SDXL 1.0: The Good, The Bad and The Ugly

A new era of generative AI for everyone

Image by the author

On 26 July 2023, Stability.AI announced the public release of SDXL 1.0, another big update to its generative models with breaking changes. The initial responses from the community is largely positive and people are now experimenting the capabilities of the new architecture.

SDXL 1.0 consists of two different models:

• SDXL-base-1.0: A base text-to-image model to generate 1024 x 1024 images. The base model uses OpenCLIP-ViT/G and CLIP-ViT/L for text encoding.
• SDXL-refiner-1.0: A image-to-image model to refine the latent output of the base model for generating higher fidelity images. The refiner model only uses the OpenCLIP-ViT/G model.

SDXL 1.0 are licensed under the permissive CreativeML Open RAIL++-M license. Read the technical report here.

This article covers some of my personal opinions and facts related to SDXL 1.0

The Good

Image quality

By default, SDXL 1.0 can generate 1024 x 1024 images natively. This alone is a big improvement over its predecessors. SDXL 1.0 is particularly well-tuned for vibrant and accurate colors. In addition, you can now generate images with proper lighting, shadows and contrast without using the offset noise trick. For example:

Prompt:
anime artwork An astronaut riding a white horse . anime style, key visual, vibrant, studio anime, highly detailed

Negative prompt:
photo, deformed, black and white, realism, disfigured, low contrast

Image by the author

Besides that, the subject is usually placed at the center of the image, making it the main focus. For beginners, this feature comes in handy as they can use the generated images right away.

Based on a local experiment, SDXL 1.0 seems to be a better base model compared to SD 1.5 and SD 2.1 especially on artistic styles. The image composition more or less stay the same while the artistic styles differs when prompting with art by <artist name>. For example:

Prompt:
An astronaut…