Testing SDXL 0.9: Image Quality, Clipdrop and a Midjourney Comparison

In June 2023, I take a look at Stability AI’s SDXL 0.9 preview and compare a few of its outputs with Midjourney. I am impressed by the apparent detail and composition, though my selections and judgments are informal rather than a controlled test.

Stability AI describes 0.9 as an improvement over its April beta. The company shows the same prompts side by side: an alien scene has more detail and color in the newer example, and a wolf in Yosemite looks sharper to me, down to its fur and eyes. These are selected demonstrations, not a measurement of typical performance.

Another comparison shows a hand holding takeaway coffee. I point out extra fingers in the earlier result and praise the newer example. A good-looking hand in one image does not mean either SDXL or Midjourney has solved hand anatomy for all prompts.

I discuss the SDXL family’s image-to-image prompting, inpainting to replace portions of an image and outpainting to extend its boundaries. Availability depends on the particular implementation and service; not every feature is demonstrated as part of the base model itself.

Stability AI reports a 3.5-billion-parameter base model and a 6.6-billion-parameter ensemble pipeline, compared with 3.1 billion for the beta. It describes two text encoders and a base-and-refiner approach, with output up to 1024 × 1024. Parameter counts alone do not establish output quality ([Stability AI](https://stability.ai/news-updates/sdxl-09-stable-diffusion)).

The 2023 guidance I read lists Windows 10 or 11 or Linux, 16 GB of system RAM and an Nvidia RTX 20-series GPU with 8 GB of VRAM. These are historical vendor recommendations, not a promise that every configuration or workflow will run identically; optimizations and model options can change memory requirements.

At the time, SDXL 0.9 was available through Clipdrop and the Stability API, with research access ahead of a planned 1.0 release. Although I call it open source, access to the 0.9 weights had specific limits: the [SDXL 0.9 Research License](https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9) grants use for non-commercial research subject to its terms. Availability of code does not automatically make those weights unrestricted.

Clipdrop displays examples across anime, realistic and stylized imagery. I consider several comparable in appearance to Midjourney. The page in the video shows a free tier with 400 watermarked SDXL images per day and a paid option; those quotas, watermarks and prices describe June 2023 and should not be assumed current.

For an informal comparison, I take a surreal “pill head” prompt from Midjourney and try it in Clipdrop. The images differ, as expected from different models and generation conditions. I find the results artistic, but a handful of selected outputs cannot establish which model is generally better.

I next try a fingerprint prompt. The first results are black and white, so I add blue and red to better match the example I am comparing. I like the new images and show one next to the Midjourney result; changing a prompt is useful experimentation, but it also means the comparison is not a fixed-condition benchmark.

For a lion prompt, an initial result looks more artistic than the realism I wanted. A subsequent set gives me the detailed, realistic appearance I was looking for. Different generations can vary even when the general subject is the same.

Clipdrop also offers tools such as background removal, cleanup, relighting, upscaling and uncropping. I demonstrate Relight on a generated portrait: I move red, blue and green lights, change their intensity, radius and distance, and watch shadows and illumination update before downloading the image.

I am excited by the apparent progress and the prospect of running models locally. Local generation avoids a hosted daily quota but still depends on hardware, software and applicable license terms; it is not literally without resource limits. My enthusiasm for the comparison reflects these examples, not an exhaustive evaluation.