SceneWorks/illustrious-xl-v1-mlx

Illustrious-XL v1.0 — MLX pre-quantized tiers

Pre-quantized, packed-load tiers of OnomaAIResearch/Illustrious-XL-v1.0 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.

Illustrious-XL is a Danbooru-tag anime SDXL finetune (OnomaAI). It is architecturally vanilla SDXL: dual CLIP-L + OpenCLIP-bigG text encoders, real classifier-free guidance + negative prompt, eps prediction, VAE scaling factor 0.13025, and full sdxl-family LoRA support. ~30 steps at guidance 7.0, native 1024×1024, and it handles wide frames up to 1536×1536.

Provenance

Upstream ships a single-file LDM checkpoint (Illustrious-XL-v1.0.safetensors), which the MLX sdxl loader cannot read. These tiers were produced offline from that checkpoint with scripts/build_sdxl_turnkey.py: StableDiffusionXLPipeline.from_single_file → diffusers component tree → per-tier quantization. The component configs are the canonical SDXL descriptors (adopted verbatim from a known-good SDXL turnkey after an architecture-key match), not from_single_file's output.

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 (f16 source mirror)	nothing

The VAE stays dense 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).

License

SDXL license — CreativeML Open RAIL++-M, per the upstream model card. Commercial use OK, ungated; behavioral-use restrictions apply.