SeedVR2-3B ONNX FP16

FP16 ONNX exports of SeedVR2-3B for WindowsML / ONNX Runtime.

This repository contains two ONNX layouts:

The repository provides model artifacts and metadata only. Use your own SeedVR2 inference pipeline or loader.

Download

Download only the variant you need. Downloading the whole repository pulls both variants.

Monolithic

hf download Merserk/SeedVR2-3B-ONNX --include "monolithic/*" --include "config.json" --include "bucket_shapes.json" --include "versions*.json"

Python:

from huggingface_hub import snapshot_download

snapshot_download(
    "Merserk/SeedVR2-3B-ONNX",
    allow_patterns=[
        "monolithic/*",
        "config.json",
        "bucket_shapes.json",
        "versions*.json",
    ],
)

Block-sharded

hf download Merserk/SeedVR2-3B-ONNX --include "blocks/*" --include "config.json" --include "bucket_shapes.json" --include "versions*.json"

Python:

from huggingface_hub import snapshot_download

snapshot_download(
    "Merserk/SeedVR2-3B-ONNX",
    allow_patterns=[
        "blocks/*",
        "config.json",
        "bucket_shapes.json",
        "versions*.json",
    ],
)

Files

monolithic/

Keep each .onnx file together with its matching .onnx.data file:

monolithic/
  seedvr2_step_cfg.onnx
  seedvr2_step_cfg.onnx.data
  seedvr2_vae_encoder.onnx
  seedvr2_vae_encoder.onnx.data
  seedvr2_vae_decoder.onnx
  seedvr2_vae_decoder.onnx.data

Default monolithic bucket:

batch=1, frames=5, height=480, width=848

Width is 848 for the monolithic export metadata to preserve 16-pixel divisibility.

blocks/

The block-sharded variant contains the DiT split into separate ONNX files:

blocks/
  seedvr2_cfg_combine.onnx
  seedvr2_dit_embed.onnx
  seedvr2_dit_block_00.onnx
  ...
  seedvr2_dit_block_31.onnx
  seedvr2_dit_out.onnx
  seedvr2_vae_encoder.onnx
  seedvr2_vae_decoder.onnx

Primary block bucket:

bucket_f1_256x256
batch=1, frames=1, height=256, width=256

For the primary block bucket, the VAE encoder accepts [1,3,1,256,256], produces latent [1,16,1,32,32], and the decoder returns [1,3,1,256,256].

Runtime notes

CUDA was used only for export acceleration. The ONNX artifacts target WindowsML-compatible ONNX Runtime execution. The block-sharded export is intended for DirectML block streaming, with CPU fallback possible for components where needed.

Metadata

• config.json lists both variants and the exact paths inside this repository.
• bucket_shapes.json lists bucket shapes for both variants.
• versions.json points to variant-specific version metadata.
• versions_monolithic.json records the monolithic export metadata.
• versions_block.json records the block-sharded export metadata.

Source

Source model: ByteDance-Seed/SeedVR2-3B.

This repository is an ONNX artifact export and is not the original SeedVR2 source repository.