A Christmas Video for the Team – Half Cloud, Half Homelab

🌐 Also available in: Deutsch · Français · Español

For our monitoring team’s internal advent calendar I wanted a short video: in the data centre the LEDs flip from green to red, alarm, heartbeat, rewind – and in the end Santa is to blame, of course, having tripped over a “Wi-Fi cable”. The result is just under 100 seconds of film that I polished over more than 70 versions with Claude Code as assistant director.

More interesting than the film itself is the question which parts came from the cloud and which ran on my own hardware. In short: the images that matter came from the cloud – everything else ran locally.

☁️ From the cloud: the scenes with people

For the actual video scenes I used ArtCraft, a subscription that gives access to several commercial models:

• Seedance 2.0 (video): data centre, monitoring centre, LED close-up, Santa with the annoyed reindeer. A 4-second clip in 720p costs about 100 credits.
• Nano Banana Pro (image editing): the group photo – colleagues as silhouettes, Santa hats, snow. 15 credits per image.

I did not drive ArtCraft through its interface but through an MCP server: Claude Code writes the prompt, uploads reference images and fetches the result. With one fixed rule: before every job Claude states the cost and waits for my “Start” – a submitted job cannot be called back.

The takeaway: for realistic people in motion the cloud is clearly ahead today. But every clip costs real money, and credits go faster than you think – in the end the balance was empty before the last idea was done.

🖥️ Local: everything around the images

A machine with an RTX graphics card (8 GB of video memory) did the rest:

• ComfyUI as the frame for the local models
• MMAudio: sound effects that match the picture – snorting reindeer, rustling cables, the stumble
• ACE-Step: music drafts in different styles; the final video uses a royalty-free epic track from a music library
• Chatterbox (text-to-speech): the narrator
• Whisper: a cross-check that “Icinga” and “Zabbix” are pronounced clearly
• RIFE: AI in-between frames for the slow motion
• ffmpeg and Python: cutting, transitions, text, iris wipe, typewriter credits, the terminal ending with systemctl status weihnachten and loudness mastering

Locally, attempts cost only power and time. That is where most iterations happen: one voice sounded tinny, another had an English accent, the music was too poppy – twelve variants of one line are no problem locally.

The takeaway: the cloud delivers the expensive raw footage, the homelab turns it into a film. That keeps costs low and control with you.

🎞️ Where local is not enough – and where it is better

At first I tried to create the video scenes locally as well, with Wan 2.1/2.2. With 8 GB of video memory the results were crude: primitive faces, unnatural walking, clips too short.

The other way round, the “smarter” AI was sometimes the worse choice:

• Upscaling to 1080p: Real-ESRGAN sharpened edges but made surfaces waxy. Classic upscaling with light sharpening looked more natural.
• In-between frames: ffmpeg’s motion estimation produced block artefacts on fast hands and on the LED patterns of the racks. RIFE was much cleaner – only when the LED flips from green to red did RIFE invent a wrong frame too. That spot now uses the original.

The takeaway: “more AI” is not automatically better. Every tool invents details that do not exist somewhere – and you have to find them.

👀 Checking what you cannot see

Claude can neither play nor hear video. So checks went the long way round: single frames and contact sheets, loudness per scene (the scenes were up to 15 dB apart before), speech recognition for pronunciation and hidden vocals, image comparison for exact joins.

Still I found errors no measurement showed: a female “Aaa” that was really an artefact at the end of the voice recording, a grey, noisy background after an image filter, a text overlay flying through the room during a crossfade. And once a 2.5-second extension simply did nothing – without an error message.

The takeaway: measuring does not replace watching. Tools find levels and formats, people find what looks “odd”.

🧾 The bill

• Cloud: a monthly subscription plus credits – almost all of it for a handful of 4-second clips and a few image edits.
• Local: no running costs, but GPU time – upscaling the whole film takes about 45 minutes.
• Time: the biggest item. Most versions went into timing, sound and details, not into new scenes.

The takeaway: the money is in a few cloud clips, the time in many local iterations.

🎯 The lessons

1. Cloud for people, local for everything else – an 8 GB homelab cannot do realistic people in motion today.
2. Confirm the cost before every cloud job – submitted is submitted.
3. Save every version under a new name – more than once an older one was the better one.
4. Use AI tools selectively – sometimes the classic method is more honest than the model.
5. Watch and listen yourself – measurements show whether something is correct, not whether it works.