| --- |
| license: apache-2.0 |
| base_model: ByteDance-Seed/SeedVR2-7B |
| pipeline_tag: image-to-image |
| library_name: mlx-gen |
| tags: |
| - mlx |
| - mlx-gen |
| - mflux |
| - apple-silicon |
| - 8-bit |
| - seedvr2 |
| - image-upscaling |
| - super-resolution |
| --- |
| # 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`](https://github.com/lpalbou/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`](https://huggingface.co/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](https://github.com/lpalbou/mlx-gen/blob/main/docs/quantization.md) 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](https://github.com/lpalbou/mlx-gen/blob/main/docs/upscaling.md) |
| 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 |
|
|
| ```bash |
| 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](https://github.com/filipstrand/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](https://huggingface.co/lpalbou). |
|
|
