What is the SDXL model

The SDXL model is the official upgrade to the v1.5 model and is released as open-source software. It features significant improvements and enhancements over its predecessor.

Using Stable Diffusion SDXL on Think DIffusion, Upscaled with SD Upscale 4x-UltraSharp

Differences between SDXL and v1.5 models

The SDXL model architecture consists of two models: the base model and the refiner model.

The base model sets the global composition, while the refiner model adds finer details.

6.5 Billion (SDXL) vs 1 Billion Parameters (V1.5 Base)

The SDXL model incorporates a larger language model, resulting in high-quality images closely matching the provided prompts. The number of parameters on the SDXL base model is around 6.5 billion, compared to just under 1 billion for the V1.5 base model so we can expect some really good outputs!

Running the SDXL model with SD.Next

As of the time of writing, SDXLv0.9 can be used with the SD.Next web user interface. This is a fork from the VLAD repository and has a similar feel to automatic1111.

SDXL is not currently supported on Automatic1111 but this is expected to change in the near future. The good news is that the SDXL v0.9 can now be used on ThinkDiffusion.

Fire up a machine and select the SD.Next option with the Rapid machine

Workflow

The workflow is slightly different with SDXL due to the 2 models that can be used. You can generate images using the base model only:

Or you can use the base model together with the refiner model (recommended for better results)

Let's generate some images and see how SDXL compares!

Positive Prompts: portrait oils, beautiful paladien good and evil fantasy character portrait, ultra realistic, wide angle, intricate details, the fifth element artifacts, highly detailed by peter mohrbacher, hajime sorayama, wayne ba, cinematic, 4k, 8k uhd, dslr, soft lighting, high quality, film grain

Negative Prompts: (semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, cropped, out of frame, worst quality, low quality

  • (1) Sampler: Euler a
  • (2) Sampling Steps: 25
  • (3) Resolution: 1024 x 1024
  • (4) CFG Scale: 11

SDXL base model only image

SDXL Base model and Refiner

SD1.5 (TD-UltraReal model 512 x 512 resolution)

Positive Prompts: side profile, imogen poots, cursed paladin armor, gloomhaven, luminescent, haunted green swirling souls, evil inky swirly ripples, sickly green colors, by greg manchess, huang guangjian, gil elvgren, sachin teng, greg rutkowski, jesper ejsing, ilya kuvshinov, cinematic, 4k, 8k uhd, dslr, soft lighting, high quality, film grain

Negative Prompts: worst quality, low quality, blurry, b&w, greyscale, monochrome, jpeg artifacts, multiple view, multiple panel, signature, artist name, watermark, username, trademark, title, text, logo, label, cropped, out of frame

  • Sampler: Euler a
  • Sampling Steps: 25
  • Resolution: 1024 x 1024
  • CFG Scale: 11

SDXL base model only image

SDXL Base model and Refiner

SD1.5 (TD-UltraReal model 512 x 512 resolution)

Positive Prompts: photo, full body, 18 years old girl, punching the air, blonde hair, blue eyes, Italian, garden ,detailed face, 8k, raw, masterpiece

Negative Prompts: worst quality, low quality, double head, deformed, mutated, extra arms, duplicated, multiple, bad face,ugly face, extra fingers, blurry, signature, watermark, ((monochrome)), ((grayscale))

  • Sampler: Euler a
  • Sampling Steps: 25
  • Resolution: 1024 x 1024
  • CFG Scale: 11

SDXL base model only image

SDXL Base model and Refiner

SD1.5 (TD-UltraReal model 512 x 512 resolution)

If you’re having issues with SDXL installation or slow hardware, you can try any of these workflows on a more powerful GPU in your browser with ThinkDiffusion.

TRY THINKDIFFUSION

If you like what you're creating from SDXL but wished you could just swap a face easily, you're in luck check out my tutorial here on Roop and Face Swapping here.

Let me know what kind of stuff you're creating with this new version, and most of all have fun going XL!!