Sign In

Techniques for making good horizonal wallpaper images

17

Introduction

This wallpaper workflow assumes the subject is a person, and the aspect ratio of the wallpaper is horizontally oriented. Many monitors tend to have DPIs of around 90-140 but some particularly high quality ones may have DPIs up to 220 and upscale the UI for readability. A wallpaper that meets 100 DPI will be attractive but still have a reasonable file size. The following techniques can help refine and find suitable images without wasting resources and time rendering unsuitable candidates. Bad candidates are quickly discarded and good candidates will be upscaled x4 to reach the 100 DPI mark.

Start with a resolution that is approx. the same or greater than the base 512x512 for using a SD 1.5 model. If either dimension is less than 512 pixels, that can introduce both good and bad effects with the image. For less work refining an image, make the image at least 512 pixels high.

Pick any Sampler and Hires. Fix that works well with the model; there are several that tend to be preferred over others, and DPM++ 3M SDE Karras and DPM++ 2M Karras are very popular Samplers. Both R-ESRGAN 4x+ Anime6B and SwinIR_4x are very popular Hires. Fixers. This example, uses DPM++ 3M SDE Karras and R-ESRGAN 4x+ Anime6B as the Sampler and Hires. Fixer respectively.

In general, the following will alter the image in order of biggest changes to smallest. This will be utilized to find promising candidates before spending a lot of time upscaling the final wallpapers.

  1. Seed (completely changes image)

  2. Sampler (completely changes image)

  3. Sampling Steps (adds/removes/alters elements and details)

  4. CFG (adds/removes/alters elements and details)

  5. Denoising Strength (alters aesthetic from abstract to sharp)

  6. Batch Size (technically a bug, but will alter the image)

  7. Hires. Steps (once convergence is reached, or with low Denoising Strength)

One: Use Seed and Sampling Steps to find promising starts

First, generate a batch of candidate images by varying the Seed and the number of Sampling Steps. The various seeds are to show different results for your prompt, and the different Sampling Steps will show where the convergence begins and the images begin to more closely resemble each other. This will vary from model to model, but using 35, 50, 65, 80, and 100 Sampling Steps can be a good start (in automatic1111 this is expressed as "35,50,65,80,100" in the XYZ plot).

An ideal wallpaper should have an interesting subject in the center, but also a variety of colors at the edges. Because application windows often cover the center subject while the computer is in use, it's the details at the edges that will often be noticed. It helps to have the colors change, and not have a consistent edge that is always one or two colors. Seed 1004 looks like a good start and will possibly result in a good wallpaper.

Two: Use CFG and Sampling Steps to narrow down the candidates

From the previous example grid, seed 1004 created several good candidates with bright colors along the bottom and shifting colors up the sides. Start a second grid that uses that Seed, but vary the CFG combined with the Sampling Steps. The Sampling Steps range is changed to 30-50 in 5 step increments to focus on the examples that looked most promising (in automatic1111 this is expressed as "30-50(+5)" in the XYZ plot). The CFG is kept in the 7.5 to 12.5 range to maintain a balance between respecting the prompt but also introducing interesting details into the art (e.g. "7.5-12.5(+1)" in the XYZ plot).

This results in several good candidates for further testing. In the above example grid, Sampler Steps/CFG combinations that look promising are 30/8.5, 35/7.5, 35/8.5, 45/12.5 and 50/7.5.

Three: Use Hires. Fix Steps and Denoising to upscale x2 for final tests

Create another set of grids that uses the Hires. Fix steps and Denoising Strength to introduce some more variations. Create a grid for each combination of Sampling Steps + CFG that looks promising. To speed things up, only upscale x2 to create the test images.

From there, the x2 upscaled image should give an idea which image is the best candidate to do the final upscale to x4 to achieve the final resolution. Further upscaling does change the image, and it can introduce anatomical errors.

As a general rule, use around 10-15 Hires Fix steps. If Sampling Steps is over 50, try to make Hires Fix steps to be 25 or about half of the Sampling Steps.

Pick at least one candidate from each of the grids (each XYZ grid is the result of testing Sampler Steps, CFG, Hires. Fix Steps, and Denoising) and fully upscale at x4 to create a final wallpaper candidate.

Some of these wallpapers had anatomical errors or clothing artifacts that were difficult to remove, but use of the Batch Size bug/feature generated similar wallpapers without the errors. The cover image of this article is an example of anatomical errors that could not be easily fixed.

The above final five wallpaper candidates are available at this post.

References

User @abraxo created this sample prompt and image.
The prompt was modified and used with the Sardonyx REDUX model.

5

Comments