AI Video Generation Hosting — Open Models, Your Hardware
Run open-source video generation models on a dedicated GPU server. Wan 2.1, HunyuanVideo, LTX-Video, AnimateDiff and CogVideoX all preconfigured on ComfyUI. 32 GB minimum, 96 GB recommended for batch work.
Models & Stack
Everything you need to run production AI workloads on dedicated hardware in the UK.
Wan 2.1 (Alibaba)
14B and 1.3B I2V/T2V variants, 720p output, MIT licensed. The current best open-source video model — quality competitive with closed-source providers for most prompts.
HunyuanVideo (Tencent)
13B parameter T2V, FP8 quantised fits comfortably on a 5090. Strong prompt adherence and motion realism.
LTX-Video (Lightricks)
Lightweight 2B model focused on real-time-ish generation. ~5× faster than HunyuanVideo at lower fidelity. Great for iteration.
AnimateDiff + Stable Video Diffusion
Older but battle-tested. SVD for I2V on a 24 GB card, AnimateDiff for stylised animation.
ComfyUI workflow runner
Pre-configured ComfyUI with the most-used video workflow JSON files. Frontend bound to your IP, FastAPI server for programmatic access.
Big NVMe scratch
Video models need 1–2 TB of scratch space for the diffusion cache and frame outputs. Standard on every video-tier server.
GPU Tier vs Video Model
Video models are VRAM-hungry. The chart shows what runs comfortably on each card.
| Model | Params | FP16 VRAM | INT4 VRAM | Recommended |
|---|---|---|---|---|
| RTX 3090 24 GB | Stable Video Diffusion / AnimateDiff | ~14 GB | Fits | from £159/mo |
| RTX 3090 24 GB | LTX-Video 2B | ~12 GB | Fits, fast | from £159/mo |
| RTX 5090 32 GB | HunyuanVideo 13B FP8 | ~30 GB | Tight but fits | from £399/mo |
| RTX 5090 32 GB | Wan 2.1 14B FP8 | ~28 GB | Fits | from £399/mo |
| RTX 6000 Pro 96 GB | HunyuanVideo 13B FP16 + batch | ~70 GB | Comfortable batch 4 | from £899/mo |
| RTX 6000 Pro 96 GB | Wan 2.1 14B FP16 + LoRAs | ~80 GB | Fits with multiple LoRAs | from £899/mo |
| 2× RTX 5090 | HunyuanVideo + ComfyUI farm | — | Two parallel workers | from £899/mo |
What People Build
Real customer workloads we run on this hardware every day.
Marketing & advertising
Generate ad creative variants in-house instead of paying per-clip rates to a closed video API. Brand-safe, on-site, infinite iteration.
Stock + asset libraries
Render 1,000 clips overnight for a stock library. Fixed monthly cost, no per-clip pricing.
Research & development
Universities and AI labs needing transparent reproducible video generation. Pin a Hugging Face commit, run experiments for months.
Creative tools
ComfyUI-based products that bake video generation into a wider workflow — AnimateDiff for animators, SVD for I2V.
Frequently Asked Questions
The questions buyers actually ask before committing to a GPU server.
How long does a clip take?
Wan 2.1 14B FP8 on a 5090 — about 3-5 minutes for a 5-second 720p clip. HunyuanVideo similar. LTX-Video ~30 seconds. Real-time generation is not yet feasible on consumer hardware.
Can I fine-tune video models?
LoRA training on Wan 2.1 and HunyuanVideo is supported on 6000 Pro 96 GB. Full fine-tunes need multi-GPU clusters.
Do you support Sora?
No — Sora is closed-source. We host open models that you can run, modify, and audit.
Can I expose ComfyUI as an API?
Yes — ComfyUI exposes a /prompt and /history HTTP API. Front it with auth and you have a video generation API.
What about audio + lipsync?
Wan 2.1 has audio variants. For lipsync we recommend Memo or LatentSync. Both pre-installed.
Related Pages
Pages our visitors typically read next.
Self-host AI video generation.
Wan, HunyuanVideo, LTX-Video and AnimateDiff on a dedicated GPU server. ComfyUI pre-installed, full root access, your data stays on your hardware.
