ComfyUITTSFFmpegRemotionHyperframesAI Agents

AI-Assisted YouTube Content Production

4 faceless channels, 1 monetised, $200–550/mo revenue

ProjectComfyUI, TTS, FFmpeg...Back to Projects

Short Demo

Live Demonstration

A quick look at the system in action — input, processing, output.

AI-Assisted YouTube Content Production — Short Demo

Short Demo · 1:15 · 20-60 second demo

[PLACEHOLDER — YouTube ID: REPLACE_ME]

Context

Problem & Motivation

Why this system needed to exist.

Creating consistent YouTube content at scale is a massive operational burden. Research, scripting, voiceover, visuals, editing, thumbnail, metadata, publishing — each video takes 10-20 hours traditionally. For faceless channels, the challenge is maintaining quality without a human presenter.

I asked: Can an AI-assisted pipeline produce watchable, monetisable content end-to-end with minimal human intervention?

The result: 4 channels, 1 monetised, $200-550/mo — proving the model works, while revealing exactly where human taste is still irreplaceable.

Solution

What I Built

The system architecture and core capabilities.

An end-to-end content production system orchestrated by AI agents:

Pipeline Stages

Topic Research & Selection — Agents scan trends, competitor gaps, keyword difficulty, RPM estimates. Output: ranked topic list with projected CPM.
Script Generation — Structured prompts with retention hooks, open loops, chapter markers. Multiple variants A/B tested.
Voice Generation — Fish S2.1 Pro (primary), Edge TTS (fallback). Voice clones from 30-60s reference per language (PT/DE/FR/JA/NL/EN-AU).
Visual Generation — ComfyUI workflows for consistent style. Image → video via SVD/AnimateDiff. Custom LoRAs for brand consistency.
Programmatic Assembly — Hyperframes/Remotion compositions: timed captions, b-roll inserts, transitions, lower-thirds, end screens.
Post-Processing — FFmpeg: loudness normalisation (-14 LUFS), colour grade, watermark, chapter markers.
Metadata & Publishing — Title/description/tags generation, thumbnail variants, scheduled upload via YouTube API.

Channel Portfolio

Channel A (Monetised)

Documentary/History · 8-12 min · $200-550/mo

Channel B

Science Explainers · 5-8 min · Pre-monetisation

Channel C

Tech/AI News · 3-5 min daily · Testing frequency

Channel D

Multilingual (PT/DE/FR) · Localisation pipeline

Architecture

System Architecture

How the components connect and data flows.

The system is a multi-agent pipeline where each stage is a specialised Hermes skill:

┌─────────────────┐
│  TOPIC SCOUT    │  →  Trending topics, RPM estimates, competition
│  (Research)     │
└────────┬────────┘
         ▼
┌─────────────────┐
│  SCRIPT WRITER  │  →  Hook → Chapters → Open loops → CTA
│  (Creative)     │
└────────┬────────┘
         ▼
┌─────────────────┐     ┌─────────────────┐
│  VOICE GENERATOR│     │  VISUAL GENERATOR│
│  (Fish TTS)     │     │  (ComfyUI)      │
└────────┬────────┘     └────────┬────────┘
         │                       │
         └───────────┬───────────┘
                     ▼
          ┌─────────────────┐
          │  COMPOSER       │  →  Hyperframes/Remotion timeline
          │  (Assembly)     │     Captions + B-roll + Transitions
          └────────┬────────┘
                     ▼
          ┌─────────────────┐
          │  POST-PROCESS   │  →  FFmpeg: LUFS, chapters, watermark
          │  (FFmpeg)       │
          └────────┬────────┘
                     ▼
          ┌─────────────────┐
          │  PUBLISHER      │  →  YouTube API, metadata, scheduling
          │  (Automation)   │
          └─────────────────┘

Workflow

End-to-End Workflow

Step-by-step from input to output.

Agent Orchestration (Hermes Skills)

youtube-content

Transcript → summary, thread, blog

ai-video-production

Full pipeline orchestration

faceless-explainer

Text → script → video

hyperframes-core

Composition contract

media-use

Unified media OS

songsee / heartmula

Audio analysis / music gen

Retention Engineering

Script psychology applied systematically (from hyperframes-composition-patterns skill):

Hook at 0-3s: curiosity gap + visual promise
Open loops every 60-90s: "But first..." "What happens next..."
Chapter markers every 2-3 min for retention graph plateaus
Mid-roll at ~60% for ad revenue optimisation
End screen: 20s to next video + subscribe

Technology

Technology Stack

Tools, frameworks, and services used.

Hermes Agents Orchestration

ComfyUI Image/Video Gen

Fish S2.1 Pro TTS Voice (Multilingual)

FFmpeg Post-processing

Hyperframes Programmatic Video

Remotion React Video

OpenMontage Auto-editing

YouTube API Publishing

n8n Scheduling/Automation

Visuals

Screenshots & Diagrams

Real captures from the running system.

Placeholder captures — replace with actual screenshots:

ComfyUI Workflow Graph

Hyperframes Composition

Channel Analytics Dashboard

Script → Video Pipeline

Decisions

Technical Decisions & Trade-offs

Key choices made during development and why.

Fish S2.1 Pro over ElevenLabs

Free tier, HTTP header auth, better multilingual prosody, emotional range. ElevenLabs is higher ceiling but cost-prohibitive at volume.

Hyperframes over pure Remotion

HTML/CSS-based compositions are faster to iterate, easier for agents to generate, lighter renders. Remotion reserved for complex 3D/WebGL needs.

Local ComfyUI over API (Replicate/Fal)

Zero per-generation cost, full workflow control, custom LoRAs, no rate limits. RTX 5070 Ti 12GB handles SDXL + SVD comfortably.

8-12 min target length

Long-form RPM ($6-8 US Education) vs Shorts ($0.05). Watch time compounds. Mid-roll eligibility at 8 min.

Challenges

Challenges & Solutions

What went wrong and how it was resolved.

Challenge: Visual Repetition Kills Retention

Solution: Built ComfyUI "style consistency" LoRA + prompt templates with forced variation seeds. Added mandatory b-roll insert every 45s in Hyperframes composition.

Challenge: TTS Monotone on Long Scripts

Solution: SSML-style chunking: split script into emotional beats, generate per-chunk with prosody tags, cross-fade in FFmpeg. Fish S2.1 handles this better than most.

Challenge: YouTube Algorithm Penalises "Mass Produced"

Solution: Human-in-the-loop at topic selection and thumbnail approval. Agent generates 5 options, human picks 1. This 5-min touch keeps channels in "creator" not "spam" tier.

Challenge: Multilingual Voice Consistency

Solution: Single voice clone per language (30-60s reference). Fish S2.1 maintains speaker identity across languages. Avoid per-video voice training.

Results

Outcome & Lessons Learned

What worked, what didn't, and what I'd do differently.

Revenue reality: $200-550/mo across 1 monetised channel. Not life-changing, but proves the pipeline works end-to-end.

Key Learnings

Content quality floor is achievable automatically; quality ceiling still needs human curation
Thumbnails drive CTR more than title — agent generates 5, human picks 1 (5 min investment)
RPM varies wildly by niche: Education $8, Crime $7.27, DIY $5.30, Gaming $2.40 (2025 data)
Conspiracy/alt-history niches run 25-40% below country rate due to brand safety
Consistency > quality for algorithm; 3x/week beats 1x/week "perfect"
Multilingual expansion (PT/DE/FR/JA/NL) is a force multiplier — same content, new audiences

What's Next

Automated thumbnail A/B testing via YouTube API
Real-time trend ingestion → same-day publish
Shorts pipeline from long-form chapters (Repurpose.io style)
Revenue attribution per topic → smarter topic selection

Note: Revenue figures are approximate monthly ranges from actual channel. No client work or inflated claims. Screenshots are from development — replace with production captures.

Deep Dive

Technical Walkthrough

Extended video covering architecture, code, and implementation details.

AI-Assisted YouTube Content Production — Technical Walkthrough

Technical Walkthrough · 4:30 · 2-5 minute technical deep dive