This repository contains a converted MLX-loadable snapshot for Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers.
It is intended for use with the sanasprint-mlx runtime. The exported snapshot preserves the tokenizer, component configs, and safetensors keys expected by the native MLX loader.
This is a weights repository, not a standalone Diffusers pipeline and not a SANA-WM/world-model checkpoint.
Format
Format: sanasprint-mlx-snapshot
Format version: 1
Weight dtype: bfloat16
Components: text_encoder, transformer, vae
Runtime: native MLX via sanasprint-mlx
Target hardware: Apple Silicon with 16GB unified memory
Generate an image directly from this Hugging Face repo:
bash
1python -m sanasprint_mlx.cli.generate \2 --prompt "a tiny astronaut hatching from an egg on the moon"\3 --height 768 --width 768 --steps 2 --seed 42\4 --snapshot RayyTien/SanaSprint-0.6B-1024px-MLX \5 --allow-download \6 --output /tmp/sanasprint-mlx.png \7 --tiled-decode
The first run downloads the model files from Hugging Face. Later runs use the local Hugging Face cache.
Measured Smoke Tests
Measured on an Apple M4 with 16GB unified memory:
768x768, 2 steps, tiled decode from a cached HF snapshot: 19.51s for a 3-image warm batch
Batch peak footprint: 5.57 GiB
Single 768x768 first run including download: 118.89s
Single first-run peak footprint: 5.63 GiB
Prompt Notes
The underlying SanaSprint model follows English visual prompts best. Traditional Chinese prompts can work, but mixed prompts with English visual keywords are more reliable for specific materials, lighting, and scene details.
Scope
Supported:
Text-to-image generation
Native MLX text encoding, transformer denoising, and VAE decode
Tiled decode and warm batch generation
Not supported by this checkpoint:
SANA-WM
Video generation
Action-conditioned world-model rollout
Diffusers SanaSprintPipeline loading directly from this converted repo
License and Attribution
These are converted model weights derived from Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers. The original model card lists NSCL v2-custom / NVIDIA License terms and Gemma terms for the text encoder. Those upstream terms continue to govern the converted weights.
