software-mansion/react-native-executorch-sdxs-512-dreamshaper

Introduction

This repository hosts the SDXS-512-DreamShaper model for the React Native ExecuTorch library. SDXS-512-DreamShaper is a distilled, single-step latent diffusion text-to-image model. It is provided as a single multi-method .pte program per backend (XNNPACK, Core ML, MLX), ready for use in the ExecuTorch runtime.

Use it via the useTextToImage hook / models.textToImage.SDXS_512_DREAMSHAPER constant. If you'd like to run these models in your own ExecuTorch runtime, refer to the official documentation.

Compatibility

If you intend to use this model outside of React Native ExecuTorch, make sure your runtime is compatible with the ExecuTorch version used to export the .pte files. These models were exported for the v0.10.0 release; no forward compatibility is guaranteed. When using React Native ExecuTorch, the library constants guarantee compatibility with the runtime used behind the scenes.

Pipeline contract

Each backend's .pte exports three methods (see the per-backend config.json for exact shapes/dtypes):

• encode — CLIP text token ids int64 [1, 77] → text embeddings float32 [1, 77, 768].
• denoise — latents float32 [1, 4, 64, 64], timestep int64 [1], embeddings float32 [1, 77, 768] → predicted noise float32 [1, 4, 64, 64].
• decode — latents float32 [1, 4, 64, 64] → image float32 [1, 3, 512, 512] in [0, 1].

The single-step scheduler (linear in the latents and the predicted noise) and CLIP tokenization run on the client.

Repository structure

• xnnpack/, coreml/, mlx/ — one multi-method .pte per backend plus a config.json declaring its variants and method contract.
• tokenizer.json, tokenizer_config.json — the CLIP tokenizer, in the repository root.
• config.json — top-level model descriptor.