Short answer: yes. You do not need to own a GPU to use SeedVR2. You can run the same family of restoration work through a hosted upscaler, upload one file, and pay per job instead of buying hardware. The harder question is not whether it works without a GPU. It is whether the hosted path fits your file, your budget, and the amount of control you need.
This guide separates the two paths honestly. Local ComfyUI gives you model control and no per-job fee, but it expects a supported GPU, a working Python environment, and time to debug nodes. A hosted workflow gives you a browser upload and a quoted credit price, but you trade away node-level control and some file-size freedom.
What SeedVR2 needs locally
SeedVR2 is a diffusion restoration model published by ByteDance as 3B and 7B checkpoints under Apache-2.0. In ComfyUI, the maintained path is a native template that loads the diffusion model and a shared VAE, then runs a sampling workflow. That is a different workload from a small GAN upscaler. A GAN upscale is one forward pass. A diffusion restore is a multi-step denoise, and every step costs compute.
The practical result is simple: a desktop GPU with enough VRAM is the normal local path. The exact floor depends on the checkpoint, precision, resolution, batch size, and whether you use VAE tiling or BlockSwap. Community 8GB experiments exist, but they are tuning exercises, not an official hardware promise. If you are on a laptop with integrated graphics, a CPU-only server, or a machine you do not want to fill with model files, local ComfyUI is the wrong first move.
There is also a maintenance cost that does not appear in a VRAM table. You need ComfyUI at a compatible version, the correct model filenames in the correct folders, the VAE in ComfyUI/models/vae, and a workflow that matches the checkpoint. A missing node or a stale custom-node pack can stop the first run before the model sees your file. The SeedVR2 ComfyUI install guide covers those steps in order.
What "online SeedVR2" actually runs
A hosted upscaler runs the model on hardware you do not manage. You upload a source, choose a target resolution or model, and the service returns a processed file. The value is not that the model becomes magic. The value is that the GPU, the driver stack, the model files, and the queue are someone else's problem for the length of one job.
On SeedVR2.net, the image tool quotes a flat credit price by target resolution:
| Image target | Credits | Typical job |
|---|---|---|
| 2K | 10 | one still image |
| 4K | 16 | one still image |
| 8K | 25 | one still image |
The video tool bills per second of source footage, with a five-second minimum. On the standard model, the published rates are 9 credits per second at 720p and 1080p, 17 at 2K, and 33 at 4K. A five-second 720p clip therefore starts at 45 credits, and a 30-second 1080p clip costs 270 credits before any model upgrade. The credit cost guide has the full table.
Those numbers matter because they let you compare the hosted path with the real cost of local hardware. A new GPU is not just a purchase. It is a power supply, cooling, desk space, driver updates, and the hours you spend learning a node graph. If you process one clip a week, the hosted path is usually the cheaper experiment. If you process hundreds of clips a day and already own the machine, local can win on marginal cost.
The one-file test that decides
Do not compare the two paths on different footage. Pick one source that represents your real work: a compressed family video, a product shot, or an AI-generated clip with soft edges. Run it through the hosted image or video tool first, because that takes minutes and gives you a real output. Then decide whether the result is good enough to justify learning the local stack.
Use the same output resolution and the same crop when you compare. A 4K file that looks sharper at a glance may simply be more compressed. Check the details that matter for your use case: faces, small text, fabric texture, straight edges, and whether repeated patterns turn into plastic. If the hosted result already changes identity or invents detail you did not want, a local install will not automatically fix that. The model family is the same. The difference is control, not truth.
The test also gives you a real cost number. If one 30-second 4K clip costs 990 credits, you can compare that with your own time. A local run that takes an afternoon to set up is not free. A hosted run that costs a few dollars and finishes before lunch may be the better first purchase, even if the per-second rate looks higher on paper.
When local still wins
Local ComfyUI is the better path when you need repeatable batch work, custom nodes, or a workflow that combines upscaling with other steps. If you already have a supported GPU and a working ComfyUI install, the marginal cost of another job is mostly electricity and time. You can keep files on your own machine, choose your own model precision, and tune VAE tiling or BlockSwap without a vendor menu.
It is also the better path when you want to learn the model. Running the native template, reading the sampler settings, and comparing 3B against 7B teaches you what the outputs can and cannot do. That knowledge transfers to every later job. The cost is the setup time and the hardware you already own.
When the hosted path wins
The hosted path wins when the job is occasional, the machine is weak, or the file is due today. It also wins when you do not want to install model files just to answer one question. A browser upload with a quoted price is a lower-risk first step than a multi-hour environment build.
It is not a free lunch. You still need to check the output, and you still pay per job. The honest comparison is hosted credits against your own hardware, electricity, and time, not hosted credits against zero. For a single 30-second 4K clip, the hosted price is easy to see. The local price is the machine on your desk and the afternoon you spend configuring it.
A practical decision rule
Use this rule before you install anything:
- One or two files, no GPU, deadline this week: use the hosted image upscaler or video upscaler.
- A repeatable batch, supported GPU, and time to tune: use local ComfyUI and the native SeedVR2 template.
- A client deliverable where identity must not drift: test both on the same clip and compare faces before you commit.
- A low-VRAM laptop: read the 8GB VRAM guide before assuming a local run will finish.
The point is not to pick a side forever. The point is to stop spending a weekend on a local install when a hosted run would answer the same question in minutes, or to stop paying per job when you already own the hardware and run the same workflow every day.
FAQ
Does SeedVR2 require a GPU? The local ComfyUI path expects a supported GPU for practical video work. The hosted path runs on managed hardware, so your own machine only needs a browser.
Can I run SeedVR2 on CPU? It is not a practical path for video. Diffusion restoration is compute-heavy, and a CPU run that eventually finishes can take far longer than the job is worth.
Is the online result the same model? The online workflow uses the same SeedVR2 family, but the exact checkpoint, precision, and settings are managed by the service. Treat the output as a result to inspect, not as a guarantee of a specific local configuration.
How much does one online job cost? A 2K image is 10 credits. A five-second 720p video starts at 45 credits. A 30-second 1080p clip is 270 credits on the standard model. The live tool quotes the exact number before you submit.
Can I try it without buying a pack? Check the current trial terms on the pricing page. Trials are limited to qualifying first jobs and are not a substitute for a full production run.
Try one real file: open the SeedVR2 image upscaler, upload a source you already know, and compare the result at the same crop before you decide whether a local install is worth it.