Video Production Workflows

Video production workflows handle standalone video generation tasks — face swapping, talking-head avatars, motion transfer, slideshows, Reddit story videos, and formula-driven style cloning. For composed multi-stage pipelines (research → video, hot topics → video), see Composed Pipelines.

Local Model Support

All video workflows support fully local execution with no external API calls. Use quality=local or quality=local-power to run entirely on-device.

Local Video Generation Models

Model	VRAM	Quality	Speed	Notes
`wan:1.3b`	8 GB	Good	Fast	Default for `local` preset
`wan:14b`	24 GB	High	Moderate	WAN 2.1
`wan2.2:14b`	24 GB	Best	Moderate	WAN 2.2 (MoE), default for `local-power`
`ltx-video:2.3`	8 GB	Good	Fast	LTX-2.3 distilled
`ltx-video:2`	8 GB	Good	Fast	LTX-2
`hunyuanvideo`	24 GB	High	Moderate	CUDA only (Tencent)
`open-sora`	16 GB	Good	Moderate	CUDA only (HPC-AI Tech)
`cogvideox:2b`	6 GB	Decent	Slow	CUDA only
`cogvideox:5b`	12 GB	Good	Slow	CUDA only

Local TTS Models

Model	Quality	Voice Cloning	Notes
`kokoro`	High	No	Default local TTS, 6 voices (3M/3F)
`chatterbox`	High	Yes	Zero-shot voice cloning from single sample
`voxtral`	High	No	Apple Silicon only (MLX)
`piper`	Decent	No	Lightweight, CPU-friendly

Local Music Models

Model	VRAM	Quality	Notes
`musicgen-small`	4 GB	Decent	Default for `local` preset
`musicgen-medium`	8 GB	Good	Default for `local-power`
`musicgen-large`	12 GB	High	Best MusicGen quality
`stable-audio-open`	6 GB	High	Stable Audio Open 1.0 via diffusers

Local LLM Models

Model	VRAM	Quality	Notes
`qwen3:8b`	5 GB	Good	Default for `local` preset
`qwen3:14b`	9 GB	High	Default for `local-power`
`qwen3:latest`	19 GB	Best	32B Q4, for high-end hardware
`gemma3:4b`	3 GB	Decent	Lightweight, for `local-light`

Override any model per-run: --input broll_model="wan2.2:14b" or --input bgm_model="stable-audio-open".

Face Swap

Workflow: video/face-swap

Swaps a persona face onto a source image or video using FAL, then uploads the result as a Fabric asset.

CLI
REST

# Direct URLs
fab-workflow video/face-swap \
  -i source_url=https://example.com/dance-still.jpg \
  -i target_url=https://example.com/persona-face.png \
  -v

# From org gallery
fab-workflow video/face-swap \
  -i source_url=https://example.com/dance-still.jpg \
  -i persona_gallery_id=YOUR_GALLERY_ID \
  -i persona_index=0

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/face-swap" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "source_url": "https://example.com/dance-still.jpg",
      "target_url": "https://example.com/persona-face.png"
    }
  }'

Pipeline

resolve_inputs → swap_face (FAL) → upload_result

Input

Parameter	Type	Default	Description
`source_url`	`string`	required	Base image/video to swap the face into
`target_url`	`string`	—	Direct URL to the persona face image
`persona_gallery_id`	`string`	—	Org gallery ID to pull the face from
`persona_index`	`int`	`0`	Which gallery item to use

One of target_url or persona_gallery_id is required.

Output

{
  "output_path": "/tmp/face-swap-result.png",
  "output_asset_id": "018f...",
  "output_url": "https://...",
  "workflow": "face-swap"
}

Avatar (Talking Head)

Workflow: video/avatar

Generates a single lip-synced talking-head video from a portrait image and audio clip. Uses the same provider dispatch as the AI Shorts pipeline but exposed as a standalone workflow.

CLI
REST

# Local files
fab-workflow video/avatar \
  --input actor.path=./portrait.png \
  --input audio.path=./voiceover.mp3 \
  --input prompt="Person speaking to camera, warm lighting"

# With a specific model
fab-workflow video/avatar \
  --input actor.path=./portrait.png \
  --input audio.path=./voiceover.mp3 \
  --input avatar_model="seedance/2.0"

# Using Fabric asset IDs (upload assets first via POST /v1/assets)
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/avatar" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "actor": {"asset_id": "018f..."},
      "audio": {"asset_id": "018a..."},
      "prompt": "Person speaking to camera, warm lighting",
      "avatar_model": "seedance/2.0"
    }
  }'

# Using direct URLs
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/avatar" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "actor": {"url": "https://example.com/portrait.png"},
      "audio": {"url": "https://example.com/voiceover.mp3"},
      "prompt": "Person explaining with natural hand gestures"
    }
  }'

Pipeline

render_avatar (resolve refs → generate talking head → persist)

Input

Parameter	Type	Default	Description
`actor`	`AssetRef`	required	Portrait image (PNG/JPG) — face must be clearly visible
`audio`	`AssetRef`	required	Audio clip to lip-sync (MP3/WAV)
`prompt`	`string`	`"Person speaking directly to camera..."`	Stylistic prompt for avatar generation
`avatar_model`	`string`	`"fal-ai/kling-video/ai-avatar/v2/standard"`	Provider/model ID

AssetRef accepts Fabric asset IDs, HTTP(S) URLs, or local file paths.

Supported Models

Model	Style	Notes
`fal-ai/kling-video/ai-avatar/v2/standard`	Lip-sync + head movement	Default, good quality
`seedance/2.0`	Full-body animation	Higher quality, slower
`bytedance/omnihuman*`	Research-grade	Experimental

Output

{
  "video_path": "/tmp/avatar.mp4",
  "asset_id": "018f...",
  "avatar_model": "fal-ai/kling-video/ai-avatar/v2/standard"
}

Talking Head (End-to-End)

Workflow: video/talking-head

Generates a complete talking-head video from just a topic or script — no pre-existing audio or portrait required. Handles script generation, actor portrait creation, optional voice cloning, TTS, and avatar rendering in a single pipeline. Unlike video/avatar (which requires pre-made audio + portrait), this workflow is fully end-to-end.

CLI
REST

# Minimal — just a topic
fab-workflow video/talking-head \
  --input topic="3 tips for better sleep"

# With explicit script and voice style
fab-workflow video/talking-head \
  --input script_text="Hey everyone, here are three tips for sleeping better..." \
  --input voice_style="warm female" \
  --input presenter_look="professional woman in her 30s, dark hair"

# With voice cloning from a sample
fab-workflow video/talking-head \
  --input topic="3 tips for better sleep" \
  --input voice_sample.path=./my-voice-sample.mp3

# With a pre-existing actor portrait
fab-workflow video/talking-head \
  --input topic="3 tips for better sleep" \
  --input actor.path=./portrait.png \
  --input voice_id="TxGEqnHWrfWFTfGW9XjX"

# From a topic (everything auto-generated)
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/talking-head" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "3 tips for better sleep",
      "duration_secs": 30
    }
  }'

# With voice cloning
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/talking-head" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "3 tips for better sleep",
      "voice_sample": {"url": "https://example.com/my-voice.mp3"}
    }
  }'

# With pre-existing actor + explicit script
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/talking-head" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "script_text": "Hey everyone, here are three tips...",
      "actor": {"asset_id": "018f..."},
      "voice_id": "TxGEqnHWrfWFTfGW9XjX"
    }
  }'

Pipeline

prepare_script
      │
  ┌───┴───┐
resolve   resolve
_actor    _voice
  └───┬───┘
      │
generate_tts → render_talking_head → finalize_output

Input

Parameter	Type	Default	Description
`topic`	`string`	—	Video topic (auto-generates script). Required when `script_text` is empty
`script_text`	`string`	—	Explicit narration text (skips script generation)
`actor`	`AssetRef`	—	Actor portrait image. Auto-generated when omitted
`presenter_look`	`string`	`""`	Actor appearance for AI generation (ignored when `actor` is set)
`voice_sample`	`AssetRef`	—	Audio sample for voice cloning (MP3/WAV)
`voice_id`	`string`	—	Explicit TTS voice ID (overrides voice_sample and auto-selection)
`voice_style`	`string`	`""`	Descriptive voice style (e.g. `"warm female"`, `"deep narrator"`)
`avatar_model`	`string`	`"fal-ai/kling-video/ai-avatar/v2/standard"`	Avatar model override
`duration_secs`	`int`	`30`	Target duration when generating a script
`language`	`string`	`"en"`	Script and TTS language

Output

{
  "video_path": "/tmp/talking-head.mp4",
  "asset_id": "018f...",
  "script_text": "Hey everyone, here are three tips...",
  "voice_id": "TxGEqnHWrfWFTfGW9XjX",
  "avatar_model": "fal-ai/kling-video/ai-avatar/v2/standard"
}

Motion Transfer

Workflow: video/motion-transfer

Animates a character image using a reference video’s motion (dance, gestures, expressions). The character performs the same movements as the driving video.

CLI
REST

# Direct URLs
fab-workflow video/motion-transfer \
  -i source_image_url=https://example.com/persona.png \
  -i driving_video_url=https://example.com/dance-reference.mp4

# With Seedance (full-body, higher quality)
fab-workflow video/motion-transfer \
  -i source_image_url=https://example.com/persona.png \
  -i driving_video_url=https://example.com/dance-reference.mp4 \
  -i motion_model="seedance/2.0"

# From org gallery
fab-workflow video/motion-transfer \
  -i persona_gallery_id=YOUR_GALLERY_ID \
  -i driving_video_url=https://example.com/dance.mp4

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/motion-transfer" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "source_image_url": "https://example.com/persona.png",
      "driving_video_url": "https://example.com/dance-reference.mp4",
      "motion_model": "seedance/2.0"
    }
  }'

Pipeline

resolve_inputs → transfer_motion (FAL/Seedance) → upload_result

Input

Parameter	Type	Default	Description
`source_image_url`	`string`	—	Character image URL
`driving_video_url`	`string`	required	Motion reference video URL
`persona_gallery_id`	`string`	—	Org gallery ID to pull the character from
`persona_index`	`int`	`0`	Which gallery item to use
`motion_model`	`string`	`"fal-ai/live-portrait"`	Provider/model for motion transfer

One of source_image_url or persona_gallery_id is required.

Output

{
  "output_path": "/tmp/motion-transfer-result.mp4",
  "output_asset_id": "018f...",
  "output_url": "https://...",
  "workflow": "motion-transfer"
}

Slideshow / Carousel

Workflow: video/slideshow

Generates listicle carousels with triple output from a single run: individual PNG slides (for native carousel upload), stitched MP4 video, and multi-page PDF (for LinkedIn document carousels). Includes cross-run topic deduplication.

CLI
REST

# Basic
fab-workflow video/slideshow \
  --input topic="5 AI tools that replaced my entire workflow" \
  --input quality=standard

# Full control
fab-workflow video/slideshow \
  --input topic="5 AI tools that replaced my entire workflow" \
  --input niche="productivity" \
  --input quality=premium \
  --input slide_layout=numbered \
  --input mixed_media=true \
  --input num_items=7

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/slideshow" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "5 AI tools that replaced my entire workflow",
      "niche": "productivity",
      "quality": "premium",
      "slide_layout": "numbered",
      "mixed_media": true,
      "num_items": 7
    }
  }'

Pipeline

generate_listicle
       │
  ┌────┼──────────────────────────────────┐
source_  generate_    generate_    generate_
images   voiceover    bgm          video_hook
  └────┼──────────────────────────────────┘
       │
  merge_slideshow_assets → compose_slides → export_pngs →
  assemble_video → generate_pdf → platform_metadata

Quality Presets

Preset	Images	TTS	Music	Script Model
`free`	Stock	Piper (local)	Skip	qwen3:8b (local)
`budget`	AI-generated	FAL Kokoro	Stable Audio	gemini-2.5-flash
`standard`	AI-generated	ElevenLabs	Stable Audio	gemini-2.5-flash
`premium`	AI-generated	ElevenLabs	Stable Audio	gemini-2.5-flash
`local`	Stock	Kokoro (local)	MusicGen (local)	qwen3:8b (local)

Input

Parameter	Type	Default	Description
`topic`	`string`	required	Listicle topic
`niche`	`string`	`""`	Content niche
`num_items`	`int`	`5`	Number of list items
`quality`	`string`	`"standard"`	Quality preset
`slide_layout`	`string`	`"overlay_gradient"`	Layout: `overlay_gradient`, `numbered`, `card`
`slide_duration`	`float`	`3.5`	Seconds per slide in the video
`transition_duration`	`float`	`0.5`	Transition time between slides
`mixed_media`	`bool`	`false`	Generate a video clip for the hook slide (Instagram mixed-media carousel)
`style_hint`	`string`	`""`	Visual style hint for image sourcing
`brand_kit`	`dict`	`{}`	Brand colors, fonts, logo
`language`	`string`	`"en"`	Language for script and voice

Output

{
  "output": {
    "slide_png_paths": ["/tmp/slide_01_hook.png", "/tmp/slide_02_item.png", "..."],
    "video_path": "/tmp/slideshow.mp4",
    "pdf_path": "/tmp/slideshow.pdf",
    "hook_video_path": "/tmp/hook_video.mp4",
    "platform_metadata": {
      "tiktok": {"format": "carousel", "description": "...", "hashtags": ["..."]},
      "instagram": {"format": "mixed_media_carousel", "caption": "..."},
      "youtube_shorts": {"format": "video", "title": "..."},
      "linkedin": {"format": "document", "post_copy": "..."}
    },
    "topic": "5 AI tools that replaced my entire workflow",
    "hook_text": "I deleted 12 apps last month. These 5 replaced all of them.",
    "slide_count": 7
  }
}

Reddit Stories

Workflow: video/reddit-stories

Converts Reddit posts into narrated videos with screenshot overlays on background video — the classic Reddit story video format. Two modes: comments (top comments as individual segments) and story (post body paginated into segments).

CLI
REST

# From a specific Reddit post
fab-workflow video/reddit-stories \
  --input reddit_url="https://reddit.com/r/AskReddit/comments/..." \
  --input mode=comments \
  --input quality=standard

# Search a subreddit for a topic
fab-workflow video/reddit-stories \
  --input subreddit=tifu \
  --input topic="worst cooking disaster" \
  --input mode=story

# Local-only (free TTS, no background video)
fab-workflow video/reddit-stories \
  --input reddit_url="https://reddit.com/r/..." \
  --input quality=free

# From a specific Reddit post
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/reddit-stories" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "reddit_url": "https://reddit.com/r/AskReddit/comments/...",
      "mode": "comments",
      "quality": "standard"
    }
  }'

# Search a subreddit for a topic
curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/reddit-stories" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "subreddit": "tifu",
      "topic": "worst cooking disaster",
      "mode": "story"
    }
  }'

Quality Presets

Preset	TTS	Background Video	Screenshots	Subtitles	Voice Mode
`free`	Piper (local)	None	1x scale	No	Single
`budget`	Kokoro (local)	Random	2x scale	Yes	Alternating
`standard`	ElevenLabs	Random	2x scale	Yes	Alternating
`local`	Kokoro (local)	Random	1x scale	Yes	Alternating

Input

Parameter	Type	Default	Description
`reddit_url`	`string`	—	Direct URL to a Reddit post
`subreddit`	`string`	—	Subreddit to search in
`topic`	`string`	—	Search query within subreddit
`mode`	`string`	`"comments"`	`"comments"` or `"story"`
`max_comments`	`int`	`6`	Max comments to include
`min_score`	`int`	`10`	Min upvote score for comments
`min_length`	`int`	`100`	Min comment length (chars)
`max_length`	`int`	`600`	Max comment length (chars)
`theme`	`string`	`"dark"`	Screenshot theme: `dark`, `light`, `transparent`
`overlay_opacity`	`float`	`1.0`	Screenshot overlay opacity (0.0–1.0)
`tts_model`	`string`	preset	TTS model override
`voice_mode`	`string`	preset	`"single"`, `"alternating"`, or `"random"`
`background_video`	`string`	preset	Video ID, category, `"random"`, or `"none"`
`quality`	`string`	`"standard"`	Quality preset
`platform`	`string`	`"youtube_shorts"`	Target platform
`language`	`string`	`"en"`	Target language

At least one of reddit_url, subreddit, or topic is required.

Formula Shorts

Workflow: global/formula-shorts

Loads a VideoFormula (extracted style DNA from reference videos) and maps its constraints — hook type, segment structure, camera work, color grade, subtitle style, voice, music — onto the AI Shorts pipeline. The formula acts as a style template; the topic is new.

CLI
REST

# Named formula (from ~/.fabric/formulas/)
fab-workflow global/formula-shorts \
  --input formula="my-style" \
  --input topic="Why cold showers boost productivity"

# Formula file path
fab-workflow global/formula-shorts \
  --input formula="/path/to/formula.json" \
  --input topic="The psychology of procrastination"

# With an inline formula dict
curl -X POST "$FABRIC_URL/v1/workflows/run?name=global/formula-shorts" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "formula": "my-style",
      "topic": "Why cold showers boost productivity"
    }
  }'

Pipeline

load_formula → apply_formula_inputs → [ai_shorts_pipeline]

The apply_formula_inputs task maps formula fields to AI Shorts overrides:

Formula Section	Mapped To
`script.hook_type`	Hook style constraint in script generation
`script.tone`	`mood` override
`script.structure.segment_types`	Segment arc (e.g., hook → context → insight → CTA)
`visuals.overall_aesthetic`	`visual_style` with color palette
`visuals.camera_work`	Camera style constraints in b-roll prompts
`audio.voice`	Voice gender and style overrides
`audio.music`	Music mood and volume
`effects.subtitles`	Subtitle highlight/inactive/outline colors, position, size
`effects.color_grade`	FFmpeg filter hint and effects style
`effects.hook_overlay`	Hook text overlay style

Input

Parameter	Type	Default	Description
`formula`	`string` or `dict`	required	Formula name, file path, or inline dict
`topic`	`string`	required	Video topic (passed through to AI Shorts)
All AI Shorts inputs	—	—	Any AI Shorts parameter can be passed; explicit values override formula defaults

Recreate From Format

Workflow: video/recreate-from-format

Takes format DNA and a recreation blueprint from research/reference-format-analysis and generates an original reel plan — script, shotlist, overlay captions, and edit recipe — that follows the same structural format without copying source content. Supports generating N variants per run.

CLI
REST

# Generate a single variant
fab-workflow video/recreate-from-format \
  -i format_dna_json_path="output/analysis/format_dna.json" \
  -i recreation_blueprint_json_path="output/analysis/recreation_blueprint.json" \
  -i topic="5 morning habits for productivity" \
  -i creator_persona="confident lifestyle coach" \
  -i tone="energetic" \
  -i platform="instagram_reels" \
  -o output/recreation/

# Generate 3 variants with TTS
fab-workflow video/recreate-from-format \
  -i format_dna_json_path="output/analysis/format_dna.json" \
  -i recreation_blueprint_json_path="output/analysis/recreation_blueprint.json" \
  -i topic="Why cold showers boost focus" \
  -i creator_persona="health-focused creator, calm authority" \
  -i tone="informative" \
  -i platform="tiktok" \
  -i variants=3 \
  -i include_tts_audio=true \
  -o output/recreation/

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/recreate-from-format" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "format_dna_json_path": "format_dna.json",
      "recreation_blueprint_json_path": "recreation_blueprint.json",
      "topic": "5 morning habits for productivity",
      "creator_persona": "confident lifestyle coach",
      "tone": "energetic",
      "platform": "instagram_reels"
    }
  }'

Pipeline

load_and_extract_constraints → generate_scripts → generate_shotlists_and_overlays → generate_edit_recipes → emit_recreation_artifacts

Per-variant outputs

Each variant gets its own directory (variant_0/, variant_1/, …) containing:

script.json — structured script with per-segment narration, visual intent, emotional intensity
script.md — human-readable formatted script
shotlist.json — beat-by-beat shot plan with framing, motion, composition type
overlay_captions.json — timed on-screen captions matching the reference format’s overlay style
edit_recipe.json — machine-readable edit plan with transitions, pauses, music guidance, emphasis points
voiceover.txt — narration text (if voiceover enabled)
voiceover.wav — TTS audio (if TTS enabled)

Two-workflow flow

reference_reel.mp4
       │
       ▼
research/reference-format-analysis
       │
       ├── format_dna.json
       ├── recreation_blueprint.json
       └── summary.md
               │
               ▼
video/recreate-from-format
       │
       ├── variant_0/script.json
       ├── variant_0/shotlist.json
       ├── variant_0/overlay_captions.json
       ├── variant_0/edit_recipe.json
       └── recreation_manifest.json