seedvr2-7b-8bit

This repository contains MLX-Gen saved weights for ByteDance-Seed/SeedVR2-7B. The checkpoint is designed for local Apple Silicon inference with mlx-gen.

It uses the mflux/MLX saved-weight layout. Quantized checkpoints include MLX quantization tensors. It is not a Diffusers or Transformers from_pretrained() checkpoint.

Source Model

Original model: ByteDance-Seed/SeedVR2-7B.

License and Access

This quantized derivative follows the Apache 2.0 license of the source model.

Quantization

This is an MLX 8-bit checkpoint for SeedVR2 7B image super-resolution.

• q8 for quantizable SeedVR2 transformer linears and VAE attention linears.
• BF16 for convolutions, normalization layers, and other non-quantizable parameters.
• The static SeedVR2 positive embedding is provided by the MLX-Gen runtime.

See the MLX-Gen quantization docs for compatibility notes and measured behavior.

Validation

Measured on an Apple M5 Max with 128 GB unified memory using mlx-gen 0.18.13. The profile upscales docs/assets/upscaling/seedvr2-5x-source.jpg from 133x113 to 658x560 with --resolution 5x --seed 42 --metadata.

Package	Storage	Generation time	Wall time	Max RSS
ByteDance-Seed/SeedVR2-7B source generation files	31.63 GiB	2.64 s	8.69 s	61.62 GiB
AbstractFramework/seedvr2-7b-8bit	8.62 GiB	2.29 s	3.36 s	8.90 GiB

See the SeedVR2 upscaling guide for the 3B/7B source, q8, and q4 comparison sheet generated from this profile.

Compatibility

Requires mlx-gen >= 0.18.13.

Generated with mlx-gen 0.18.13.

Use mlxgen download and mlxgen upscale for this SeedVR2 package.

Usage

python -m pip install -U mlx-gen

mlxgen download --model AbstractFramework/seedvr2-7b-8bit

mlxgen upscale \
  --model AbstractFramework/seedvr2-7b-8bit \
  --image-path input.png \
  --resolution 2x \
  --seed 42 \
  --metadata \
  --output upscaled.png

Attribution

MLX-Gen is based on mflux by Filip Strand and the original mflux contributors. This model card is generated by MLX-Gen so derived checkpoints keep that attribution visible.

Quantized and contributed by @lpalbou.