VFX Compositing

Fabric provides three AI-powered compositing capabilities that collapse traditional multi-step VFX pipelines into single inference calls. All three use Seedance 2.0’s reference-to-video system under the hood, with automatic fallbacks when Seedance is unavailable.

These capabilities work as standalone SDK stages, as tasks within the long-form and ai-shorts pipelines, or as building blocks for custom workflows.

Pipeline Architecture

                      ┌─────────────────────┐
                      │  Multi-View Character │
                      │  generate 4 angles    │
                      └─────────┬─────────────┘
                                │
          ┌─────────────────────┼─────────────────────┐
          │                     │                     │
    Green Screen          Layer Editing         B-Roll Generation
    footage + bg        base video + elements    (enhanced refs)
    → composite         → sequential apply       → ref2v with 4 views
          │                     │                     │
          └─────────────────────┼─────────────────────┘
                                │
                         Final Video

Multi-View Character Consistency

Traditional character consistency in AI video uses a single reference image. Multi-view generation produces 4 perspective views (front, 3/4, profile, back) from a single character photo, giving Seedance dramatically stronger consistency signals across different camera angles and shots.

How It Works

A single character photo is passed to Seedance ref2v with angle-specific prompts
The first frame of each generated video is extracted as a still image
All 4 views are uploaded to fal CDN and stored as _multiview_character_refs
Downstream b-roll and avatar calls pass all 4 views as role: "subject" references
Seedance uses @Image1-@Image4 role assignments for each angle

Usage

Multi-view is automatic in the ai-shorts pipeline when using premium/ultra quality and an actor image is provided. No additional configuration needed.

# Automatic: premium/ultra with actor image triggers multi-view
fab ai-shorts "AI replacing developers" \
  --quality ultra \
  --actor-image ./my_actor.jpg

# Long-form also benefits
fab video/long-form "The Metaverse" \
  --quality premium \
  --presenter-look "young tech executive in a navy blazer"

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("video/ai-shorts", input={
    "topic": "AI replacing developers",
    "quality": "ultra",
    "actor_image_url": "https://example.com/actor.jpg",
    # multi-view character prep runs automatically
})

from fabric_workflow_sdk.stages.seedance import (
    generate_multiview_character,
    build_multiview_reference_roles,
)

# Generate 4 perspective views
views = await generate_multiview_character(
    "https://cdn.example.com/actor.jpg",
    num_views=4,
    view_angles=["front", "3/4 left", "profile right", "back"],
    resolution="720p",
)
# views = [{"url": "...", "angle": "front"}, ...]

# Build role assignments for downstream ref2v calls
roles, refs = build_multiview_reference_roles(
    character_views=views,
    environment_urls=["https://cdn.example.com/bg1.jpg"],
)
# roles = {"@Image1": "character front view ...", "@Image2": "character 3/4 view ...", ...}
# refs = [{"url": "...", "role": "subject"}, {"url": "...", "role": "subject"}, ...]

Multi-View vs Single Reference

Approach	References	Consistency	Cost
Single ref (default)	1 character image	Good for front-facing shots	1 image gen
Multi-view (premium/ultra)	4 character images	Strong across all angles	~$0.20 (4 x $0.05)

Multi-view is most impactful for videos with diverse camera angles, multiple scenes, or character movement. For static talking-head videos, single reference is sufficient.

Green Screen Compositing

Upload green screen footage and a background image — Seedance keys out the green screen, preserves the original camera motion, and composites the subject onto the new background in one step.

How It Works

Foreground video is passed as a motion reference (camera trajectory preservation)
Background image is passed as an environment reference
Seedance handles keying, edge preservation, and lighting integration
If Seedance fails, falls back to ffmpeg chromakey filter + overlay composite

Usage

# Long-form video with green screen footage
fab video/long-form "The History of Visual Effects" \
  --quality premium \
  --greenscreen-footage-url "https://cdn.example.com/presenter_greenscreen.mp4"

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("video/long-form", input={
    "topic": "The History of Visual Effects",
    "quality": "premium",
    "greenscreen_footage_url": "https://cdn.example.com/presenter_greenscreen.mp4",
    # Pipeline auto-generates per-act backgrounds and composites
})

from fabric_workflow_sdk.stages.seedance import greenscreen_composite_seedance

# Composite green screen footage onto a background
result_path = await greenscreen_composite_seedance(
    foreground_video_url="https://cdn.example.com/presenter_greenscreen.mp4",
    background_image_url="https://cdn.example.com/office_background.jpg",
    prompt="Professional office setting, warm lighting",
    preserve_camera_motion=True,
    duration=10,
    aspect_ratio="16:9",
)

Pipeline Integration

When greenscreen_footage_url is provided to the long-form pipeline:

generate_scene_backgrounds — generates one background image per act from the script’s visual cues
produce_chapters — for each chapter, composites the green screen footage with the act’s background instead of generating b-roll
Falls back to standard b-roll generation for any chapter where compositing fails

Fallback Chain

Seedance ref2v (motion + environment refs)
  → ffmpeg chromakey + overlay composite
    → Seedance i2v with background as start frame

Layer-by-Layer Scene Editing

Add visual elements (vehicles, weather effects, objects) from reference images into existing footage. Each element is placed at a specific spatial position while the original camera work is preserved.

How It Works

Base video is passed as a motion reference (camera anchoring)
Each element image is passed as a typed reference (subject, weather, vehicle, environment)
Spatial placement directives tell Seedance where to integrate each element
Layers are applied sequentially by default for highest quality — each layer builds on the previous result

Usage

# Long-form with scene layers
fab video/long-form "Future of Transportation" \
  --quality premium \
  --scene-layers '[
    {"url": "https://example.com/cybertruck.jpg", "role": "vehicle", "placement": "background right", "description": "silver cybertruck"},
    {"url": "https://example.com/snow.jpg", "role": "weather", "placement": "atmospheric", "description": "light falling snow"}
  ]'

from fabric_platform import FabricClient

fabric = FabricClient()
run = fabric.run_workflow("video/long-form", input={
    "topic": "Future of Transportation",
    "quality": "premium",
    "scene_layers": [
        {
            "url": "https://example.com/cybertruck.jpg",
            "role": "vehicle",
            "placement": "background right",
            "description": "silver cybertruck driving on highway",
        },
        {
            "url": "https://example.com/snow.jpg",
            "role": "weather",
            "placement": "atmospheric",
            "description": "light falling snow",
        },
    ],
})

from fabric_workflow_sdk.stages.seedance import layer_edit_seedance

# Add a cybertruck and snow to existing footage
result_path = await layer_edit_seedance(
    base_video_url="https://cdn.example.com/highway_scene.mp4",
    layers=[
        {
            "url": "https://cdn.example.com/cybertruck.jpg",
            "role": "vehicle",
            "placement": "background right",
            "description": "silver cybertruck",
        },
        {
            "url": "https://cdn.example.com/snow.jpg",
            "role": "weather",
            "placement": "atmospheric",
            "description": "light falling snow",
        },
    ],
    prompt="Highway scene at dusk with light snowfall",
    duration=10,
    aspect_ratio="16:9",
)

Layer Specification

Each layer is a dict with these fields:

Field	Type	Required	Description
`url`	string	yes	Image URL of the element to add
`role`	string	yes	`subject`, `vehicle`, `weather`, or `environment`
`placement`	string	no	Spatial directive: `foreground left`, `background center`, `atmospheric`, etc.
`description`	string	yes	What the element is: `red cybertruck`, `falling snow`, `neon sign`

Sequential vs Parallel Application

By default, layers are applied sequentially — each layer sees the result of all prior layers, producing the most coherent composite. For workflows with many layers, consider:

Mode	Quality	Speed	When to use
Sequential (default)	Highest	Slower (N calls)	1-3 layers, quality-critical
Parallel (planned)	Good	Faster (1 call + composite)	4+ layers, speed-critical

Parallel layer application will be available when sub-graph fan-out ships (Phase 33).

Cost Reference

Operation	Model	Cost per call
Multi-view (per view)	`bytedance/seedance-2.0/multiview`	$0.05
Multi-view (4x bundle)	`seedance/multiview-4x`	$0.20
Green screen composite	`bytedance/seedance-2.0/greenscreen-composite`	$0.15
Green screen (fast)	`bytedance/seedance-2.0/fast/greenscreen-composite`	$0.08
Layer editing	`bytedance/seedance-2.0/layer-edit`	$0.15
Layer editing (fast)	`bytedance/seedance-2.0/fast/layer-edit`	$0.08
ffmpeg fallback	local	$0.00

Example Cost Breakdown

Workflow	Multi-View	Green Screen	Layers	Total Added
ai-shorts (ultra)	$0.20	—	—	+$0.20
long-form 5ch (premium + greenscreen)	$0.20	5 x $0.15	—	+$0.95
long-form 5ch (premium + 2 layers)	$0.20	—	5 x 2 x $0.15	+$1.70

Quality Presets

These features are available at the premium and ultra quality tiers:

Tier	Multi-View	Green Screen	Layer Editing
free/budget/cheap	—	—	—
standard	—	—	—
premium	automatic	via input	via input
ultra	automatic	via input	via input

Protobuf Contracts

Typed stage I/O contracts are defined in proto/fabric/stages/v1/stages.proto:

MultiViewInput / MultiViewOutput / MultiViewImage
GreenScreenInput / GreenScreenOutput
LayerEditInput / LayerEditOutput / SceneLayer

These contracts are optional — the API always accepts arbitrary JSON — but provide autocompletion and type checking in SDKs.

Long-Form Video Pipeline — green screen and layer editing integrate here
AI Shorts Pipeline — multi-view character prep integrates here
Shot Design & Consistency — cinematography system used by all video generation
Model Configuration — quality presets and provider selection

VFX Compositing

Pipeline Architecture

Multi-View Character Consistency

How It Works

Usage

Multi-View vs Single Reference

Green Screen Compositing

How It Works

Usage

Pipeline Integration

Fallback Chain

Layer-by-Layer Scene Editing

How It Works

Usage

Layer Specification

Sequential vs Parallel Application

Cost Reference

Example Cost Breakdown

Quality Presets

Protobuf Contracts

Related