The 3B versus 7B question is usually framed as a quality question. In practice it is a memory and deadline question. The larger checkpoint is stronger on detail, but only some jobs can use that strength and only some machines can run it in time. This guide turns the choice into a short decision table instead of "it depends".
What the two checkpoints are
Both are ByteDance SeedVR2 upscaling models run inside ComfyUI. In the local setup they share one VAE and differ in the diffusion checkpoint:
| File | Folder | Role |
|---|---|---|
seedvr2_3b_int8_convrot.safetensors | ComfyUI/models/diffusion_models | 3B diffusion checkpoint, INT8 |
seedvr2_ema_vae_fp16.safetensors | ComfyUI/models/vae | Shared VAE, FP16 |
The 3B INT8 checkpoint runs on 8GB of VRAM with block swap. The 7B FP8 checkpoint is larger, slower, and stronger on detail, and it needs more memory. That is the whole trade in one sentence: more detail, more time, more memory.
If you are installing from scratch, the installation guide covers ComfyUI 0.28.0 and the folder layout, and the model download page is where the files come from. The low-VRAM guide covers block swap for the 3B on 8GB cards.
The short answer
| Your situation | Run this | Why |
|---|---|---|
| 8GB VRAM, batch due today | 3B INT8 | It is the checkpoint that fits with block swap |
| 8GB VRAM, no deadline pressure | 3B INT8, or hosted 7B-class output | Local 7B is not the practical path on this card |
| Compressed phone video or old JPEGs | 3B or hosted | Extra detail capacity mostly sharpens existing artifacts |
| Clean RAW stills, 4K or 8K target | 7B FP8 if memory allows | Fine detail is where the larger model is stronger |
| One image, needs to be right | Hosted upscaler | No install, no VRAM ceiling, 10 credits at 2K |
| Long video, tight budget | Standard video model, not the local model | Per-second rates and no GPU time cost |
The rest of this post explains each row.
Row by row: source quality
Source quality is the first filter because the model can only work with what is there.
Clean sources favor the 7B. If the source is a high-bitrate file, a camera RAW export, or a scan with real optical detail, the 7B's stronger detail reconstruction has something to reconstruct. This is the case where paying for the larger checkpoint in time or credits makes sense.
Compressed sources rarely do. Old JPEGs, WhatsApp video, and screen recordings have artifacts that are larger than the detail the model would add. Both checkpoints will sharpen those artifacts. The 3B does it faster and on less memory, so if the source is heavily compressed, the larger model is mostly extra cost.
Faces and text are the deciding content. A group photo with small faces, or a product shot with a label, is where you will actually see a difference. Test one of those files, not a landscape.
Row by row: target resolution
The target resolution sets how much reconstruction work the model does.
| Target | Credits on SeedVR2.net | Which local model is usually enough |
|---|---|---|
| 2K | 10 | 3B is generally sufficient for clean sources |
| 4K | 16 | 7B becomes worth testing on detailed sources |
| 8K | 25 | 7B if memory allows; otherwise upscale in two steps |
At 2K, the output is close enough to the source scale that the difference between checkpoints is hard to justify on an 8GB card. At 4K and 8K, the model is filling in more, and that is where a larger checkpoint has more room to matter. "Generally sufficient" is deliberate: test your own file before assuming either direction.
Row by row: time and memory
Time is the constraint that decides most real jobs.
- Memory: the 3B INT8 runs on 8GB with block swap. The 7B FP8 is larger and needs more memory. If you have 8GB, the local decision is effectively made for you.
- Speed: the 7B is slower. Block swap adds time on top of that. A batch that is comfortable on the 3B can become an overnight job on the 7B.
- Complexity: both need the same VAE and the same ComfyUI version, so the install work is similar. The difference is what happens when you press queue.
This is why "which is better" is the wrong question for a single-GPU setup. The better question is which one finishes before your deadline.
Row by row: deadline
| Deadline | Local 3B | Local 7B | Hosted |
|---|---|---|---|
| Same day, 1–10 images | Reasonable | Risky on 8GB | Fastest path |
| Same day, 100 images | Plan the hours | Usually not viable | Priced per image, parallel |
| This week, 1–10 images | Comfortable | Worth testing | Optional |
| This week, 100+ images | Overnight batches | Only with headroom | Compare credits against your time |
The hosted route is not a fallback for people who cannot install ComfyUI. It is a legitimate answer to the memory question: SeedVR2.net runs the models server-side, so the 7B's memory requirement never reaches your desk. You pay in credits instead of GPU time and setup.
The hosted image upscaler charges 10 credits at 2K, 16 at 4K, and 25 at 8K. Video bills per second on the standard model at 9 credits per second for 720p and 1080p, 17 for 2K, and 33 for 4K, with a 5-second minimum and a 10-minute maximum per clip.
Estimate the credits for your own job
Pick a resolution and length. The calculator uses the same credit rates as checkout, then shows which pack covers the job.
PricingPage.pricingCalculator.title
PricingPage.pricingCalculator.description
PricingPage.pricingCalculator.estimateTitle
PricingPage.pricingCalculator.creditsValue
PricingPage.pricingCalculator.perVideoSummary
- PricingPage.pricingCalculator.timeLabel
- PricingPage.pricingCalculator.timeValue
- PricingPage.pricingCalculator.sizeLabel
- PricingPage.pricingCalculator.sizeValue
PricingPage.pricingCalculator.videoTimeEstimateHint
PricingPage.pricingCalculator.videoFootageScale
PricingPage.insufficientCredits.failedTaskRefund
PricingPage.pricingCalculator.deliveryHint
PricingPage.pricingCalculator.recommendedPackage
$9.99
PricingPage.pricingCalculator.recommendedPackageHint
A first image upscale at 2K is free, and a guest can see a 720px preview with a $4.99 unlock for the full image. That is enough to compare a hosted result against your local 3B output before committing either way.
A five-question procedure
Ask these in order and stop at the first one that decides it:
- Does the 7B fit in my VRAM with room to finish the job? If no, use the 3B or the hosted tool. Do not spend the afternoon fighting an out-of-memory error.
- Does the source contain fine detail I need to keep? If no, the larger model has little to work with; use the 3B.
- Is the target 4K or 8K? If yes, test the 7B on one file. If the target is 2K, the 3B is usually enough.
- How many files, and how long does one take? Multiply before you queue. A batch is a time decision, not a quality decision.
- When is delivery? If the answer is today, pick the path that finishes, not the path with the best possible output.
Settings that matter more than the model
Before switching checkpoints, check these. In most disappointing upscales, one of them is the actual cause:
- Target resolution. Rendering a compressed source at 8K amplifies artifacts. Dropping to 2K or 4K often looks better.
- Source preparation. Cropping a bad region and upscaling it separately beats upscaling the whole frame and hoping.
- Content selection. The worst file in the folder is the one to test first, not the one to test last.
- Local versus hosted. The choice of where to run changes your memory ceiling and your deadline far more than the 3B versus 7B choice does.
Model size is a real lever. It is just not the first one to pull.
Limitation
This post does not include measured speed or quality benchmarks for either checkpoint, and it does not state a VRAM requirement for the 7B beyond "more than the 3B" — that number depends on your ComfyUI build, resolution, batch size, and block-swap settings. It also does not cover every node-level parameter, which varies by workflow. The credit prices are the published rates at the time of writing; confirm them on the pricing page and in the live preflight. The honest summary is that the right choice depends on your card and your deadline, and one test file will tell you more than any comparison table, including this one.