README.md

---
license: apache-2.0
library_name: mlx
tags:
  - mlx
  - mlx-swift
  - super-resolution
  - image-upscaling
  - diffusion
  - apple-silicon
pipeline_tag: image-to-image
---

# SeedVR2-3B (MLX-Swift) — fp16

**MLX-Swift** weights for **SeedVR2-3B**, ByteDance's one-step diffusion **super-resolution /
restoration** model (ICLR 2026). For on-device upscaling on Apple Silicon via the
[`seedvr2-mlx-swift`](https://github.com/xocialize/seedvr2-mlx-swift) package (built for
**MLXEngine / ForgeUpscaler**). int8 variant: [`SeedVR2-3B-mlx-int8`](https://huggingface.co/mlx-community/SeedVR2-3B-mlx-int8).

- **Files:** `transformer.safetensors` (DiT, fp16, ~7.9 GB) · `vae.safetensors` (3D-causal-conv VAE, fp16) · `pos_emb.safetensors` (precomputed text embedding) · `config.json`.
- **Precision:** fp16. Parity vs the mflux reference (CPU): transformer `t_out` max_abs **2.1e-4**, VAE encode/decode **3.5e-3 / 7.2e-3**, RNG/scheduler **0.0**.

## Usage

```swift
import SeedVR2MLX   // github.com/xocialize/seedvr2-mlx-swift
let upscaler = try SeedVR2Upscaler(directory: weightsDir)   // downloaded from this repo
let out = upscaler.upscale(processedImage: img, seed: 42)   // [-1,1], dims padded to /16
```

(Preprocess — resolution/softness bicubic resize — and LAB color-correction are host-side;
VAE tiling for large images is handled by the host, e.g. ForgeUpscaler's tile processor.)

## Provenance & license

Chain: **ByteDance Seed** — *SeedVR2: One-Step Video Restoration via Diffusion Adversarial
Post-Training* (ICLR 2026), [ByteDance-Seed/SeedVR](https://github.com/ByteDance-Seed/SeedVR),
**Apache-2.0** → PyTorch fp16 redistribution [`numz/SeedVR2_comfyUI`](https://huggingface.co/numz/SeedVR2_comfyUI)
→ MLX reference impl [`filipstrand/mflux`](https://github.com/filipstrand/mflux) → MLX-Swift
port + weight conversion by **MVS Collective (xocialize)**. These are format-converted weight
artifacts (not a new model); Apache-2.0 applies. Credit ByteDance Seed (original), cite the paper.