ByteDance's SeedVR2 Runs 4x Faster Video Upscaling on Consumer GPUs

The community ComfyUI port of ByteDance's one-step SeedVR2 restorer passes 800K monthly downloads. It runs 7B video upscaling on 8GB GPUs and is free.

·
·
·
ByteDance's SeedVR2 Runs 4x Faster Video Upscaling on Consumer GPUsPRO
  • The ComfyUI SeedVR2 port hit about 812K monthly downloads and roughly 3k GitHub stars.
  • It's built on ByteDance's SeedVR2 paper, a one-step diffusion video restorer accepted at ICLR 2026.
  • The one-step design, trained against a discriminator, gives over 4x faster inference than multi-step diffusion restorers.
  • v2.5 adds GGUF quantization, BlockSwap and VAE tiling, so 7B models run on about 8GB GPUs.
  • A blind test found it close to Topaz: it wins on textures, loses on faces, and tends to produce artifacts.
  • It's free (MIT code, Apache 2.0 weights) and installs via ComfyUI Manager or a multi-GPU CLI.

ByteDance’s Seed team developed SeedVR2, a generative video-restoration model that reconstructs detail while upscaling footage. A widely used ComfyUI port by NumZ and AInVFX’s Adrien Toupet packages the model for local workflows, including low-VRAM systems and multi-GPU processing. It gives developers an open alternative to proprietary tools such as Topaz and older convolutional upscalers such as Real-ESRGAN.

At publication, the model repository showed 812,155 downloads for the previous month, while the ComfyUI project had roughly 3,000 GitHub stars and 255 forks. The available checkpoints include 3-billion and 7-billion parameter variants.

The port’s v2.5 release line is a breaking rewrite focused on memory management, multi-GPU rendering, and production workflows. Existing v2.4 workflows may require replaced or reconfigured nodes. Subsequent patches fixed a system-memory leak during long renders and added AMD ROCm and Apple Silicon support.

One denoising step cuts the expensive loop

Conventional diffusion restoration repeatedly removes noise from each temporal batch, often requiring dozens of denoising steps. SeedVR2 distills that process into one denoising step. The term covers the diffusion stage only: footage still passes through a variational autoencoder, or VAE, that converts pixels to and from the model’s latent representation.

SeedVR2 reaches one-step inference by training a distilled restoration model against a discriminator, a second network that distinguishes restored output from real high-quality footage. The SeedVR2 paper identifies three techniques that keep this training stable:

  • Progressive distillation: Training reduces the denoising-step count gradually, preserving restoration ability as the model approaches a single step.
  • Adaptive window attention: The transformer computes attention within local windows whose dimensions adjust to the output resolution, reducing artifacts when inference and training sizes differ.
  • Feature-matching loss: The generator learns from intermediate discriminator features as well as the final real-or-generated judgment, which stabilizes large-scale adversarial training.

The authors report more than four times the inference speed of their multistep baseline at comparable quality. That comparison applies to the paper’s baseline, so proprietary tools and community implementations require separate benchmarks. The paper is listed as an ICLR 2026 camera-ready publication.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Trending
  • No trending articles

Comments

avatar

Next Reads