Image Generation NPU Models – SD 1.5
This repository contains the ONNX models and runtime libraries required to run the image generation pipeline on AMD NPUs.
The folder structure is organized to mirror the main components of the diffusion pipeline (UNet, VAE decoder, text encoder, tokenizer and scheduler), plus the platform-specific runtime libraries.
Repository structure
.
├─ libs/
├─ scheduler/
├─ text_encoder/
├─ tokenizer/
├─ unet/
└─ vae_decoder/
libs/
This folder contains the dynamic libraries (.dll) required at runtime by the NPU backend.
They must be placed in a location where the application can load them (e.g., in the same folder as the executable or in the system PATH).
unet/
This folder contains the UNet model used in the diffusion process. The UNet is exported and structured specifically to leverage the AMD NPU accelerator for the denoising steps.
vae_decoder/
This folder contains the VAE decoder model used to map latent representations back to the image space. The VAE decoder is also structured to make use of the NPU accelerator for efficient image reconstruction.
text_encoder/
This folder contains the text encoder model used to convert the input prompt into conditioning embeddings for the diffusion model.
tokenizer/
This folder contains the tokenizer configuration and vocabulary files required to preprocess the text prompt before it is fed to the text encoder.
scheduler/
This folder contains the scheduler configuration (timesteps, betas, alphas, etc.) used during the diffusion sampling process.
Release 1113
This release corresponds to the 1113 build.
Included in this version:
- Updated UNet and VAE Decoder models optimized for AMD NPU execution.
- Synchronized text encoder, tokenizer, and scheduler components aligned with the 1113 pipeline.
- Updated runtime DLLs in the
libs/folder. - Improved model folder structure for compatibility with Procyon and NPU execution environments.
Notes:
- All ONNX models in this release are validated with the 1113 test package.
- Ensure that the DLLs from
libs/are correctly placed in the application’s search path. - This release is intended for NPU execution; GPU versions are hosted separately.
Notes
- UNet and VAE decoder models are optimized and structured to run on AMD NPUs.
- The other components (text encoder, tokenizer and scheduler) are shared between GPU and NPU pipelines, but are provided here for completeness.
- Please refer to the associated application or benchmark documentation for detailed integration and usage instructions (e.g., how to set model paths, environment variables and library search paths).