SeedVR Upscale

SeedVR2 is the second-generation architecture of the open-weight SeedVR video super-resolution family developed by the ByteDance Seed Team. It introduces a Swin-MMDiT diffusion transformer and an adversarial one-step denoising framework to upscale video footage without temporal drifting.

SeedVR uses a single-step conditional diffusion process rather than traditional iterative 20-to-50-step diffusion. By executing upscaling in a single forward pass through a Causal Video Autoencoder and Shifted Window Multi-Modal Diffusion Transformer (Swin-MMDiT), it locks fine structural details across consecutive frames and eliminates the frame-to-frame flickering and hallucinated morphing common to multi-step upscalers.

Set the scale factor between 1.0x and 10.0x depending on source fidelity. The default 2.0x factor delivers the optimal balance of speed and organic micro-texture recovery for standard AI clips. Clean, high-bitrate archival footage scales cleanly up to 4.0x. Pushing scale factors beyond 4.0x on low-bitrate, heavily compressed video is discouraged, as it can over-sharpen compression artifacts into rigid geometry.

SeedVR excels at reconstructing natural micro-textures—such as skin pores, hair strands, fabric weave, and foliage—in soft or clean footage. However, because it is a faithful super-resolution model rather than a generative repair model, it preserves and sharpens existing source geometry. Avoid using it to fix severe anatomical errors, flat plastic AI skin textures, extreme motion blur, or illegible fine text.

SeedVR was developed by the ByteDance Seed Team and introduced at CVPR 2025, with the refined SeedVR2 architecture presented at ICLR 2026. The model is released as an open-weight framework for temporally consistent video super-resolution.