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

sdxs-512-dreamshaper

This repository hosts the sdxs-512-dreamshaper models exported for the React Native ExecuTorch library as ExecuTorch .pte programs, ready to run on device.

Upstream model: SDXS-512-DreamShaper

Variants

Path	Backend	Precision
coreml/sdxs_512_dreamshaper_coreml_fp16.pte	coreml	fp16
vulkan/sdxs_512_dreamshaper_vulkan_fp16.pte	vulkan	fp16
xnnpack/sdxs_512_dreamshaper_xnnpack_fp32.pte	xnnpack	fp32
xnnpack/sdxs_512_dreamshaper_xnnpack_int8.pte	xnnpack	int8

XNNPACK int8

xnnpack/sdxs_512_dreamshaper_xnnpack_int8.pte uses dynamic int8 activations with int8 per-channel weights on every conv and Linear of denoise and decode; encode stays fp32. 16 images (8 prompts x 2 seeds), CLIP ViT-B/32 image-prompt score:

	CLIP score	Galaxy S26 Ultra, per image
fp32	34.20	1323 ms
int8	34.94	845 ms

Unshipped backends

A backend directory carrying a NOTES.md and no .pte was exported and not published. Its note records what went wrong.

• mlx (why)

Repository structure

config.json                                    41 B
coreml/config.json                             2.0 kB
coreml/sdxs_512_dreamshaper_coreml_fp16.pte    840 MB
mlx/NOTES.md                                   2.8 kB
tokenizer.json                                 3.5 MB
tokenizer_config.json                          399 B
vulkan/config.json                             2.0 kB
vulkan/sdxs_512_dreamshaper_vulkan_fp16.pte    840 MB
xnnpack/config.json                            3.7 kB
xnnpack/sdxs_512_dreamshaper_xnnpack_fp32.pte  1.6 GB
xnnpack/sdxs_512_dreamshaper_xnnpack_int8.pte  776 MB

Compatibility

These files are published for the ExecuTorch v1.4.1 runtime. ExecuTorch gives no forward compatibility guarantee, so an older runtime may fail to load them.

To use them in React Native ExecuTorch, pass the model constant shipped in the library's model registry to the corresponding task pipeline. See the documentation.

To load these files in your own ExecuTorch runtime, read the compatibility note first.

Model details

Each backend's .pte exports three methods, and the diffusion loop is split between the program and the client. The program owns encode, denoise and decode; the client owns CLIP tokenization and the single-step scheduler, which is linear in the latents and the predicted noise. Neither of those is in the graph, so a caller that only runs the three methods will not produce an image.

Exact shapes and dtypes for the three methods are declared in each config.json.