File size: 2,726 Bytes
22f491b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | ---
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
- 4-bit
- seedvr2
- image-upscaling
- super-resolution
---
# seedvr2-7b-4bit
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 4-bit checkpoint for SeedVR2 7B image super-resolution.
- q4 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-4bit` | 4.79 GiB | 2.21 s | 3.24 s | 5.10 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-4bit
mlxgen upscale \
--model AbstractFramework/seedvr2-7b-4bit \
--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).
|