Running ComfyUI on Vast.ai cheaply, bandwidth costs included
Rent a 24 GB GPU on Vast.ai for ComfyUI without surprise bills: picking a host, the download charge most guides skip, disk sizing, and stop versus destroy.
Not everyone has a 24 GB graphics card, and you don’t need one. Renting an RTX 3090 or 4090 on Vast.ai costs roughly $0.25 to $0.40 an hour, which makes Wan 2.2 video and Qwen image editing affordable for anyone.
But the hourly GPU price is not the whole bill. There are two other costs that catch people out: download bandwidth and disk while stopped. This guide covers how to keep all three low. It is based on setting up the AI Character Kit’s workflows on Vast.ai many times.
How does Vast.ai pricing actually work?
Vast.ai is a marketplace. Independent hosts list their machines, and each host sets their own prices. A listing typically has:
- GPU price per hour, charged while the instance is running.
- Storage price, charged for the disk you reserve, whether the instance is running or stopped.
- Internet traffic price, charged per GB uploaded and downloaded. Some hosts charge nothing. Some charge a few cents per GB.
The details are in the official Vast.ai pricing docs. Read them once; it saves money every time after.
Why does bandwidth matter so much for ComfyUI?
Because modern video models are huge, and you download them onto every new instance.
The four AI Character Kit workflows pull about 93 GB of model files: Krea 2, Qwen-Image-Edit 2511, both Wan 2.2 experts and Wan Animate 2, plus text encoders, VAEs and LoRAs. Here is what that costs at different traffic prices:
| Host traffic price | Cost to download 93 GB |
|---|---|
| Free | $0.00 |
| $0.01 per GB | about $0.93 |
| $0.03 per GB | about $2.79 |
| $0.05 per GB | about $4.65 |
At $0.03 per GB you pay close to $3 before rendering anything. That is ten hours of GPU time on a cheap 3090. Do that a few times a month and bandwidth becomes your biggest cost.
Rule one: sort or filter by download price, and pick a host with free or very cheap download. It matters more than saving a few cents on the GPU hour.
If you only need one workflow, download only its models. Workflows 01 and 02 together are about 35 GB, workflow 03 adds about 33 GB, and workflow 04 about 22 GB more.
Which machine should you pick?
| Card | Typical price | Notes |
|---|---|---|
| RTX 3090 / 3090 Ti | $0.20 to $0.35/hr | Best value. What the kit was tested on. |
| RTX 4090 | $0.35 to $0.60/hr | About 1.5 to 2 times faster. |
| RTX 5090 | $0.60 and up | Fastest, often more than you need. |
These were typical prices in October 2026. They move all the time.
Beyond the card, filter for:
- 24 GB VRAM
- At least 32 GB of system RAM. Video models swap between RAM and VRAM. With less RAM, Wan 2.2 crawls or crashes.
- At least 120 GB of disk for all four workflows. That leaves room for ComfyUI, Python packages and your outputs.
- On-demand, not interruptible. Interruptible instances are cheaper, but they can pause in the middle of a render when someone outbids you. For a 3 minute video job, that is a bad trade.
- Good internet speed. A fast connection turns a 40 minute model download into 15 minutes, and those minutes are billed GPU time.
How do you set it up step by step?
- Make an account on Vast.ai and add $5 to $10 of credit. That is plenty for several sessions.
- Pick a template. The PyTorch (Vast) template is a solid base. Vast’s templates docs explain how templates and on-start scripts work.
- Set your filters as above: 3090 or 4090, 24 GB VRAM, 32 GB RAM, 120 GB disk, on-demand.
- Check the traffic price on each listing before you click rent.
- Add an on-start script in the template if you have one. An on-start script runs when the instance boots, so it can install ComfyUI and download your models while you get a coffee. The AI Character Kit includes a Vast.ai setup script for this, written for the PyTorch (Vast) template.
- Rent it. Setup with all models takes about 15 to 40 minutes, depending on the host’s connection.
- Open ComfyUI from the instance’s Open button or the portal link, and drag in your workflow files.
What if you are setting it up by hand?
You can download each model straight from its official Hugging Face page into the right folder. For example, for the Wan 2.2 high noise model:
cd ComfyUI/models
wget -c -P diffusion_models "https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors"The -c flag resumes a broken download instead of starting over, which saves bandwidth on flaky hosts. The ComfyUI manual install guide covers installing ComfyUI itself.
Stop or destroy: which one saves money?
- Stop when you take a break. The GPU charge ends, but you keep paying the storage price for your disk. Your models stay, so you don’t pay bandwidth again when you restart. One catch: on a busy host, the GPU may be rented by someone else while you are stopped, and you have to wait to restart.
- Destroy when you are done for good. All charges end, and everything on the disk is deleted.
A simple rule: if you will be back within a day or two, stop. If not, destroy, and accept the bandwidth cost next time (which is why the free-download host matters).
Download your outputs before you destroy. Files in ComfyUI/output are gone the moment the instance is destroyed. Use the Jupyter file browser or the ComfyUI queue sidebar to save anything you want to keep.
What does a real session cost?
A typical evening making a character, a dozen scenes and a few videos:
| Item | Cost |
|---|---|
| RTX 3090 on-demand, 3 hours at $0.30/hr | $0.90 |
| Model download, free-traffic host | $0.00 |
| Disk while running | a few cents |
| Total | about $1 |
The same session on a host charging $0.03 per GB of download costs nearly $4. Same GPU, same work.
What goes wrong most often?
- Setup seems stuck. Open the instance logs. Model downloads can take 30 minutes or more on slow hosts. If nothing moves for a long time, destroy it and pick a host with faster internet.
- Out of disk. 120 GB is the minimum for all four workflows. If you went smaller, delete the models of a workflow you are not using.
- Out of memory on video. Restart ComfyUI between image and video workflows to clear VRAM, or lower Wan 2.2 to 49 frames.
- The first render is slow. Normal. Models load from disk the first time.
What should you run once it is up?
Start with an image workflow to check everything works, then move to video. The settings that fit on a rented 3090 are the same as at home: see Wan 2.2 image to video on 24 GB and the Qwen Image Edit 2511 beginner guide.
The AI Character Kit includes a Vast.ai setup script for the PyTorch (Vast) template and a step by step rental guide.