trigger": "curiosity",
    "tone": "conversational",
    "structure": {
      "segment_types": ["hook", "problem", "insight", "solution", "cta"],
      "segment_durations": [3, 5, 8, 6, 4],
      "total_segments": 5,
      "alternation_pattern": "actor→broll→actor→broll→actor"
    },
    "narration_density": 1.5,
    "word_choice_patterns": "Direct, fast-paced, informal phrases",
    "cta_style": "Follow for more",
    "hook_template": "[Person] did [unexpected action], and we found the reason why."
  }
}

The hook_type uses the same taxonomy as Hook Generation: question, contrarian, cliffhanger, pain_point, shock, social_proof, story, fomo, challenge, counter_intuitive, curiosity_gap, authority, confession, list, comparison.

Effects

{
  "effects": {
    "subtitles": {
      "style": "karaoke_highlight",
      "position": "center",
      "font_style": "bold_sans",
      "colors": {
        "active": "#FFFF00",
        "inactive": "#FFFFFF",
        "outline": "#000000"
      },
      "size": "large"
    },
    "hook_overlay": {
      "has_text_overlay": true,
      "style": "bold_caps",
      "position": "center"
    },
    "transitions": "cut",
    "color_grade": {
      "ffmpeg_filter_hint": "warm_saturated",
      "lut_description": "Vibrant saturated colors with deep blacks and natural skin tones"
    },
    "post_effects": "none"
  }
}

Multi-Video Consensus

When you pass 2+ reference videos, the extraction pipeline:

1. Runs the full analysis on each video independently
2. Sends all analyses to Gemini with a consensus prompt
3. Keeps only what’s consistent across all videos — that’s the formula
4. Excludes elements that vary between videos

This is useful for extracting a creator’s or brand’s signature style from multiple examples rather than copying a single video.

fab-workflow formula-extract \
  --input video_paths='["/path/to/video1.mp4", "/path/to/video2.mp4", "/path/to/video3.mp4"]' \
  --input formula_name="creator-style"

How Formulas Drive Generation

The formula doesn’t replace the AI Shorts pipeline — it constrains it. Each formula field maps to an existing pipeline input:

Formula Field	Pipeline Input	Effect
script.hook_type	extra_rules	Forces a specific hook pattern
script.emotional_trigger	extra_rules	Sets the emotional through-line
script.tone	mood	Controls narration tone
script.structure	extra_blocks	Constrains segment sequence and durations
script.hook_template	extra_rules	Pattern for the opening line
visuals.overall_aesthetic	visual_style	Unified visual directive for all prompts
visuals.film_stock_match	Shot design override	Overrides continuity brief film stock
visuals.camera_work	extra_rules	Camera/lens constraints for b-roll prompts
audio.voice.gender	voice_gender_override	Forces voice gender selection
audio.music.mood_prompt	music_mood_override	Custom music generation prompt
audio.music.volume_level	bgm_volume	Background music level
effects.subtitles.colors	subtitle_style	Subtitle highlight/inactive/outline colors
effects.color_grade	color_grade_hint	FFmpeg color grading constraints

User-provided inputs always override formula values — the formula is a default, not a lock.

Creating Formulas Manually

You can create a formula by hand without a reference video. Create a JSON file at ~/.fabric/formulas/{name}.json with any subset of the schema fields above.

# List saved formulas
ls ~/.fabric/formulas/

# Use a manually created formula
fab-workflow formula-shorts \
  --input formula="my-custom-style" \
  --input topic="Your topic here"

Supported Video Sources

The extraction pipeline accepts:

• YouTube — https://youtube.com/watch?v=... or https://youtu.be/...
• TikTok — https://tiktok.com/@user/video/... or https://vm.tiktok.com/...
• Instagram — https://instagram.com/reel/...
• Direct URLs — Any https:// link to an .mp4 file
• Local files — Absolute paths to video files on disk

Downloads use yt-dlp and are cached in temp storage during extraction.