I Built an AI Video Pipeline That Runs Completely Offline. Here's the Blueprint.

I spent last year building a fully offline AI video generation pipeline. No cloud services, no API keys, no monthly subscriptions — just a used GPU and open-source software.

Here's exactly what I built, what it costs, and the full blueprint so you can build your own.

The Hardware

You don't need a data center. I run everything on:

• GPU: RTX 3090 24GB (used, ~$700)
• RAM: 32GB DDR4
• Storage: 1TB NVMe SSD
• Total cost: ~$900 for the upgrades (I already had the rest of the PC)

If you're on a tighter budget, an RTX 3060 12GB (~$200 used) works for 512x512 video at 24fps. You just can't do 1080p.

The Software Stack

All free. All open source.

Component	Tool	What It Does
Core Engine	ComfyUI	Node-based Stable Diffusion pipeline
Video Model	AnimateDiff	Text/image → short video clips
Background Removal	rembg (RMBG 1.4)	Batch remove image backgrounds
Transcription	faster-whisper	Speech → text
Audio	Piper TTS	Text → voiceover speech
Video Editing	DaVinci Resolve (free)	Final assembly

Step-by-Step Setup

# 1. Install ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt

# 2. Download models
# Put AnimateDiff models in ComfyUI/models/animatediff_models/
# Put motion modules in ComfyUI/models/motion_models/

# 3. Install custom nodes (from ComfyUI Manager)
# - AnimateDiff Evolved
# - ComfyUI-VideoHelperSuite
# - WAS Node Suite

# 4. Install supporting tools
pip install faster-whisper rembg piper-tts

Generating Your First Video

The workflow is simpler than most tutorials make it look:

1. Load the AnimateDiff workflow from the ComfyUI examples
2. Pick a checkpoint (I use majicMIX_realistic_v7)
3. Write a prompt
4. Set frame count to 32 (about 5 seconds at 6fps)
5. Hit "Queue Prompt"

Expected render times (32 frames):

• RTX 3090: ~2 minutes
• RTX 3060: ~5-6 minutes

Quality won't blow you away on the first try. Video prompt engineering is different from images. After weeks of tweaking, these settings worked best for me:

Setting	Value	Why
CFG Scale	7.0	Lower = more natural motion
Sampling steps	25	More = smoother but slower
Motion scale	1.1	0.8 is subtle, 1.5 is chaos

Production Pipeline

Here's how I go from idea to finished video entirely offline:

1. Generate AI clip(s) in ComfyUI (~2 min per 5s clip)
2. Download royalty-free music from Pixabay (free, no attribution)
3. Generate voiceover with Piper TTS (~2s for 30s of audio)
4. Run audio through faster-whisper for captions
5. Assemble everything in DaVinci Resolve
6. Export as 1080p H.264

Total time for a 2-minute video: about 30 minutes. Compare that to rendering on Runway where 2 minutes of video at $0.05/s = $6 every single time. For me, the economics flipped in the first month.

Where It Struggles

I'm not going to pretend local AI video is perfect. Here's where it falls short:

1. Coherent motion beyond 5 seconds: AnimateDiff starts hallucinating around frame 48. Keep clips short and stitch them together.
2. Hands and faces: Same problem as every diffusion model. The smaller the subject in frame, the weirder the hands.
3. Training data bias: The model tends toward certain aesthetics. Getting specific styles (anime, claymation, watercolor) takes significant LoRA training.
4. No real-time generation: If you need instant previews, cloud is still faster.

But for batch production — social media clips, B-roll, background videos — it's absolutely usable today.

Cost Comparison Over 1 Year

Item	Cloud (Runway Pro)	Local
Monthly subscription	$95/mo × 12 = $1,140	$0
Per-second generation	+ variable ($0.05/s)	$0
Hardware (one-time)	$0	$900
Year 1 total	~$1,140+	$900
Year 2 total	~$1,140	$0

After 18 months, local has saved me over $2,000. My data never touched a third-party server. And I can generate as much as I want without watching a credit meter.

The hardest part was the first weekend of setup. Everything after that is just hitting "Queue Prompt."

*I write about local AI tools and workflows at aixhdd.com. Hardware prices shown are from the used market (mid-2026).