RelayVSR vs SeedVR2 tested: 66-second upscale nears SeedVR2 quality
friedlc · reddit · 2026-11-07
Using one MiniMax H3 render (12.25s, 294 frames, 5.7 min on RTX 4090), the author benchmarked four routes to 1080p:
- SeedVR2 7B: +13.1 min; natural skin, best far-face retention
- Native H3 1080p finish: 13.3 min total; softest, most filmic close-up, but far small faces turn mushy
- RelayVSR x4 (from 464×272): +66 s, surprisingly close to SeedVR2 with slightly plastic skin
- RelayVSR x2 (from 720p): +100 s (35 s once loaded), most raw detail but crunchy eyes/hair at default detail 1.0
Actual winner: raise the low pass to 928×544 then native 1080p finish — coherent far faces, 16.5 min total, used for hero shots. The bigger low pass also follows prompts better.
Pitfalls: never run x4 on 720p input (5248×3072 output, 3+ hours without finishing, interrupt fails, ComfyUI restart needed); set VSR Decode detail to 0.65 to tame crunch. ComfyUI flow: Load Video → RelayVSR Conditioning → KSampler (1 step, cfg 1) → VSR Decode → Create Video; int8 models 5.2 GB each + 128 MB flash VAE.
Verdict: RelayVSR isn't real HD — close-ups read fuzzy — but 1 min per clip is unbeatable for previews and sweeping takes.
More from Multimodal
- Dev turns Runway AI video into a playable Unreal Engine game via agent — davidfromkansas · 2026-11-07
- Consistent character dance video pipeline tested in ComfyUI with AnimateDiff, Wan, ReActor — LimberCircle · 2026-11-07
- Lovable to bring OpenAI Images 2.6 into its app-building platform — MyCreativeOwls · 2026-11-07
- Standalone workflow reader, image sorter, and workflow-preserving webp converter for ComfyUI — Southern-Chain-6485 · 2026-11-07
- Gradium's tiny on-device Phonon TTS tops Hugging Face's Open TTS Leaderboard — mattturck · 2026-11-07
- Maket launches Text to Furniture, generating 3D furniture models from text — testingcatalog · 2026-11-07