| --- |
| license: other |
| license_name: stabilityai-ai-community |
| license_link: LICENSE.md |
| base_model: stabilityai/stable-diffusion-3.5-medium |
| pipeline_tag: text-to-image |
| tags: |
| - text-to-image |
| - stable-diffusion |
| - stable-diffusion-3 |
| - mlx |
| - sceneworks |
| --- |
| |
| # Stable Diffusion 3.5 Medium — SceneWorks MLX quant matrix |
|
|
| Pre-built, turnkey MLX tiers of **Stability AI Stable Diffusion 3.5 Medium** — the 2.5B MMDiT-X (dual-attention) mid-tier, renders up to 1440² for |
| [SceneWorks](https://github.com/SceneWorks/SceneWorks), hosted as separate downloadable artifacts so |
| the app loads a tier directly with **no on-device conversion peak** (sc-8513, epic 8506). |
|
|
| Each subdirectory is a complete `from_snapshot`-loadable diffusers tree — packed/dense |
| `transformer/` plus the dense triple text encoder (CLIP-L + CLIP-G + T5-XXL), tokenizers, and 16-ch |
| VAE: |
|
|
| | Tier | `transformer/` precision | Use | |
| | --- | --- | --- | |
| | `q4/` | Q4 (asymmetric group-affine, group-64) — **default** | lowest memory | |
| | `q8/` | Q8 (asymmetric group-affine, group-64) | near-lossless | |
| | `bf16/`| dense bf16 (the source backbone, no quantization) | highest fidelity, high-memory Macs | |
|
|
| Only the MMDiT transformer is quantized; the text encoders and VAE are reused dense in every tier. |
| The packed weights auto-detect their quantization on load. |
|
|
| ## License & attribution |
|
|
| This model is a derivative of **Stable Diffusion 3.5 Large** and is distributed under the |
| **[Stability AI Community License](LICENSE.md)** (free for research, non-commercial, and commercial |
| use by organizations with under \$1M in annual revenue; see the license for terms). |
|
|
| > This Stability AI Model is licensed under the Stability AI Community License, Copyright © Stability |
| > AI Ltd. All Rights Reserved. |
|
|
| **Powered by Stability AI.** |
|
|
