--- license: openrail++ tags: - mlx - apple-silicon - diffusion - stable-diffusion-xl - sdxl - sdxl-lightning - text-to-image base_model: SG161222/RealVisXL_V5.0_Lightning library_name: mlx-gen pipeline_tag: text-to-image --- # RealVisXL V5.0 Lightning — MLX pre-quantized tiers Pre-quantized, packed-load tiers of [SG161222/RealVisXL_V5.0_Lightning](https://huggingface.co/SG161222/RealVisXL_V5.0_Lightning) for on-device Apple-Silicon inference with [SceneWorks / `mlx-gen`](https://github.com/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. A few-step distilled SDXL-Lightning photoreal checkpoint (openrail++, commercial-OK, ungated) — a standalone sibling of RealVisXL V5.0 tuned for ~5-step generation, roughly 6× faster than the 30-step base. Runs CFG-free by default; text-to-image only. ## 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). Convolutions, GroupNorms, and the CLIP token/position embeddings also stay dense; only the true Linear projections are packed. Quantization is byte-identical to `mlx-gen`'s load-time `nn.quantize` (bf16 cast, group 64). ## Usage ```rust use mlx_gen::{LoadSpec, WeightsSource, Quant}; let spec = LoadSpec::new(WeightsSource::Dir("…/realvisxl-lightning-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_Lightning](https://huggingface.co/SG161222/RealVisXL_V5.0_Lightning). See `LICENSE`.