Model reference · open weights
stable-diffusion-3.5-medium-amdnpu-medusa is an open-weight image model from Stability AI.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the stabilityai/stable-diffusion-3.5-medium-amdnpu-medusa model card, 2026-10-29
What it is
| Released by | Stability AI |
|---|---|
| Released | 2026-10-26 |
From the model card
"Stable Diffusion 3.5 Medium is an improved version of SD3 Medium, featuring better prompt adherence, better generation quality, and more diverse outputs. SD3.5 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features improved image quality, typography, and prompt understanding." (stabilityai/stable-diffusion-3.5-medium)
This version of the model has been optimized to run on AMD NPUs.
For more information, refer to the original model card under stabilityai/stable-diffusion-3.5-medium.
M1.0 variant: This repository contains the M1.0-optimized NPU model derived from amd/stable-diffusion-3.5-medium-amdnpu.
The folder structure is organized around the shared (common) pipeline components and the text-to-image transformer (normal).
├─ common/
│ ├─ scheduler/
│ ├─ text_encoder/
│ ├─ text_encoder_2/
│ ├─ text_encoder_3/
│ ├─ text_encoder_3_int4/
│ ├─ tokenizer/
│ ├─ tokenizer_2/
│ ├─ tokenizer_3/
│ ├─ vae_decoder/
│ └─ vae_encoder/
└─ normal/
└─ transformer/
The scheduler folder contains the scheduler configuration used during the diffusion sampling process.
The text_encoder, text_encoder_2, text_encoder_3, and text_encoder_3_int4 folders contain the CLIP and T5 text encoder models (including the INT4 T5 variant) used to convert the input prompt into conditioning embeddings.
The tokenizer, tokenizer_2, and tokenizer_3 folders contain the tokenizer configuration and vocabulary files required to preprocess text prompts for the respective text encoders.
The transformer folder contains the MMDiT transformer model used in the diffusion process. The transformer is exported and structured specifically to leverage the AMD NPU accelerator for the denoising steps.
The vae_decoder 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.
The vae_encoder folder contains the VAE encoder model used for image-to-latent conversion.
Note: Transformer and VAE decoder models are optimized and structured to run on AMD NPUs. The other components (text encoders, tokenizers and scheduler) are shared between GPU and NPU pipelines, but are provided here for completeness.
| Model Details | Description |
|---|---|
| Person or organization developing model | AMD |
| Original model authors | Stability AI |
| Model date | October 2026 |
| Model version | 1.8.0 |
| Model type | Diffusion-based text-to-image generation model |
| Information about training algorithms, parameters, fairness constraints or other applied approaches, and features | This is a model that can be used to generate and modify images based on text prompts. |
| License | Stability AI Community |
| Where to send questions or comments about the model | Community Tab and AMD Developer Community Discord |
To get started with this model, visit github.com/amd/sd-sandbox.
AMD is committed to conducting our business in a fair, ethical and honest manner and in compliance with all applicable laws, rules and regulations. You can find out more at the AMD Ethics and Compliance page.
Please visit the original model card for more details: stabilityai/stable-diffusion-3.5-medium.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works
Running it yourself
Rent a machine by the hour. How to run this model is on its model card.