ComfyUI Guide: Z-Image Image2Image Detailing and Upscale

I am going to show a relatively simple way to add details using the Z image model. This workflow loads an image, then upscales and adds details.

Scaling for Z-Image Upscaling and Detailing

Decide how much to scale the image. Usually 2x is safe. You do not want to go too high because the Z image model I think is trained to a maximum of 48x48. If you go over that you will see artifacts start to appear.

If you want a higher scale, increase the rows and columns of the tiling nodes. Do that for the untiled too to match when restitching.

Sharpen and Grain in Z-Image Upscaling and Detailing

Go to the sharpen and add grain stages one at a time. Make sure the image is not oversharpened. You want it a little sharpened because you lose a bit of sharpness when you upscale using Lanczos. If you use nearest neighbor it is just going to be super pixelated.

You get a little sharpness back by sharpening after the upscale. Grain is a way to add fake detail into the model so it thinks there is texture and the image is not flat. Do not add too much grain because you will get a grainy image. Do not add too little because the model will think it is super smooth.

With a 3D picture the model will just make it look 3D still, so we want it to look real. You can adjust these depending on the size of the image. The grain will look bigger or smaller. In this case the input image is 1080 by 1448, and this is what worked for this image. You might bump it up or down depending on the image.

Prompting for Z-Image Upscaling and Detailing

You could add an autoprompter, but it is pretty easy to label the basic parts of the image. For example, an Asian woman wearing a red turtleneck sweater.

Steps, Shift, and Artifacts in Z-Image Upscaling and Detailing

I set steps to 18. We are not actually running for 18. We are running from 14 to 18, so it is four steps. The reason to bump it to 18 is because if you are denoising fully you would run around nine steps at a shift of three, which is the default.

In this case, bumping to 18 allows a higher shift, which is a max of six, without getting artifacts. If you want a more pronounced difference increase the shift, and if you want less of a change decrease the shift. You can also increase or decrease the steps.

Minimal Change Example

If you set it to 17 it will only run for one step, so there is hardly any change. Running one step is relatively fast, but you get basically no change at all. After and before it is just slightly sharper, and that is it.

Increasing Steps and Tweaking Shift

Bumping the steps to 15 makes it slightly more realistic. The 3D effect is mostly gone. There is slightly a more plastic look in the highlights. You could lower the steps by one or tweak the shift.

With a shift of four it is ever so slightly more realistic and we retain most of the detail. The buckles are pretty coherent. There is a little extra detail which may be wanted or not.

If there is one concern, the more tiles you have, the more inconsistent the image could be because we are not using any control net. In this case the right side of the head is slightly more out of focus than the left side. The background changed quite a bit.

Final Thoughts on Z-Image Upscaling and Detailing

2x scaling is a safe starting point, and for higher scales match your tiling and untiled settings for restitching. Sharpen after upscaling and add just enough grain to introduce texture without making the image look noisy. Keep prompts simple and adjust steps and shift to balance realism and detail. Be mindful that more tiles can introduce inconsistencies without control net.