Most wasted credits do not come from a bad model. They come from starting a fifty-file batch before looking closely at one result. This is the sequence we use on SeedVR2.net: verify the source, price the target, run one cheap or free test, inspect it at full size, and only then commit the batch. Every step below lists what it costs, so you can decide where to stop.
The workflow in one pass
- Check the source file: real pixel dimensions, compression, and whether the detail you want still exists.
- Choose the target resolution and price it.
- Run a single test — the free 2K image upscale, a guest preview, or one video clip.
- Inspect the result at 100% against the source.
- Adjust once: same settings at a lower target, a crop, or a different model.
- Scale to the batch only after step 4 passes.
The order matters more than any individual setting. Steps 1 to 5 cost little or nothing. Step 6 is where the credits go.
Step 1: Check the source before you pay
Open the file and confirm its native dimensions rather than trusting the filename. A photo named portrait-final-4k.jpg is often 1,200 pixels wide. You can check quickly with the resolution checker.
Three things decide whether an upscale can work:
- Native size. Going from 800px to 2K is a 2.5x scale. Going from 800px to 8K is roughly 10x. The model can reconstruct plausible detail at both, but it is inventing more at the second.
- Compression. Heavy JPEG blocking and video compression are the first things an upscaler sharpens. If the artifacts are large, a higher target resolution makes them larger and more visible, not smaller.
- Content type. Faces, hair, fabric texture, and small text are the four areas where a bad upscale is obvious. If your deliverable depends on readable text, test one file with text before anything else.
If the source is a 480p video, read the 480p to 4K guide before choosing a target. The short version is that 4K from 480p is a reconstruction decision, not a resolution setting.
Step 2: Price the target
Image upscaling is a flat fee per image by target resolution:
| Target | Credits per image | Typical use |
|---|---|---|
| 2K | 10 | Web, listings, social, most client previews |
| 4K | 16 | Print, large displays, archival masters |
| 8K | 25 | Large-format print and heavy cropping headroom |
Video bills per second of source footage, with a 5-second minimum and a 10-minute maximum per clip. On the standard model:
| Target | Credits per second | 5-second minimum |
|---|---|---|
| 720p | 9 | 45 |
| 1080p | 9 | 45 |
| 2K | 17 | 85 |
| 4K | 33 | 165 |
720p and 1080p cost the same on the standard model, so there is rarely a reason to pick 720p when the source is already 1080p. The Ultimate model is 131 credits per second at 4K, which is a different budget conversation entirely; the model comparison covers when that tier is worth it.
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
Pack pricing changes the per-credit cost:
| Pack | Price | Approximate cost per credit |
|---|---|---|
| 200 credits | $9.99 | about $0.050 |
| 1,000 credits | $19.99 | about $0.020 |
| 1,500 credits | $29.99 | about $0.020 |
| 3,000 credits | $49.99 | about $0.017 |
| 5,000 credits | $79.99 | about $0.016 |
| 15,000 credits | $199 | about $0.013 |
The gap between the 200-credit pack and the 15,000-credit pack is roughly four times per credit. A small test should not drive your pack choice; the batch size should. The credit cost explainer has the full rate table and the subscription comparison.
Step 3: Run one cheap test
This is the step people skip, and it is the cheapest one.
- Images: every account gets one free 2K image upscale. Use it on the hardest file in the set, not the prettiest one.
- Guests: without an account you can see a 720px preview. Unlocking the full guest image is $4.99, which is useful for a one-off job but not a batch strategy.
- Video: one free standard 720p clip up to 5 seconds with a verified email. Cut a 5-second segment that contains the worst motion, the worst compression, or a face.
A 5-second 720p test on the standard model would otherwise cost 45 credits. Running it as the trial means your first paid decision is based on evidence.
If you are working locally in ComfyUI instead of on the hosted tool, the test still applies. Queue a single frame or a single short clip, not the whole folder. Local runs do not spend credits, but they spend GPU time, and a failed 200-frame queue costs more of your day than a failed single frame.
Step 4: Inspect at 100%
Do not judge an upscale from a fitted-to-window preview. At 100% zoom, compare the result to the source in four places:
- Eyes and hair. These show whether the model is reconstructing structure or smearing it.
- Text and logos. Sharpened text can look crisp while being wrong. Distorted letterforms are a stop signal.
- Flat areas. Skin, sky, and painted walls reveal banding, blotching, and over-sharpening halos.
- Noise. Grain that looked natural in the source can become a repeating pattern after upscaling.
For video, scrub rather than pause on one frame. Temporal consistency is the thing a single still cannot show: check that fine texture does not crawl or shimmer between frames, especially in hair, foliage, and fabric.
Step 5: Adjust once, not five times
One test is enough to make a decision if you read it honestly. Three outcomes:
- Clean and better: keep the settings and scale.
- Clean but no visible gain over the source: drop one resolution tier. A 2K file that looks identical to a 4K file saves 6 credits per image.
- Artifacts or invented detail: do not scale. Crop the problem region, fix the source if you can, or accept a lower target. If text is unreadable, no upscale setting will recover it.
The most common mistake in this step is re-running the same bad file at 8K hoping the extra resolution helps. It does not; it makes the artifact bigger.
Step 6: Scale to the batch
Now do the arithmetic before uploading.
| Job | Credits |
|---|---|
| 10 images at 2K | 100 |
| 50 images at 4K | 800 |
| 100 images at 2K | 1,000 |
| 60 seconds of video at 1080p | 540 |
| 60 seconds of video at 4K | 1,980 |
| 10 minutes of video at 1080p | 5,400 |
| 10 minutes of video at 4K | 19,800 |
Two practical notes. First, group files by target resolution so a single 8K outlier does not force the whole batch into an expensive tier. Second, run the worst file first; if it passes, the rest usually will. Our batch photo workflow covers file naming and folder prep so that a failed file is easy to identify and re-run.
For a batch of 100 images at 2K you need exactly 1,000 credits — the $19.99 pack. At 4K the same 100 images cost 1,600 credits, which is the $29.99 pack plus 100 credits, or the 3,000-credit pack if you expect more work later. Subscription plans change the math again: Starter is $19 per month for 1,200 credits, Creator is $39 for 3,500, and Studio is $99 for 10,000. If your batch is a one-off, buy a pack; if it is monthly, compare the subscription against the pack you would otherwise buy.
Running the same sequence in ComfyUI
The hosted tool enforces this sequence for you, but the local workflow is the same shape. Install ComfyUI 0.28.0, place seedvr2_3b_int8_convrot.safetensors in ComfyUI/models/diffusion_models and seedvr2_ema_vae_fp16.safetensors in ComfyUI/models/vae, then run one image before queuing a folder. The 3B INT8 checkpoint runs on 8GB of VRAM with block swap, which makes it the practical starting point on modest cards.
The low-VRAM guide covers the block-swap setup, and the complete ComfyUI guide covers video. Locally the cost is time rather than credits, so test one frame before queueing a batch.
A short checklist before you press go
- Source dimensions confirmed, not assumed.
- Target resolution priced at the per-image or per-second rate.
- One free or paid test completed on the hardest file.
- Result inspected at 100%, including text and flat areas.
- Batch size multiplied out and matched to a pack or plan.
- Worst file runs first.
That is the whole workflow, sequenced so the expensive step happens last.
Limitation
This post cannot tell you how a specific file will look after upscaling, and it does not publish render times or quality benchmarks for particular hardware. The credit figures are the published rates at the time of writing, and the live preflight in the image upscaler and video upscaler remains the source of truth. It also does not cover sampler, denoise, tile, or other node-level settings, which depend on the workflow you are running. For those, use the ComfyUI guides linked above and confirm the values against the workflow file you actually downloaded.