I tested 11 video upscaling methods for AI video: local, Topaz and cloud compared
lopas99 · reddit · 2026-10-28
To get Seedance 2.5 videos to 4K, the author spent two days and $100 benchmarking 11 video upscaling options (free local, paid Topaz on an RTX 5090, and cloud), judged specifically for realistic anime-style character videos.
All free local options were discarded:
- Wan 2.2 ComfyUI workflow: barely added detail, 43 minutes for 480p→1080p, too slow.
- H3 Video Upsampler: 2-3 minutes but heavy smearing and temporal issues.
- SeedVR2: 6 minutes, worst-looking results for this use case.
- NVIDIA RTX Video Super Resolution: fast (37s to 4K) but output identical to input — only useful for cleaning compressed streams, adds no detail.
Topaz (paid, local):
- Starlight Mini (15 min): conservative, adds clarity while preserving blurred backgrounds — made the final two. But the local app has a known bug (reported on Topaz forums since Nov 2025, still in 1.7.1): a horizontal line across the frame near the bottom with a blurrier strip below it. The same clip via fal has no line, so it's the app, not the model.
- Starlight Precise 2.6: partial results shown (post truncates).
The author stresses results are specific to their use case and invites others to share experiences.
More from Infra
- First hands-on with Bittensor subnet SN53: open-weight models online in 2 minutes via $TAO — markjeffrey · 2026-10-28
- CI demand surges 9x in a week as AI-generated code floods pipelines, Celesto shares scaling fixes — aniketmaurya · 2026-10-28
- Capsem: Rust hypervisor from DeepMind team runs concurrent OCI containers for isolated AI agents — rakyll · 2026-10-28
- SemiAnalysis: OpenAI reportedly negotiating ~1GW TPU commitment while seeking sites — dylan522p · 2026-10-28
- Thermo-backed CTO says AI is in its dial-up phase, touts disaggregated compute for an ultrafast 'broadband moment' — beffjezos · 2026-10-28
- DeepSeek-V4.1-Flash Q2 runs at 41 tok/s on one RTX 3090, 90% token match vs API — musicurgy · 2026-10-28