| --- |
| license: apache-2.0 |
| library_name: mlx |
| tags: |
| - mlx |
| - mlx-swift |
| - super-resolution |
| - image-upscaling |
| - diffusion |
| - quantized |
| - apple-silicon |
| pipeline_tag: image-to-image |
| --- |
| |
| # SeedVR2-3B (MLX-Swift) — int8 |
|
|
| **int8-quantized** MLX-Swift weights for **SeedVR2-3B** (ByteDance, ICLR 2026) — one-step |
| diffusion **super-resolution**. The **on-device** variant: ~half the size, near-lossless. |
| For the [`seedvr2-mlx-swift`](https://github.com/xocialize/seedvr2-mlx-swift) package |
| (MLXEngine / ForgeUpscaler). fp16 base: [`SeedVR2-3B-mlx`](https://huggingface.co/mlx-community/SeedVR2-3B-mlx). |
|
|
| - **Quantization:** transformer DiT Linears int8, group size 64 (the small `vid_in.proj` and the VAE are kept fp16). Declared in `config.json` (`"quantization": {"bits": 8, "group_size": 64}`); the loader applies it automatically. |
| - **Files:** `transformer.safetensors` (int8, **~3.9 GB** vs 7.9 GB fp16) · `vae.safetensors` (fp16) · `pos_emb.safetensors` · `config.json`. |
| - **Quality:** int8 `t_out` cosine vs fp16 = **0.9999749** (near-lossless; int4 by contrast degrades this model to ~22.7 dB — use int8 on-device). Reload round-trip cosine **1.0**. |
|
|
| ## Usage |
|
|
| ```swift |
| import SeedVR2MLX |
| let upscaler = try SeedVR2Upscaler(directory: weightsDir) // detects int8 from config, applies quantize on load |
| let out = upscaler.upscale(processedImage: img, seed: 42) |
| ``` |
|
|
| ## Provenance & license |
|
|
| **ByteDance Seed** (*SeedVR2*, ICLR 2026, [ByteDance-Seed/SeedVR](https://github.com/ByteDance-Seed/SeedVR), **Apache-2.0**) |
| → fp16 [`numz/SeedVR2_comfyUI`](https://huggingface.co/numz/SeedVR2_comfyUI) → MLX ref |
| [`filipstrand/mflux`](https://github.com/filipstrand/mflux) → MLX-Swift port + int8 conversion by |
| **MVS Collective (xocialize)**. Format/precision-converted weights (not a new model); Apache-2.0 applies. |
|
|
