* This blog post is a summary of this video.

Nvidia and Stability AI Release Optimized Stable Diffusion XL for Faster AI Image Generation

Author: PixovertTime: 2024-03-23 17:15:01

Table of Contents

Stable Diffusion XL 1.0 Collaboration Between Stability AI and Nvidia Improves Performance and Image Quality

Stable Diffusion XL has been a huge hit since it launched, and we now have a new version - Stable Diffusion XL 1.0 with TensorRT. This optimized version is a collaboration between Stability AI and Nvidia, and it provides substantial improvements in speed and efficiency.

On the Stability AI Hacking Face page, there are some sample images generated with the new SDXL 1.0. The image quality looks on par with the original SDXL, with crisp examples like a fighter jet, cheetah in motion, and a cat ready to race.

Compared to the original Stable Diffusion XL, the collaboration with Nvidia resulted in a 13% faster timing at 1024x1024 resolution over 30 steps when using a standard GPU. The improvements get even better with Nvidia's specialized AI accelerators - up to a 41% timing improvement with the H100!

Significant Performance Improvements

Looking closer at the performance numbers, SDXL 1.0 displayed a 20% higher image throughput over 30 steps at 1024x1024 resolution on a standard GPU. With Nvidia's A10 accelerator, the improvement jumps to 33%, while the powerful A100 sees a 26% speedup. The $35,000 H100 takes the cake with an incredible 70% better image throughput. Of course, not everyone can afford the top-of-the-line H100. But the benchmark clearly shows the optimizations provide excellent speedups, especially when leveraging Nvidia's AI specialized hardware.

Image Quality On Par with Original

The sample images from SDXL 1.0 showcase the continued stellar image generation capabilities. We see imaginative, intricate images like a fighter jet, cheetah, cat, and athlete generated with precision and detail. So while performance gets a healthy upgrade thanks to TensorRT optimizations for Nvidia hardware, Stability AI ensures there is no compromise on SDXL's industry-leading image quality.

Increase Image Size 64x with Stable Diffusion 1.5 for Detailed Blow Ups

Stable Diffusion 1.5 delivers an unexpected capability - enlarging images massively while preserving clean details. As an example, a 512x512 image of the Mona Lisa is transformed into a scenic vision of Atlantis, then scaled up 64 times to deliver a staggeringly detailed final image.

This end-to-end workflow leverages depth mapping, creative AI generation, and multi-step blowing up powered by models like RealESRGAN. The result is an image enlarged 56 times yet retaining intricate details - whether it's the textures on Atlantis statues or precision-etched stairs.

Starting Small with Mona Lisa...

The input is a common 512x512 image - in this case, the Mona Lisa portrait. This is converted into a depth map to establish shapes and contours. Then Stable Diffusion's imaginative image generation takes over, transforming the depth map into a majestic Atlantis background scene with steps, statues, and stone architecture. So already the foundation is set with a high-quality 512x512 AI-generated image, primed for scaling up to far larger dimensions.

...Ending with Gigantic Atlantis Statue

The AI-powered image blow up process utilizes models like RealESRGAN for each incremental size increase. With each step up in size, details are preserved while image artifacts are minimized. In the end, the 512x512 starting image is enlarged to a staggering 29,000 x 29,000 pixels - 56 times bigger in each dimension! Yet the final image retains crisp, clear details in even tiny crevices and textures. This technique demonstrates Stable Diffusion 1.5's versatility in creating and intelligently manipulating images while keeping intricate details intact.

New Advanced Stable Diffusion Course Covers Diffusion Methods

For those looking to take their Stable Diffusion skills even further, check out the new Advanced Stable Diffusion course on Udemy under the Pixelbook Studio channel. Spanning over 4 hours, this course dives deeper into generative AI using ControlNet, ControlLoris and other advanced techniques within Automatic1111's Stable Diffusion web UI.

As part of the launch discount for YouTube subscribers (see video description for coupon), the course is available for only $12.99 for 3 days - so enroll quickly to lock in the reduced pricing.

Fine-Tune Control with ControlNest and Hypernetworks

A major focus of the course is leveraging ControlNet and ControlLoris to gain more granular steering over your Stable Diffusion image generations. You'll learn approaches like hypernetwork embeddings which alter AI model weights on-the-fly to specialize behavior for niche prompt domains.

Latest Advanced Text-To-Image Methods

The course also showcases bleeding edge text-to-image techniques to push creative boundaries. Explore approaches like automatic masking and editing flows to streamline working with generated images. There are also tips to efficiently explore new prompt concepts while avoiding model bias and artifacts.

Nvidia RTX 4060 TI GPU Now Widely Available for AI and Creative Work

Last month saw the consumer launch of Nvidia's new RTX 4060 TI graphics card, bringing the latest Ada Lovelace architecture to more affordable price points. Fortunately, supply constraints seen with earlier RTX 4000 series models are not an issue with the 4060 TI.

There are ample stock options on Amazon from reliable brands like Asus, MSI, Zotac and Gigabyte. The roughly $500 starting price represents a solid value for AI creators and video editors given its combination of rendering speed, 16GB of VRAM and advanced capabilities like hardware-accelerated AI inference.

Most Models Retail Between $500 and $550

Pricing for the RTX 4060 TI models from Asus, MSI and other brands largely clusters around Nvidia's $500 MSRP. There are occasional deals under $500, such as one MSI Ventus model priced at $480.

Great for Video Editing, 3D Modeling or Game Streaming

With high-speed 16Gbps GDDR6 memory and upgraded NVENC encoding, the RTX 4060 TI makes an excellent choice for creative workflows like 4K+ video editing, 3D modeling/CAD, game streaming, and more. It can also accelerate AI tasks like inference/enhancements.


The latest updates to Stable Diffusion XL and emergence of new Scale models signify an exciting time for generative AI. We're seeing rapid performance improvements to enable higher resolution, larger batchgenerations while image quality continues to impress.

For creators, expanding mainstream availability of new GPUs like the RTX 4060 TI at just over $500 makes adopting AI more accessible. We can expect continued enhancements to models like Stable Diffusion along with hardware advancements to push boundaries on outputs.


Q: How much faster is Stable Diffusion XL 1.0?
A: Up to 70% faster image generation compared to previous version according to Nvidia's benchmarks, with the highest speed gains on the Nvidia H100 GPU.

Q: What image size can I increase using Stable Diffusion 1.5?
A: The example in the video increased a 512x512 image to over 8000x8000 pixels, a 64x enlargement. Stable Diffusion 1.5 can increase small images to gigantic sizes with convincing detail.

Q: What advanced Stable Diffusion topics are covered in the new course?
A: The course covers advanced techniques like ControlNest and Hypernetworks for better image control, as well as methods to get better results from text prompts.

Q: Where can I buy an Nvidia RTX 4060 TI GPU?
A: The RTX 4060 TI launched last month and is now widely available from Amazon, Newegg, and other major retailers for around $500.

Q: What are the best RTX 4060 TI models for AI and video editing?
A: The MSI and Zotac models offer a good balance of price and reliability. Over 16GB of VRAM is recommended for best performance.

Q: Does the RTX 4060 TI come with any gaming bundles?
A: Select models come with the Overwatch 2 Invasion ultimate bundle, check listings for details.

Q: How much is the advanced Stable Diffusion course discount?
A: YouTube subscribers get a special discounted price for the new advanced Stable Diffusion course on Udemy. Channel members get an additional discount.

Q: What hardware do I need to run Stable Diffusion XL 1.0?
A: You'll need an Nvidia GPU for best performance - the optimizations are specifically for Nvidia's TensorRT platform. An RTX 4060 TI is sufficient for most users.

Q: Can I increase image size with DALL-E or Midjourney?
A: No, the extreme image scaling demo was done in Stable Diffusion 1.5. Other AI image models may not support this feature.

Q: Where can I learn more about the new Nvidia GPUs?
A: Check the video description for links to Amazon product listings and recommendations for the latest Nvidia graphics cards.