stable-diffusion-3.5-medium-amdnpu-medusa: File List

stable-diffusion-3.5-medium-amdnpu-medusa: File List

At a Glance

Item Value
Repository stabilityai/stable-diffusion-3.5-medium-amdnpu-medusa
Publisher guide Stability AI: models and licenses
Published 2026-10-27
License stabilityai-ai-community
Formats ONNX
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Stable Diffusion 3.5 Medium, optimized for operation on AMD NPUs, has been released. This is a text-to-image generation model employing the Multimodal Diffusion Transformer (MMDiT) architecture, which generates images based on text prompt inputs.

Specifications

This model is specifically optimized for operation on AMD NPU accelerators. The main specifications and configuration are as follows:

  • Architecture: Multimodal Diffusion Transformer (MMDiT)
  • Model Type: Diffusion-based text-to-image generation model
  • Optimization Target: AMD NPU
  • Distribution Format: ONNX
  • License: Stability AI Community
  • Variations: This repository includes the “M1.0-optimized NPU model" derived from amd/stable-diffusion-3.5-medium-amdnpu.

The repository structure is organized into a common folder for shared components and a normal folder for text-to-image conversion.

  • Transformer: Located in the normal/transformer folder, it is exported and structured to utilize AMD NPU accelerators during the denoising steps.
  • VAE decoder: Located in the common/vae_decoder folder, it is structured to use NPU accelerators for efficient image reconstruction.
  • Text encoders: Includes common/text_encoder, text_encoder_2, text_encoder_3, and text_encoder_3_int4. These are CLIP and T5 text encoders (including an INT4 version of T5) used to convert input prompts into conditioning embeddings.
  • Tokenizers: Includes common/tokenizer, tokenizer_2, and tokenizer_3, containing configurations and vocabulary files for preprocessing prompts for each text encoder.
  • Scheduler: The common/scheduler folder contains scheduler settings used in the diffusion sampling process.
  • VAE encoder: The common/vae_encoder folder contains the model for converting images into latent representations.

While the Transformer and VAE decoder are optimized for NPU operation, other components (text encoders, tokenizers, scheduler) are designed to be shared across GPU and NPU pipelines.

Performance

The stable-diffusion-3.5-medium model, which serves as the base for this version, is positioned as an improved version of the previous SD3 Medium.

According to the model card, this model is reported to have improved prompt adherence, generation quality, and output diversity. It is also stated that improvements in typography (text rendering) and prompt understanding have been achieved.

Strengths and Use Cases

This model is suitable for generating images based on text prompts or modifying existing images. Inheriting the characteristics of the base Stable Diffusion 3.5 Medium model, it enables high-fidelity generation relative to prompts and high-quality, diverse outputs. In particular, it is well-suited for image generation involving complex instructions due to improved typography accuracy and prompt understanding.

The most significant feature is its optimization for AMD NPU accelerators. This is intended to enable efficient text-to-image generation processes in environments such as PCs equipped with compatible hardware.

Distributed Files

Weight files published in stabilityai/stable-diffusion-3.5-medium-amdnpu-medusa, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
common/text_encoder/model.onnx 248MB
common/text_encoder_2/model.onnx 1.39GB
common/text_encoder_3/model.onnx 1MB
common/text_encoder_3_int4/model.onnx 1MB
common/tokenizer/model.onnx 1MB
common/tokenizer_2/model.onnx 1MB
common/tokenizer_3/model.onnx 1MB
common/vae_decoder/dynamic/dd/replaced.onnx 0MB
common/vae_encoder/model.onnx 69MB
common/vae_encoder/optimized.onnx 69MB
common/vae_encoder/optimized_1024x1024.onnx 69MB
common/vae_encoder/optimized_512x512.onnx 69MB
normal/transformer/dynamic/dd/replaced.onnx 2MB

How to Get It

This model is distributed in ONNX format. It can be obtained from the following Hugging Face repository:

  • Repository: stabilityai/stable-diffusion-3.5-medium-amdnpu-medusa

Additionally, the sd-sandbox repository provided by AMD is recommended as an environment for verifying the operation of this model.
– github.com/amd/sd-sandbox

Since this model is not gated, no license agreement procedure is required for downloading.

Can You Run It Locally?

The publisher distributes this model as ONNX.

License — stabilityai-ai-community: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (stabilityai-ai-community). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Other Models for the Same Task

Recent image generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.

See all image generation models →

What to Read Next

Sources