Cloud GPU for AI short drama series
A short drama series is the same characters in dozens of short shots, episode after episode. On CloudGPU you rent a GPU with ComfyUI by the minute, generate the shots for an episode, copy them off and destroy the instance. Billing is per minute, with a 1-minute minimum, and stops when you destroy it.
See current rates · Deploy a GPU
The workflow for one episode
- Lock the characters. Train a character LoRA once from 20–40 reference images, then reuse it in every episode so faces and costumes stay the same. See how to train a Wan 2.2 LoRA with musubi-tuner.
- Generate the shots. Wan 2.2 in ComfyUI turns a prompt or a first frame into a 5-second clip. Vertical 9:16 works the same way: swap width and height in the latent node. See how to run Wan 2.2 in ComfyUI.
- Add voices and lip sync. Use CosyVoice 3 or IndexTTS-2.5 for dialogue, then LatentSync or MuseTalk to match the mouths.
- Upscale and cut. Upscale with SeedVR2 or Real-ESRGAN, then edit the episode in your usual editor.
Which card fits each step is in AI film and video production: which card for which step.
What a minute of finished video costs
Measured on one RTX 4090 in ComfyUI on 2026-11-04 (settings in the RTX 5090 benchmark guide):
| Model and output | GPU time per 5 s clip | GPU cost per 1 min of finished video |
|---|---|---|
| Wan 2.2 TI2V-5B, 1280×704, 20 steps | 169–175 s | about $0.23 (35 GPU-minutes) |
| Wan 2.2 T2V-A14B FP8, 832×480, 20 steps | 354–361 s | about $0.47 (72 GPU-minutes) |
That is GPU time per clip divided by clip length, times 60 seconds, at the RTX 4090 rate of $0.39/hour (prices as of 2026-11-07 15:47 (UTC+8)). Short dramas need retakes, so budget for several takes per shot. Model downloads and idle time are extra.
Wan 2.1 and 2.2 are released under Apache 2.0, so the clips you make with them can be used commercially. Check the licence of every other model you add to the workflow.
MiniMax H3: video with native audio
MiniMax H3 generates the picture and the soundtrack together. On /deploy, the MiniMax H3 (ComfyUI) template downloads the H3 files after the instance is handed over and adds an example workflow. In our test on one RTX 4090 48G (1344×768, 24 fps, int8 files, 4-step turbo LoRA), a 5-second clip took 85–87 s once the model was loaded, about $0.20 of GPU time per minute of finished video. Details are in which GPU you need for MiniMax H3. H3’s weights come under the MiniMax H3 Community License, not an open-source licence: read it before you use the weights or publish what you make with them.
Getting your material on and off the machine
- Upload reference images and scripts through JupyterLab’s upload button, or with
scp/rsyncover SSH. - Pull models from Hugging Face or GitHub after running
cloudgpu-net on(guide). - If your weights don’t fit, choose a larger system disk on /deploy before you start the instance. A larger disk adds a per-hour charge.
- Copy finished clips off before you destroy the instance. Its data is deleted when it is destroyed.
Queue a whole episode from a script
The ComfyUI template gives the instance a public HTTPS endpoint for ComfyUI’s API while it runs. Export your workflow in API format, then send one POST /prompt per shot with that shot’s prompt and seed, poll GET /history/{prompt_id}, and download each clip with GET /view. The endpoint adds no login, so add one if you need it.
FAQ
Which GPU should I rent for AI short dramas?
Start with an RTX 4090 for Wan 2.2 TI2V-5B and the 14B FP8 files in ComfyUI: in our test both ran on one RTX 4090, peaking at about 21 GB of VRAM. Choose a card with more VRAM for longer clips, higher resolution or LoRA training. Which cards you can start right now is shown on /deploy.
How much does a minute of finished video cost?
In our RTX 4090 test, Wan 2.2 TI2V-5B at 1280×704 took about 35 GPU-minutes per minute of finished video, and Wan 2.2 14B FP8 at 832×480 took about 72. Multiply by the current RTX 4090 rate on the pricing page. Retakes add to that.
Can I make vertical 9:16 videos?
Yes. Set the width and height in the Wan latent node the other way round, for example 704×1280 instead of 1280×704.
How do I keep characters consistent across episodes?
Train a character LoRA once and load it in every shot’s workflow. Keep the same trigger word and reference outfit descriptions in your prompts.
Can I use the videos commercially?
Wan 2.1 and 2.2 are Apache 2.0. Other models in your workflow, such as voice, lip-sync or upscaling weights, have their own licences, and some are non-commercial. Read each one before you publish.
Prices are on https://cloudgpu.app/pricing and are not repeated here. Which cards you can start is shown on https://cloudgpu.app/deploy.