If you already own a 24 GB graphics card, a local open model costs you about a cent of electricity per short clip. If you do not, a cloud app is cheaper for most people until they make hundreds of clips a month. Privacy, content rules and setup time matter as much as price.
What a cloud clip costs
Cloud apps sell video as credits inside a plan. Only some publish how many credits a clip uses.
| App | Published plan price | Credit cost of a clip |
|---|---|---|
| OurDream AI | $19.99 a month, or $9.99 a month billed yearly, with 1,000 dreamcoins | 100 coins for a 10-second “Spicy 1.0” clip, per its generate page |
| Promptchan | Video on the Pro plan, $33.99 a month with 800 Gems, per its upgrade page | Not fixed; “a handful of gems” |
| Candy AI | From $3.99 a month on a 12-month plan, with 100 tokens a month on Premium, per its subscriptions page | Not published |
By OurDream’s own numbers, the monthly plan covers 10 such clips, about $2.00 per 10-second clip, or about $1.00 on the yearly price before bonus coins. Its “Prime” mode is listed at 10 coins per second at 720p with audio, which works out to the same 100 coins for 10 seconds or 50 coins for a 5-second clip. That is arithmetic from the vendor’s published numbers, not a clip test. For the other apps, use the cost per clip calculator once you see the real credit cost inside the app.
Remember to count unusable takes. In any image-to-video workflow, some clips drift in the face, melt a hand, or miss the camera move you asked for. If only half of your generations are keepers, your effective cost per usable clip doubles.
Check current plans on OurDream AI
Publishes per-clip dreamcoin costs on its generate page; video requires a paid plan.
What a local clip costs
The Wan 2.2 repository says its TI2V-5B model needs at least 24 GB of VRAM, such as an RTX 4090. Its model card says it makes a 5-second 720p clip in under 9 minutes on one consumer GPU. The ComfyUI docs say it should fit on 8 GB with offloading, which will be slower; the two sources disagree.
NVIDIA lists 450 W total graphics power for the RTX 4090. Nine minutes at 450 W is under 0.07 kWh, about a cent at $0.17 per kWh. Even if you include another 100 W for the rest of the desktop during those nine minutes, power stays near $0.014 per clip.
The real cost is the hardware write-off and storage:
- Graphics card write-off: If you buy a used or new 24 GB card SOLELY for AI video and write off $1,000 of its value over 1,000 clips, hardware adds $1.00 per clip. If you already own the card for gaming or 3D work, the marginal hardware cost is zero.
- Disk space and system RAM: A single open-weight video workflow in ComfyUI requires the diffusion transformer weights, the UMT5 text encoder, and the Wan VAE—often 15 GB to 35 GB of disk space per model variant. Offloading weights between system memory and VRAM also benefits from 32 GB to 64 GB of system RAM.
- Your time: At nine minutes per 5-second clip on a single GPU, generating twenty variations takes three hours of queue time on your machine.
Break-even math: cloud plan vs buying a GPU
Suppose a cloud subscription costs $20 a month for twenty 5-second clips ($1.00 per clip). Over a year, that is $240 for 240 clips. Buying a dedicated 24 GB GPU only beats that monthly subscription in pure dollar terms if you generate hundreds of clips every month or if you plan to resell the card later. Where local hardware wins immediately is privacy, uncapped retries overnight, and freedom from sudden changes to a vendor’s credit pricing or content filters.
Side by side
| Cloud app | Local open model | |
|---|---|---|
| Up-front cost | None; monthly plan | A 24 GB GPU, or a slower offloaded setup |
| Cost per clip | Credits; often not published | About a cent of power plus the card’s write-off |
| Setup | Minutes, in a browser | Install ComfyUI, download several model files |
| Content rules | The app’s own rules, which can change | No filter is documented in the official model pages |
| Privacy | Prompts and outputs pass through the vendor; most say they moderate them | Nothing leaves your computer |
| Billing trace | A card statement entry; some vendors publish a discreet descriptor | None beyond buying the card |
Privacy and billing descriptors
When you use a cloud generator, your prompts, starting frames, and finished clips sit on the vendor’s servers, and your payment card shows a recurring charge. Two vendors in our app comparison name their billing descriptor on their public pages: Candy AI lists “EverAI” and Secrets AI lists “S LABS INC.” Kupid AI names “Servalot Ltd” as its payment processor. Running open-weight video models locally keeps every prompt and frame on your own drive and leaves no subscription trail on a bank statement.
Which one fits you
- A few clips a month, no GPU: a cloud app. Check whether video needs a higher plan; on Promptchan it is Pro only.
- You own a 24 GB card: try Wan 2.2 TI2V-5B locally first. Our setup guide lists the files and folders.
- Privacy comes first: local. Nothing is uploaded, but you carry full responsibility for what you make.
- You need long scenes now: OurDream documents scenes up to 60 seconds each in the cloud.
The same rule in both places
Local models do not stop you from misusing them, and cloud apps ban it in their terms. Either way, never animate photos of real people without their consent, and never anything involving minors. See our consent and safety policy and the guide to what the law says.
Facts checked against the linked official pages on October 31, 2026.