Add SANA MLX mirror (verbatim repackage of NVIDIA diffusers checkpoint into SanaPipeline::from_snapshot layout)
eef346e verified | license: other | |
| license_name: nvidia-open-model-license | |
| license_link: LICENSE | |
| library_name: mlx | |
| pipeline_tag: text-to-image | |
| tags: | |
| - sana | |
| - text-to-image | |
| - mlx | |
| - sceneworks | |
| - non-commercial | |
| # SANA 1600M 1024px — SceneWorks MLX mirror | |
| Un-gated, repackaged mirror of [`Efficient-Large-Model/Sana_1600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_diffusers) | |
| for the [SceneWorks](https://github.com/SceneWorks) native MLX worker (`mlx-gen-sana`). | |
| The tensors are **copied verbatim** (dtype-preserving — no conversion, no | |
| re-quantization) into the directory layout that `SanaPipeline::from_snapshot` | |
| expects: | |
| ``` | |
| transformer/diffusion_pytorch_model.safetensors # SANA 1.6B Linear-DiT trunk (F16) | |
| vae/diffusion_pytorch_model.safetensors # 32x DC-AE f32c32 decoder (F32) | |
| text_encoder/gemma-2-2b-it.safetensors # gemma-2-2b-it CHI caption encoder (BF16) | |
| text_encoder/tokenizer.json | |
| ``` | |
| The gemma-2-2b-it text encoder is sourced from the un-gated | |
| [`SceneWorks/gemma-2-2b-it`](https://huggingface.co/SceneWorks/gemma-2-2b-it) | |
| mirror (merged from its two shards into one file) so the SANA snapshot is | |
| self-contained. | |
| ## License — NON-COMMERCIAL | |
| SANA is distributed under the **NVIDIA Open Model License** (see [`LICENSE`](LICENSE) | |
| and [`NOTICE`](NOTICE)). This is a non-commercial license: the model and its | |
| outputs are for **research and evaluation use only**. By downloading these | |
| weights you agree to the NVIDIA Open Model License terms. The bundled gemma-2-2b-it | |
| encoder is under the Google Gemma Terms of Use. | |
| This mirror adds no additional grant and alters no weights. | |