RTX VSR vs SeedVR2 video upscaling on a 12GB card
RTX VSR vs SeedVR2 on one RTX 4070: 23.69 s against 1,243.58 s for the same clip. Use RTX VSR by default and SeedVR2 for fine-detail shots.
In an RTX VSR vs SeedVR2 run on one 12 GB RTX 4070, RTX VSR took 23.69 s service time and SeedVR2 took 1,243.58 s. Use RTX VSR for routine 1080×1920 delivery, and keep SeedVR2 for fine-detail shots. These are rankings for one 4070 and one service, not a universal quality score.
Same clip, two passes
The source was a 5.167 s clip of 124 frames at 480×832, enlarged to 1080×1920. Both passes ran on that one clip, so the gap is the method, though it is a single run.
| Measure | RTX VSR | SeedVR2 |
|---|---|---|
| Service time | 23.69 s | 1,243.58 s |
| Peak whole-card used | 1,325 MiB | 11,730 MiB |
SeedVR2 was 52.5× slower and left only 281 MiB free at peak. Both VRAM figures are whole-card readings under Windows WDDM, not per-process allocations, and client wall time includes queue wait.
What the output looked like
SeedVR2 gives the best cat fur and whiskers, but on smooth CG skin it invents grain and warps eye outlines. RTX VSR sharpened more faithfully. Neither pass fixes source composition or identity.
Across three clips, the upscale step alone added this much:
RTX VSR +24-51 s 1.3 GBSeedVR2 +490-1,244 s 11.7 GBWhich to use when
| Situation | Pick | Why |
|---|---|---|
| Routine 1080×1920 delivery | RTX VSR HIGH | 23.69 s and 1,325 MiB |
| Fur, fabric or other fine detail | SeedVR2 3B FP8 | measured 8-21 minute upscale is justified |
| Smooth CG faces | RTX VSR | SeedVR2 warped the eye outlines |
FlashVSR was never run here, so no result is claimed for it. The reports on it are cross-machine, and they conflict on temporal stability. Review moving video and faces, not only still crops. For other jobs that share this card, see local models on one 12 GB GPU.
Verdict
Make RTX VSR the default pass, since it costs seconds and about 1.3 GiB, and spend the SeedVR2 wait only on a shortlisted fur or fabric shot. If the routine pass repeats often, Claude Code hooks can run it for you.
Try it on your own card: read the SeedVR2 8-12 GB guide
Sources:media-gen/notes/h3-upscale-research-2026-09-29.md, media-gen/notes/hoodie-eval-2026-09-29.md, media-gen/notes/h3-upgrade-eval-2026-09-27.md.
Responses
Keep reading
Qwen image edit mask vs Krea inpaint vs GrabCut
One header, one mask, three ways to darken the background: Qwen left 4.2% of the mask light, Krea 18.5%, GrabCut 0.3%. Qwen's fringe is the catch.
Claude Code local model: run it through Ollama
Point Claude Code at a local model with Ollama: one launch command or three environment variables, a 64k context window, and the features that fall short.
Computer use for agents: agent-desktop install and first run
Computer use for agents with agent-desktop: install with npm, grant macOS Accessibility, snapshot an app, click a ref, scope reads and run several agents.