Long-Form Video Pipeline

The Long-Form Video pipeline generates YouTube-ready videos from 5 to 120+ minutes. It produces chapter-based output with visual consistency, background music, subtitles, thumbnails, and YouTube metadata with chapter timestamps.

Powered by Seedance 2.0 at premium/ultra quality tiers for cinema-grade video generation with native audio and lip-sync.

Pipeline Architecture

                    route_script_input
                          │
            ┌─────────────┼─────────────┐
            │             │             │
        [topic]      [synopsis]    [full_script]
         research      research      parse into
         → structure   → expand      acts
         → generate    → score       → score
            │             │             │
            └─────────────┼─────────────┘
                          │
                    estimate_cost  (--dry-run stops here)
                          │
                generate_anchor_keyframes  (Seedance refs)
                          │
              generate_scene_backgrounds  (green screen)
                          │
                  produce_chapters (parallel × 3)
                    ├─ voiceover (locked voice)
                    ├─ b-roll OR green screen composite
                    └─ assembly
                          │
                  apply_scene_layers  (VFX layer editing)
                          │
              generate_chapter_transitions (Seedance i2v)
                          │
                    concatenate_all
                          │
                  add_background_music (speech ducking)
                          │
                transcribe_and_subtitle
                          │
                  video_output_gate  (ffprobe validation)
                          │
              generate_thumbnail + youtube_metadata
                          │
                  collect_long_form_output

Quick Start

CLI

# Topic mode — research + script generation
fab video/long-form "The Rise and Fall of Theranos" --structure documentary --duration 20min

# Synopsis mode — expand a story idea
fab video/long-form --story "A programmer discovers a security flaw in the world's largest bank" \
  --structure hero_journey --duration 15min

# Full script mode — just produce the video
fab video/long-form --story @my_script.txt --quality premium

# Topic + story angle
fab video/long-form "Chernobyl disaster" \
  --story "From the perspective of a firefighter first on scene" \
  --structure documentary --duration 25min

# Dry run — estimate cost without producing
fab video/long-form "SpaceX" --dry-run --quality premium --duration 10min

Python SDK

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("video/long-form", input={
    "topic": "The Rise and Fall of Theranos",
    "structure": "documentary",
    "duration_target": "20min",
    "quality": "premium",
    "mood": "dramatic",
})
result = fabric.wait_for_run(run["id"])
print(result["output"]["video_path"])
print(result["output"]["youtube_metadata"]["youtube_title"])

TypeScript SDK

import { FabricClient } from "@fabric-platform/sdk";
const fabric = new FabricClient();

const run = await fabric.workflows.runs.submit("video/long-form", {
  topic: "The Rise and Fall of Theranos",
  structure: "documentary",
  duration_target: "20min",
  quality: "premium",
});
const result = await fabric.workflows.runs.waitForRun(run.id);
console.log(result.output.video_path);

Three Input Modes

Mode	Trigger	What happens
Topic	Only topic provided	Full pipeline: deep research, structure selection, script generation
Synopsis	story < 500 words	Research supplements the story idea, LLM expands into full script
Full Script	story > 500 words	Script parsed into acts, video produced directly (no research)

When both topic and story are provided, topic provides the research subject and story acts as the narrative angle.

Script Structures

Structure	Description	Best for
documentary	Rise-fall-aftermath with cold open	Biographical, institutional stories
hero_journey	Monomyth transformation arc	Personal stories, startup journeys
true_crime	Mystery pacing with layered revelations	Crime, investigation, conspiracy
listicle	Ranked items with escalating interest	Top-N lists, comparisons
problem_solution	Pain, failed approaches, solution	Tutorials, how-to content
explainer	Layered complexity with aha moments	Science, economics, complex topics
auto (default)	LLM selects best fit for your topic	When unsure

Visual Controls

Parameter	Values	Default	Description
aspect_ratio	16:9, 9:16	16:9	Landscape (YouTube) or portrait (Shorts)
broll_density	low, medium, high, all_broll	medium	Presenter vs b-roll ratio
visual_style	any string	auto	Visual aesthetic direction
presenter_look	any string	none	Actor description (requires --talking-heads)
include_talking_heads	flag	false	Add AI presenter segments

B-roll density controls how much screen time goes to the presenter vs b-roll footage:

• low — 80% presenter, 20% b-roll (talking head style)
• medium — 50/50 balanced
• high — 80% b-roll, 20% presenter
• all_broll — pure b-roll narration (documentary style, no presenter)

Seedance 2.0 Provider

At premium and ultra quality tiers, the pipeline uses ByteDance’s Seedance 2.0 for video generation. Key advantages:

• Multi-reference input — up to 12 reference images for visual consistency across chapters
• Native audio generation — synchronized audio + video in a single generation
• Phoneme-perfect lip-sync in 8+ languages
• Video extension — smooth transitions between chapters
• Second-by-second timeline prompts — frame-level control within each clip

Anchor keyframes are generated before chapter production and passed as references to every Seedance call, maintaining visual DNA across a 20+ minute video.

Quality Presets

Tier	B-Roll Model	Avatar	Lipsync	Est. Cost (10min)
free	skip (stock)	skip	skip	$0
budget	Veo 3.1 Fast	Kling Avatar	FAL Lipsync	~$2
standard	Veo 3.1 Fast	Kling Avatar	FAL Lipsync	~$5
premium	Seedance 2.0	Seedance 2.0	native	~$8
ultra	Seedance 2.0	Seedance 2.0	native	~$12

Cost Estimation

Use --dry-run to see a cost breakdown before committing:

fab video/long-form "SpaceX" --quality premium --duration 10min --dry-run

Output shows per-component costs: TTS, b-roll clips, transitions, BGM, thumbnail.

YouTube Integration

The pipeline generates YouTube-ready metadata:

• Chapter timestamps — derived from act boundaries, formatted for YouTube descriptions
• SEO title — under 60 characters, optimized for search
• Description — summary + chapter timestamps + related video suggestions
• Tags — 15-20 SEO-optimized tags
• Thumbnail — 1280x720 with hook text overlay

Voice Continuity

The voice ID is locked after the first chapter generates successfully. All subsequent chapters use the exact same voice, preventing drift across a multi-chapter video.

You can also provide your own voice via:

• --voice <voice_id> — use a specific TTS voice ID
• --voice-sample <path> — clone from an audio sample (ElevenLabs or FAL Chatterbox)

Prompt Customization

All LLM prompts are stored as external template files in prompts/long_form/. Override any template by placing a file at ~/.fabric/prompts/long_form/{name}.txt.

Available templates: script_generation, synopsis_expansion, script_parsing, retention_scoring, structure_detection, youtube_metadata, seedance_cinematic.

Input Reference

Field	Type	Default	Description
topic	string	—	Video topic (triggers research mode)
story	string	—	Full script or synopsis
structure	enum	auto	Narrative archetype
duration_target	string	15min	Target duration
quality	enum	standard	Cost tier (free/budget/standard/premium/ultra)
platform	string	YouTube	Target platform
mood	string	auto	Emotional tone
audience	string	general	Target audience
research_depth	int	5	Research depth 1-10
aspect_ratio	enum	16:9	Video aspect ratio
include_talking_heads	bool	false	Include AI presenter
presenter_look	string	—	Actor description
visual_style	string	—	Visual direction
broll_density	enum	medium	B-roll vs presenter ratio
voice_style	enum	—	Voice preset
voice_id	string	—	Explicit TTS voice ID
voice_sample_path	string	—	Audio sample for cloning
subtitle_style	string	—	Subtitle styling
intro_style	enum	cold_open	Intro sequence type
outro_strategy	enum	subscribe_tease	CTA strategy
greenscreen_footage_url	string	—	Green screen footage URL for AI background replacement
scene_layers	array	—	Visual elements to add: [{url, role, placement, description}]
dry_run	bool	false	Estimate cost only

Output Reference

Field	Type	Description
video_path	string	Path to final assembled video
script	object	Full structured script with acts
chapter_paths	string[]	Per-chapter video paths
thumbnail_path	string	Generated thumbnail
youtube_metadata	object	Title, description, chapters, tags
word_count	int	Total script words
estimated_duration_min	float	Estimated duration
cost_estimate	object	Per-component cost breakdown

VFX Compositing

The long-form pipeline supports two VFX modes that activate via input parameters:

Green screen compositing — provide greenscreen_footage_url and the pipeline automatically generates per-act backgrounds from the script’s visual cues, then composites the footage onto each background with camera motion preservation. Falls back to ffmpeg chroma-key when Seedance is unavailable.

fab video/long-form "History of Cinema" \
  --quality premium \
  --greenscreen-footage-url "https://cdn.example.com/presenter.mp4"

Layer-by-layer scene editing — provide scene_layers to add elements (vehicles, weather, objects) into every chapter. Layers are applied sequentially after chapter production for highest quality.

fab video/long-form "Future Cities" \
  --quality premium \
  --scene-layers '[{"url": "https://example.com/drone.jpg", "role": "vehicle", "placement": "sky", "description": "delivery drone"}]'

See the VFX Compositing guide for full documentation, SDK examples, and cost breakdown.

Tip

For short-form video (under 2 minutes), see the AI Shorts Pipeline.