AI video generation for businesses, without sending your assets anywhere
An AI video generation pipeline that runs on machines you own or rent: open-source models, ComfyUI, upscaling and a style trained on your brand. Short clips for promotional content, no subscription to an outside tool, no files uploaded to third-party servers.
What a private AI video generation pipeline is
The AI video generation you see in online tools (type a prompt or upload an image, get a clip back) can also be done in-house. Today's open-source models produce 5-10 second clips at 720p, good enough for social, campaigns and product pages. The difference is where the compute runs: on a GPU you control, not a vendor's cloud.
What I build is not a single model but a chain: generation in ComfyUI, upscaling and frame interpolation, and when needed a style LoRA trained on your material. The result is a repeatable workflow your team (or your agency) can run every week, with the same look from one clip to the next.
AI video production for businesses: when it makes sense
You produce many short clips in the same style
Promotional content, campaign variants, product animations. At high volume, paying a credit-based tool for every attempt gets expensive, and the style drifts between generations.
Your assets can't leave the building
Unreleased products, photos of people, NDA material or an agency's client assets. With a local pipeline, source images, prompts and videos stay on your machines.
You're an agency and want your own service
Instead of reselling a subscription, offer clients production in their own trained style, with clear rules on where their files go.
You want to know it's worth it before buying hardware
We start with a test on a few real cases, on dedicated rented GPUs. If the clips aren't good enough, the project ends there with a contained investment.
Making promotional videos with AI, step by step
Drafts and fast previews
Shots, motion and prompts are explored with fast models: LTX-Video for near-instant previews, Wan 2.2 TI2V-5B for 720p drafts on a single 24 GB GPU. Most ideas are discarded here, cheaply.
Final version with Wan 2.2 14B
Approved drafts are regenerated with Wan 2.2 A14B, the largest model in the family (text-to-video or image-to-video, up to 720p). It takes a few minutes per clip depending on the GPU, so it only runs on what has already been chosen.
Upscaling and frame interpolation
The generated video goes through an upscaler, for example from 720p to 1080p, and a frame interpolator (such as RIFE) that adds in-between frames for smoother motion. All inside the same ComfyUI workflow.
Brand style LoRA (if needed)
A LoRA is a small adaptation trained on a set of your images or clips: palette, lighting, framing, how the product looks. It gives you a consistent, recognisable style; it does not teach the model facts or business data.
What it costs
I don't have a price list: cost depends on clip volume, hardware (your GPU, a dedicated server or hourly rental) and whether you need a style LoRA. Scope and cost of the initial test are clear before we start, and on the first call I'll also tell you if, at your volume, a credit-based tool is the cheaper option.
Open-source AI video models I run locally
| Wan 2.2 TI2V-5B | Wan 2.2 A14B | LTX-Video | |
|---|---|---|---|
| Role in the pipeline | 720p drafts | Final versions | Fast previews |
| Made by | Alibaba (Wan-AI) | Alibaba (Wan-AI) | Lightricks |
| Characteristics | Dense 5B model, text and image to video, 720p at 24 fps | Mixture-of-Experts: two 14B experts, 14B active per step; 480p and 720p | Built for speed: distilled versions generate close to real time on datacenter GPUs |
| Licence | Apache 2.0 | Apache 2.0 | Free for commercial use under $10M annual revenue; above that, an agreement with Lightricks is required |
Your assets never leave
One rule: source images, prompts, LoRAs and final videos stay on infrastructure you choose, whether a workstation in your office, a dedicated server or rented GPUs in a European datacenter under a contract in your name. No calls to outside generation APIs.
The style LoRA trained on your brand is your file: it stays with you when the project ends, and you can keep using it with the same pipeline.
Stefano Salvucci
12+ years across software development, architecture and automation. Co-founder of BusinessinCloud, YTILI Fellow 2019 (U.S. Department of State and German Marshall Fund), registered with the Order of Engineers of Rome. I work as a fractional CTO: I design the pipeline, put it into production and train your team to run it.
Questions about AI video generation
How is this different from an online AI video generator?
An online generator is handy for trying things out, but your files end up on the vendor's servers, you pay per credit and you don't control the style. A local pipeline with open-source models costs more up front, but your data stays yours, the cost per clip falls with volume and the style can be trained on your brand.
How long are the clips and at what resolution?
The typical case is a 5-10 second clip generated at 720p, then taken to a higher resolution with upscaling and smoothed with frame interpolation. Longer videos are made by editing several generations together.
How long does Wan 2.2 take to generate a video?
With Wan 2.2 A14B, the 14B version, a clip takes a few minutes; the exact time depends on the GPU and settings. That's why drafts are made with lighter models (Wan 2.2 5B or LTX-Video) and the 14B model only runs for final versions.
What do I need to run Wan 2.2 in ComfyUI?
An NVIDIA GPU with enough memory: the TI2V-5B model runs on a single 24 GB card such as an RTX 4090, while the 14B model is better on a GPU with more memory or with offloading and quantisation. ComfyUI has native nodes for Wan models; the real work is building a stable, repeatable workflow.
What is AI video upscaling and why do I need it?
It's a model that raises video resolution by reconstructing detail instead of stretching pixels. Generating directly at high resolution is much slower: it's better to generate at 720p and then upscale, alongside interpolation that adds in-between frames.
Does LoRA fine-tuning use my company's data?
The LoRA is trained on images or clips you choose (products, settings, palette) and is used for visual style. Training happens on the same private infrastructure, and the resulting file is yours.
Can I use these videos commercially?
It depends on the model licence. Wan 2.2 is Apache 2.0. LTX-Video is free for commercial use under $10M annual revenue. I'll tell you the licence of every model before we use it.
How do we start?
With a free 30-minute intro call. Tell me what videos you need, how many and what constraints you have on data, and I'll tell you whether a private pipeline makes sense or whether an existing tool is enough for you.
Let's talk about your videos
30 minutes, free: we assess the best fit for your volumes.