xFlow

Self-hosted orchestration for channel content. Local inference, procedural assembly, human review at every gate.

A self-hosted system that takes a topic or a live news source and returns a finished After Effects project for a human to review and publish - narration, stills, video, music, timeline. Generation runs locally: ComfyUI for image, video, TTS, and recognition; local models for prompting and metadata. Nothing leaves the machine unless it has to.

Two spreadsheets run it. One is a per-channel DSL - voices, visual style, LoRAs, music, timeline parameters - so a new channel is a new row, not new code. The other is a state machine tracking where every video sits across stages, so scripting, production, and upload batch independently with a human QA gate between each. The timeline is assembled procedurally in Houdini, not cut by hand: movement, transitions, and overlays fall out of channel parameters as an After Effects script.

The diagram shows the system at two levels. Overview is the four-stage backbone; Detailed expands each block with its steps and stack.

01 Script Writing

News or non-news branch routes through an agent chain. Output is script JSON and a new row in the tracking sheet.

n8n LLM newsapi.org

Awaiting review

02 Production

Audio, stills, recognition, video, cull and fix, timeline, AE script.

n8n ComfyUI LLM Whisper FFMPEG Houdini

Ae Timeline Ready.

03 Manual Validation

Run generated JS in After Effects, review composition, sign off, final render.

After Effects human

Ready for upload

04 Uploadseparate workflow

Publish all approved videos

n8n YouTube API

Channels, config DSL

Drives every block of the pipeline.

Identity

Channel ID, Name, Language. Channel Info is a system prompt injected into every LLM step where channel character matters.

Script behaviour

Narration style, News channel, Podcast. Switches Stage 1 between news-pull, podcast, and free-topic modes.

Stills

Stills provider, Visual style, Image prompt suffix, Core LoRAs, Animation style. Which ComfyUI workflow runs, what style is prepended or appended, which LoRAs load.

Cover thumbs

Cover provider, Cover style, Cover prompt suffix, Cover LoRAs. Same controls, separate generation path for YouTube thumbnails.

Voice and audio

Voice TTS, Voices, Music folder, Audio settings. TTS engine, voice characters used, music pool, ComfyUI-side audio settings (voice seed, etc.).

Timeline, WF

Houdini settings, Workflow settings. Houdini timeline-assembly parameters, per-channel workflow overrides.

01 Script Writing

News mode pulls latest from newsapi.org.newsapi
Non-news routes to Non-fiction, Fiction, or Podcast.channel-driven
Short or Long form selector.channel-driven

Agent chain: Planner, Writer, Editor.

Write script JSON to dataset.
Append title and metadata to a tracking-sheet row.Sheets

Awaiting review

02 Production

Create dataset folders, place music, prepare config JSON.
YouTube metadata: titles, descriptions, tags, categories.Cloud, Local LLM
Cover thumbnail prompts.Local LLM
TTS from script using channel's Voices.ComfyUI
Whisper transcription for precise timings.ComfyUI
Audio mix: voice plus random-shuffled music master.FFMPEG
Block splitter turns text into scene blocks by time.
Still prompts per block.
Stills and Cover thumbnails (xN).ComfyUI
Image recognition pass over every still.ComfyUI, Local LLM
Video prompts built from recognition (not intent).
Video generation.ComfyUI
Upscale (optional).
Cull bad shots.manual, opt-in
Smart Fix fills empty slots from remaining shots.
Timeline assembly produces AE JS script.Houdini

Ae Timeline Ready.

03 Manual Validation

Run JS in After Effects. Composition assembled.
Human review and sign off.
Final render.

Ready for upload

04 Uploadseparate workflow

Scan the tracking sheet for approved videos.
Publish approved videos to YouTube.n8n, YouTube API

Design notes

STATE MACHINE

Tracking sheet decouples stages

Batch-write scripts, then batch-run production, then batch-upload. Human QA gate between any pair of stages. Pause or resume across runs.

FEEDBACK LOOP

Recognition drives video prompt

Stills are generated, then described, then the description (not the original intent) drives the video prompt. Compensates for drift between the generated still and the clip.

QUALITY RECOVERY

Cull and Smart Fix

Failure is assumed. Bad shots are pulled by eye; Smart Fix then repopulates the timeline from remaining usable generations without re-running the whole scene.

PROCEDURAL TIMELINE

Houdini as timeline engine

Timeline composition is procedural, not editorial. Movement, transitions, overlays driven by channel's Houdini settings. Output is a JS script for After Effects.

CONFIG-DRIVEN

Channels as DSL

One spreadsheet drives every block. Channel-by-channel customization without code changes. New channel equals new row plus asset references.

Worked example

One episode end to end. Breakfast World on a micro-world channel where tiny civilizations live inside everyday food.

01 Script

A topic seed enters the agent chain (Planner, Writer, Editor) and comes out a script.

An educational survey of the inhabited world that is served, each morning, upon a single breakfast tray - a geographical tour of its regions, peoples, and customs, from the Pancake Plateaus to the Bacon Ridges, recorded in the manner of a 1950s field study by observers who know the world will be gone by nine o'clock.

02 Audio

TTS voiceover mixed with a shuffled music bed into one master track that drives the timeline.

master mix, voiceover plus music bed (excerpt)

03 Covers

Five thumbnail options generated from the script; one chosen.

Published thumbnail: Empires of Breakfast

04 Stills

A still is generated per block, then described; the description drives its video prompt. Click any frame to play its clip.

05 Result

Composition assembled in After Effects, reviewed, rendered, and published as Breakfast World.

Contact

Get in touch