license: openrail++
tags:
- mlx
- apple-silicon
- diffusion
- stable-diffusion-xl
- sdxl
- text-to-image
base_model: SG161222/RealVisXL_V5.0
library_name: mlx-gen
pipeline_tag: text-to-image
RealVisXL V5.0 — MLX pre-quantized tiers
Pre-quantized, packed-load tiers of SG161222/RealVisXL_V5.0
for on-device Apple-Silicon inference with SceneWorks / mlx-gen
(the sdxl generator). Each tier is a self-contained diffusers turnkey snapshot (U-Net + both
CLIP text encoders + VAE + tokenizers + scheduler + model_index.json) that loads directly — no
in-app quantization pass, no dense transient.
RealVisXL V5.0 is a photoreal SDXL finetune (openrail++, commercial-OK, ungated). Same architecture as SDXL base: dual CLIP-L + OpenCLIP-bigG text encoders, real classifier-free guidance + negative prompt, sdxl-family LoRA support. ~30 steps at guidance 7.0, native 1024×1024.
Tiers
| dir | precision | what's quantized |
|---|---|---|
q4/ (default) |
group-wise affine Q4, group size 64 | U-Net Linears + both CLIP encoders |
q8/ |
group-wise affine Q8, group size 64 | U-Net Linears + both CLIP encoders |
bf16/ |
dense (full-precision master) | nothing — verbatim source mirror |
The VAE stays dense (f32) in every tier — the SDXL VAE is int8/fp16-unstable, so it is never quantized. Convolutions, GroupNorms, and the CLIP token/position embeddings also stay dense (gather lookups and convs, not matmuls); only the true Linear projections are packed.
Quantization is byte-identical to mlx-gen's load-time nn.quantize (bf16 cast, group 64), so a
tier loaded here matches an in-app-quantized dense checkpoint bit-for-bit.
Usage
use mlx_gen::{LoadSpec, WeightsSource, Quant};
let spec = LoadSpec::new(WeightsSource::Dir("…/realvisxl-mlx/q4".into())).with_quant(Quant::Q4);
let g = mlx_gen::load("sdxl", &spec)?;
License
openrail++ — inherited from the source model
SG161222/RealVisXL_V5.0. See LICENSE.