SeedVR
ByteDance's diffusion-transformer video restoration model — open weights, one-step successor SeedVR2, and a growing ComfyUI/Replicate ecosystem.
About
SeedVR is ByteDance Seed's diffusion-transformer for generic real-world video restoration. In June 2025 the team released inference code and model weights (github.com/ByteDance-Seed/SeedVR, weights on Hugging Face), which made it one of the first serious diffusion video-restoration models you could actually run. The follow-up, SeedVR2 (accepted to ICLR 2026), pushes the idea to one-step restoration via diffusion adversarial post-training, with 3B and 7B variants, adaptive window attention for high-resolution inputs, and claims of comparable-or-better quality against multi-step approaches in a single step.
The ecosystem has moved fast: a ComfyUI node, a video tutorial, and a Replicate Cog wrapper all exist, with community notes suggesting 7B FP16 needs around 24GB VRAM while quantized builds can squeeze into 8GB. The weights carry a research license — fine for personal use, but check terms before distributing or commercializing. All quality claims are author-reported.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (SeedVR2, ICLR 2026): comparable or better than existing multi-step VR approaches in a single step
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: official release repo ByteDance-Seed/SeedVR (June 2025: inference code + HF weights confirmed via the SeedVR2 repo updates); SeedVR2 accepted to ICLR 2026; ComfyUI and Replicate integrations documented. Strengths: major-lab backing, genuinely open weights, one-step diffusion is a real advance, fast-growing ecosystem. Docked for: (1) quality claims are author-reported with no independent audit; (2) weights are research-licensed, not commercially open; (3) heavy VRAM requirements (24GB for 7B FP16 per community reports). License: research license; confirm terms before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ByteDance Seed; diffusion-transformer video restoration
- · Inference code + HF weights released June 2025 (ByteDance-Seed/SeedVR)
- · SeedVR2 accepted to ICLR 2026: one-step restoration, 3B/7B variants, adaptive window attention
- · ComfyUI node + Replicate Cog wrapper; 7B FP16 ~24GB VRAM per community reports; claims author-reported