Generation first, upscale second
Pair Z-Image generation with a SeedVR2 upscale
Z-Image and SeedVR2 solve different steps, and conflating them is the most common pipeline mistake. Z-Image creates the base image; SeedVR2 restores and enlarges the keeper. Generate first, keep the best frame, then upscale it with a 2K crop test before any 4K push. This page maps the two-step loop so you do not waste a good generation on a bad upscale.
This is a SeedVR2.net pipeline explainer. Z-Image is an upstream image-generation model and SeedVR2 is a restoration model; this page describes how to chain them, not a benchmark of either. The hosted tool is a separate SeedVR2.net service.
Upscale the keeper without local tiling
Once you have generated a keeper with Z-Image, upload it here for a 2K or 4K upscale. The hosted job removes local VRAM from the equation so you can judge the upscale rather than your tiling settings.
First eligible 5-second 720p Standard run: free
Verify your email to use your first video run without credits or a card. A failed run does not use the trial; longer clips and other settings show their exact credit cost before submission.
Browser upload limit: 500 MB. For larger files, paste a publicly accessible HTTPS URL. Each job supports videos up to 10 minutes.
Quick scene presets
Presets pick a model for a common scenario. You can still switch models manually below.
Upscaled preview
Compare the enhanced output before exporting the final file.
Upload a clip, paste a direct video URL, or load the sample. No example video is downloaded until you ask for it.
Tip: keep your original FPS and avoid pre-scaling. The upscaler handles cadence better than most NLE plugins.
Test before processing the full video
Upscaling cannot recreate detail removed by heavy messaging compression. Test the hardest 5–10-second segment first; a higher-priced model is not automatically a better fit.
The two-step pipeline, by the numbers
These four points capture the Z-Image to SeedVR2 loop. They are workflow guidance, not a benchmark of either model.
- Step one
- Z-Image generation
- Step two
- SeedVR2 upscale
- Test order
- 2K before 4K
- On 8GB
- VAE tiling
Create the base image and pick the single best keeper.
Restore and enlarge the keeper, not every generation.
Inspect a 2K crop first; 4K only after it passes.
Enable tiling and process one image per run.
Keep generation and upscaling separate
Z-Image's job is to create a base image from a prompt or reference. SeedVR2's job is to restore and enlarge detail that already exists. Treating SeedVR2 as a generator, or Z-Image as an upscaler, produces the worst of both.
The pipeline is simple: generate several candidates with Z-Image, keep the one with clean structure, then upscale that keeper with SeedVR2. Upscaling every generation wastes credits and hides the one output that was actually worth enlarging.
Pick the keeper before you upscale
Upscaling does not fix a bad generation. If the base image has broken hands, garbled text, or confused edges, SeedVR2 will make those problems larger and more confident, not correct.
Choose the keeper by structure: clean edges, readable text, and coherent anatomy. Then upscale only that one. A single well-chosen keeper beats a batch of mediocre enlargements.
- Generate candidates first.
- Keep the one with clean structure.
- Upscale only the keeper.
Test 2K before pushing 4K
Run a 2K upscale of the keeper and inspect a 100 percent crop around faces, text, and edges. If the 2K pass already shows ringing, plastic skin, or broken letters, 4K will only amplify those artifacts.
4K is a final export choice, not a default. When the 2K crop is clean, 4K is reasonable for thumbnails, print previews, and high-resolution product media.
The 8GB version of this pipeline
On an 8GB card, run the upscale step as one image at a time with VAE tiling enabled when the log shows decode pressure. Close other GPU apps so the full 8GB is available to the upscale, and keep batch size at one for stills.
Generation and upscaling have different memory profiles, so running both on the same tight card can thrash. If the local loop stalls, keep generation local and move the upscale to the hosted tool.
When to use this pipeline
Use the Z-Image to SeedVR2 loop when you generate concept art, product mockups, or AI imagery and need a cleaner, larger final asset. It is the generation-to-delivery path for creators who already have a base image.
If you are starting from a damaged real photo rather than a generated image, skip the generation step entirely and go straight to photo enhancement or image upscaling.
Pipeline guidance limits
This describes how to chain Z-Image generation with SeedVR2 upscaling, not a benchmark of either model. Model behavior changes; validate the loop on your own source and hardware.
- Z-Image and SeedVR2 versions change independently.
- Upscaling cannot invent detail the generation never produced.
- Running both steps on one tight card can thrash.
Z-Image to SeedVR2 questions
Should I generate with Z-Image and upscale with SeedVR2?
Yes, in that order. Generate the base image with Z-Image, keep the single best candidate, then upscale that keeper with SeedVR2. Treating SeedVR2 as a generator or Z-Image as an upscaler produces the worst of both.
Will SeedVR2 fix a bad Z-Image generation?
No. Upscaling makes existing problems larger and more confident. If the base image has broken hands or garbled text, regenerate or edit first; upscale only a keeper with clean structure.
Should I go straight to 4K?
No. Test a 2K upscale and inspect a 100 percent crop first. If the 2K pass shows ringing or broken text, 4K will amplify it. 4K is a final export choice, not a default.
Can I run this pipeline on 8GB VRAM?
With care: run the upscale step one image at a time, enable VAE tiling when decode hits memory pressure, and keep batch size at one. If the local loop thrashs, keep generation local and move the upscale to the hosted tool.