FlashSR MLX 8bit
speech-swift — Apple SDK soniqo.audio — website blog — blog
Source du modèle
- Type
- UNet
- Taille du fichier
- 649 MB
- Modèle de base
- jakeoneijk/FlashSR_Inference
- Version
- vmodel
- Mis à jour
- 25 oct. 2026
- Téléchargements
- 32
- Source
- Hugging Face
Description de la source
FlashSR-MLX-8bit
- speech-swift — Apple SDK
- soniqo.audio — website
- blog — blog
MLX port of FlashSR — a single-step distilled latent-diffusion audio super-resolution model — quantized to INT8 weight-only for on-device inference on Apple Silicon. Upsamples any-rate input (mono) to 48 kHz in one diffusion pass with a specialised BigVGAN-flavour SR vocoder. FlashSR is the distilled student of AudioSR.
Model
| Total parameters | 638 M (VAE 223 M + UNet 258 M + Vocoder 157 M) |
| Diffusion steps | 1 (distilled v-prediction student, cosine schedule, t=999) |
| Quantization | INT8 weight-only, group size 64, mode mlx_affine_flat |
| Format | MLX safetensors (single combined bundle) |
| Sample rate | 48 kHz mono out (any-rate mono in) |
| Frame length | 5.12 s (245 760 samples) per forward |
| Bundle size | 649 MB on disk |
| Source | jakeoneijk/FlashSR_Inference |
Files
| File | Size | Description |
|---|---|---|
model.safetensors | 649 MB | INT8-quantized VAE + UNet + SR Vocoder weights |
config.json | ~70 KB | Sub-model configs + quantization metadata + original shape table for dequant-on-load |
The three sub-models share one safetensors file with vae.*, ldm.*, voc.*
key prefixes. config.quantized_shapes records each tensor's pre-flatten shape
so mx.dequantize can rebuild conv weight tensors at load time.
Performance (Apple Silicon, M-series, 5.12 s @ 48 kHz)
| Metric | Value |
|---|---|
| Real-time factor (wall / audio) | 1.33 |
| Load time | 1.34 s |
| SNR vs FP16 reference | +40.2 dB |
| Cosine similarity vs FP16 | 1.0000 (bit-transparent) |
| Peak amplitude preservation | 1.000 |
INT8 is the transparent-quality variant — choose it when you want maximum fidelity vs the upstream PyTorch reference without paying the full 1.2 GB FP16 download.
Usage
from huggingface_hub import snapshot_download
import mlx.core as mx
import numpy as np
import scipy.io.wavfile as wf
from scipy.signal import resample_poly
bundle = snapshot_download("aufklarer/FlashSR-MLX-8bit")
# See https://github.com/soniqo/speech-swift for production usage.
# Toy Python demo (requires the matching MLX FlashSR runtime):
sr, audio = wf.read("lr.wav")
audio = audio.astype(np.float32) / 32767.0
audio_48 = resample_poly(audio, 48000, sr).astype(np.float32)
from flashsr import FlashSR # from this repo's export/
model = FlashSR(bundle)
hr = model(mx.array(audio_48), seed=42)
mx.eval(hr)
wf.write("hr.wav", 48000, (np.clip(np.array(hr), -1, 1) * 32767).astype(np.int16))
Source
- Upstream: jakeoneijk/FlashSR_Inference
- Paper: FlashSR: One-step Versatile Audio Super-resolution via Diffusion Distillation
- Teacher model: AudioSR
License
CC-BY-NC 4.0 — inherited from upstream FlashSR weights. Non-commercial use only.
Sources
1 source
Sources
1 sourceVérifié 25 oct.
Artefacts du modèle
1 artefact
Artefacts du modèle
1 artefactExtraits de sources
3 extraits
Extraits de sources
3 extraits--- license: cc-by-nc-4.0 language: - en tags: - mlx - audio - music - audio-super-resolution - flashsr - latent-diffusion - quantized - int8 base_model: jakeoneijk/FlashSR_Inference library_name: mlx pipeline_tag: audio-to-audio --- # FlashSR-MLX-8bit - [speech-swift](https://github.com/soniqo/speech-swift) — Apple SDK - [soniqo.audio](https://soniqo.audio) — website - [blog](https://soniqo.audio/blog) — blog MLX port of [FlashSR](https://github.com/jakeoneijk/FlashSR_Inference) — a **single-step** distilled latent-diffusion audio super-resolution model — quantized to **INT8** weight-only for on-device inference on Apple Silicon. Upsamples any-rate input (mono) to **48 kHz** in one diffusion pass with a specialised BigVGAN-flavour SR vocoder. FlashSR is the distilled student of [AudioSR](https://github.com/haoheliu/versatile_audio_super_resolution). ## Model | | | |---|---| | Total parameters | 638 M (VAE 223 M + UNet 258 M + Vocoder 157 M) | | Diffusion steps | **1** (distilled v-prediction student, cosine schedule, t=999) | | Quantization | INT8 weight-only, group size 64, mode `mlx_affine_flat` | | Format | MLX safetensors (single combined bundle) | | Sample rate | 48 kHz mono out (any-rate mono in) | | Frame length | 5.12 s (245 760 samples) per forward | | Bundle size | 649 MB on disk | | Source | [jakeoneijk/FlashSR_Inference](https://github.com/jakeoneijk/FlashSR_Inference) | ## Files | File | Size | Description | |---|---|---| | `model.safetensors` | 649 MB | INT8-quantized VAE + UNet + SR Vocoder weights | | `config.json` | ~70 KB | Sub-model configs + quantization metadata + original shape table for dequant-on-load | The three sub-models share one safetensors file with `vae.*`, `ldm.*`, `voc.*` key prefixes. `config.quantized_shapes` records each tensor's pre-flatten shape so `mx.dequantize` can rebuild conv weight tensors at load time. ## Performance (Apple Silicon, M-series, 5.12 s @ 48 kHz) | Metric | Value | |---|---| | Real-time factor (wall / audio) | **1.33** | | Load time | 1.34 s | | SNR vs FP16 reference | **+40.2 dB** | | Cosine similarity vs FP16 | **1.0000** (bit-transparent) | | Peak amplitude preservation | 1.000 | INT8 is the **transparent-quality** variant — choose it when you want maximum fidelity vs the upstream PyTorch reference without paying the full 1.2 GB FP16 download. ## Usage ```python from huggi...
Source context: 32 downloads · 1 likes · Pipeline audio-to-audio · Library mlx · Repo aufklarer/FlashSR-MLX-8bit
Source context: 34 downloads · 0 likes · Pipeline audio-to-audio · Library mlx · Repo aufklarer/FlashSR-MLX-8bit
