AMD has released a custom-optimized version of the Stable Diffusion 3 Medium model, marking a significant step in bringing high-quality generative AI capabilities directly to user devices.
Highlights
- Local AI Generation: AMD launches a locally run version of Stable Diffusion 3 Medium, optimized for Ryzen AI 300-series laptops with no cloud dependence.
- High-Resolution Outputs: The model uses a two-stage pipeline to generate 1024×1024 images and upscale them to 2048×2048 (4MP) using XDNA Super Resolution.
- Efficiency First: Runs efficiently on devices with 24GB RAM, consuming just 9GB thanks to BF16 optimization—making powerful AI generation accessible on consumer hardware.
- Enhanced Prompting: Improved adherence to detailed prompts allows for more precise and controllable image outputs, supporting structured input and negative prompting.
- No Subscriptions: Available for free through Amuse 3.1 (via Tensorstack), with zero subscription fees or cloud requirements—targeting creators who value privacy and mobility.
- Performance Without Internet: Fully offline capability benefits remote professionals, rural users, or those in secure/low-bandwidth environments.
Built in collaboration with Tensorstack and integrated into the Amuse 3.1 desktop application, the new release enables offline image generation at 4-megapixel resolution—exclusively on laptops powered by AMD’s latest Ryzen AI processors.
Local Generation at Scale
Unlike most generative image models that rely on cloud infrastructure, AMD’s SD3 Medium model runs entirely on-device, powered by the new XDNA 2 NPUs featured in Ryzen AI 300-series chips.
This shift allows users to generate images locally—without internet connectivity—offering low-latency performance, increased privacy, and flexibility for creators working in bandwidth-limited or secure environments.
How It Works?
The system follows a two-step pipeline,
- Stage 1 – Generates images at 1024×1024 resolution.
- Stage 2 – Uses XDNA Super Resolution to upscale outputs to 2048×2048 (4MP) via the NPU.
This provides print-quality visuals entirely offline—an upgrade from earlier local solutions like Stable Diffusion XL Turbo, which required more memory and heavier hardware.
Efficient and Scalable Performance
The new implementation stands out for its low resource requirements. It runs on laptops with just 24GB of RAM, consuming only 9GB during operation, thanks to AMD’s use of block floating point 16 (BF16)—a memory-efficient format that helps reduce computational load without compromising output quality.
According to AMD, this is the first BF16-optimized version of SD 3.0 Medium, tailor-made for on-chip acceleration via the XDNA 2 NPUs.
BF16 support enables complex image generation workflows on consumer-grade laptops without requiring cloud access or high-end GPUs.
Precision Prompting and User Control
The model also introduces enhanced prompt adherence, giving users more control over the final output. AMD encourages a structured input method,
- Start with: image type
- Then add: layout or composition
- Finally include: contextual details
Support for negative prompts and fine-tuned adjustments—even through subtle changes like punctuation—provides advanced users with tools for deeper stylistic control.
Accessible and Subscription-Free
The model is available for free through Tensorstack’s Amuse 3.1 beta, with no subscription fees and no requirement for cloud access. This positions AMD’s offering as a competitive solution for independent creators, designers, and professionals who need flexible tools without ongoing costs.
Asset creation for branding, marketing, stock visuals, or on-the-go design tasks—from rural studios to airplane cabins.
