Cloud GPU for AI film and video production

CloudGPU rents GPUs by the minute for AI film and video work. You pick a card and the ComfyUI template (or a raw GPU with root SSH), run your shots, copy the files off, and destroy the instance. Billing is per minute, with a 1-minute minimum, and stops when you destroy it.

A short film is several jobs, and each needs a different amount of VRAM: storyboard frames, video clips, a character LoRA, voices, music and a final upscale. With hourly billing on a single card you pay for the largest card for the whole project. Per-minute billing lets you rent a large card for the heavy steps and a smaller one for the rest.

See current rates · Deploy a GPU

Which card for which step

VRAM figures come from each project’s own documentation (read 2026-11-04). The guides linked in the table quote them in full. Which cards you can start right now is shown on /deploy. Each card there also lists its vCPU and RAM, which matters for models that offload weights to system memory.

Step Open models and tools VRAM class to start with Example cards
Storyboard frames FLUX.1, SDXL in ComfyUI 24 GB RTX 4090, RTX 4090D
Video clips (5B–8B class) Wan 2.2 TI2V-5B (24 GB with the official script), HunyuanVideo 1.5 (14 GB minimum with offload) 24–32 GB RTX 4090, RTX 5090
Video clips (14B class) Wan 2.2 T2V / I2V-A14B: the official script asks for 80 GB; ComfyUI’s FP8 files (14.29 GB per expert) and offloading go lower 32 GB and up, 80 GB for the official script RTX 5090, RTX 4090 48G, A100 80G
Video with native audio MiniMax H3 (open Base checkpoints), LTX-2.5 (no official VRAM figure) 80 GB class for H3, or 24–32 GB with int8, block streaming and a lot of system RAM RTX PRO 6000 96G, A100 80G, RTX 5090
Character consistency Wan / H3 LoRA with musubi-tuner or ai-toolkit (musubi-tuner recommends 24 GB or more for video training) 24–48 GB RTX 4090 48G, RTX 5090
Voice and lip sync CosyVoice 3, IndexTTS-2.5, LatentSync (8 GB for 1.5, 18 GB for 1.6), MuseTalk 24 GB RTX 4090, RTX 3090
Music ACE-Step 1.5 (XL without offload: 20 GB or more), MiniMax Music 3 (full precision under 24 GB) 24 GB RTX 4090, RTX 4090D
Upscale and interpolation SeedVR2, RIFE, Real-ESRGAN 8–24 GB (SeedVR2 FP16 at 24 GB+) RTX 4090, RTX 5090

Step-by-step guides:

What you get

Item Detail
Templates ComfyUI (no model weights preloaded) or a raw GPU with root SSH and JupyterLab
ComfyUI API A public HTTPS endpoint for ComfyUI’s /prompt API while the instance runs, so a script can queue shots. The endpoint adds no login, so add one if you need it
Downloads Run cloudgpu-net on before pulling weights from Hugging Face or GitHub (guide)
Billing Per minute, 1-minute minimum, from a prepaid balance. Destroying the instance stops billing. Through the GPU API you can set a time limit or spend cap on an instance, so a forgotten render doesn’t keep running
Payment Alipay, Google Pay, Apple Pay, Visa/Mastercard or USDT (payments)

Hosted video models, if you don’t need your own GPU

For finished clips without managing a machine, the CloudGPU model API also serves hosted video models such as Wan 3.0, which has no open weights. Create a job with POST /v1/videos, poll GET /v1/videos/{id}, and download the MP4. Models and prices are on /api, and the request format is on /api/docs.

Check the model licence before you ship

Open video models come under different licences. Wan 2.1 and 2.2 are Apache 2.0. MiniMax H3 and HunyuanVideo 1.5 use their own community licences, and some music and interpolation weights are non-commercial. The guides above link each licence. Read it for the files you use before you publish a film made with them.

FAQ

Which GPU should I rent for AI video?

Start with the step that needs the most memory. Wan 2.2 TI2V-5B runs on a 24 GB card with the official script, and HunyuanVideo 1.5 lists 14 GB as its minimum with offloading. The official script for the Wan 2.2 14B models asks for 80 GB; ComfyUI’s FP8 files and offloading lower that. MiniMax H3 at full precision needs 80 GB or more, or a 24–32 GB card with int8, block streaming and a lot of system RAM. The guides on this page quote each model’s own figures.

Is ComfyUI preinstalled?

Yes, the ComfyUI template installs ComfyUI without model weights. You download the checkpoints your workflow needs into ComfyUI/models over SSH or JupyterLab.

How is GPU time billed?

Per minute, with a 1-minute minimum, from a prepaid balance. Destroying the instance stops billing. Current rates are on the pricing page.

Can I use different cards for different steps?

Yes. Each instance is billed on its own, so you can render clips on a large card, destroy it, and run voice, music or upscaling on a smaller one. Copy your files off before destroying an instance, because its data is deleted.

Can I call ComfyUI from a script?

Yes. The ComfyUI template gives the instance a public HTTPS endpoint for ComfyUI’s API. Queue a workflow with POST /prompt, poll GET /history/{prompt_id} and download results with GET /view.

Sources (read 2026-11-04)

Prices are on https://cloudgpu.app/pricing and are not repeated here. Which cards you can start is shown on https://cloudgpu.app/deploy.

Deploy the ComfyUI templateCheck live prices