Media Arts and Technology

10.15 XY Plot

An XY plot in ComfyUI is a grid of generated images that lets you compare how changing two settings affects the results.
The assignment includes the following:

1) Import any of the above images into ComfyUI to get the workflow. Test it with different parameters, different prompts and different sampler settings and checkpoints. Become familiar with the interface. When testing, aim for innovation in results.

2) Shaw has prepared a number of workflows. They are located in the left menu in "Templates". Select XY Plot comparisons. Test them with your own text prompts, image prompts and paramaters.

3) If you feel up to it, ask Chatgpt to write a diffrent XYplot that compares something else like the scheduler or chckpoints or sampler name, or denoise, or ....Send me the workflow in case it works, and ask Shaw in case it does not work how to make it work. Also let me know which settings you will want to compare so that we can have a range of different comparisons in the class.
Post and describe your results.

--------------
Today (Oct 12, 2026) I added three more examples:
In this version, you can choose any 3 random seed as you wish. I used 5, 9345, and 20. Doing XY plot is like a "throwing a wide net", a fishing metaphor. You dont know what you are going to get but once you get something of interest you can then go back and use related values.

In this series I am comparing denoise 0.7 to 0.9. 0.7 is stays quite close to the original image, in 0.9 the text description dominates, so I will also try denoise 0.8. The processing takes a very lng time for this workflow (300 sec) and it produces very large files which I cannot import into the student forum so I did some screen shots (They dont have the workflow in them !!)
denoise 0.8:

For this assignment I kept going with my question from last week, whether a machine can make a mistake on purpose. Last time I mostly changed the prompt, but the XY plot let me change one setting at a time and compare the results side by side. That made it easier to see exactly where the "mistakes" come from: the prompt, the settings, or the input image.
Seed × CFG grid
Prompt: interestingness shape rock, celestial
Settings: sd_xl_base_1.0, seeds 42–45, CFG 3/6/9, steps 20, euler, normal

I started with the classes' seed × CFG plot to get used to how it works. The seed decided everything about the image. Each column kept the same rock and the same camera angle at every CFG value, so seed 42 was always a flat stone sitting in a rock bed, seed 43 was always a tall balanced rock in a desert, and seed 45 was always a single boulder floating over a gray floor. CFG mostly changed how dramatic it looked and how literally it took the word "celestial." At 3 the lighting was flat, the backgrounds were plain, and "celestial" barely showed up. By 6 and 9 the contrast went up, the shadows got darker, and "celestial" turned into the more obvious celestial rock with starry skies and the Milky Way behind every rock.
Seed × CFG grid
Prompt: interestingness geometric shape, celestial shape, photograph scene with picture frame look
Settings: sd_xl_base_1.0, consecutive seeds 761575561803253–56, CFG 3/6/9

Next I changed the prompt to see if the same pattern held, and it did. Each seed was its own design: a folded paper star, a paper mountain range inside an arch, a triangle inside a circle, and a framed poster on a shelf. At CFG 3 the designs were looser and more experimental, especially the star in the first column. By CFG 9 the same designs had become cleaner and more symmetrical, and "celestial" showed up the same way as with the rocks, as starry night skies and planets inside the frames. I also used consecutive seed numbers on purpose, and they still gave completely different images, so seeds that are "close" as numbers aren't close visually at all.
Steps × Sampler grid
Prompt: a pencil drawing of a house with mistakes, shaky lines, crooked walls, a wall scribbled out and redrawn, eraser smudges, drawn quickly by hand
Negative: blur, text, watermark, polished, digital art
Settings: sd_xl_base_1.0, seed 42, CFG 4.5, scheduler normal, steps 4/16/28/40, samplers euler/dpmpp_2m/heun

For my own plot I had Claude write one comparing steps against sampler, with the seed fixed so only those two things change. I used the house prompt from last week, since it asks for mistakes. At 4 steps the samplers looked nothing alike: euler gave a soft, smudged sketch, dpmpp_2m gave the same house covered in colored noise, and heun broke down into green and yellow speckles that barely look like a house. From 16 steps on, all three settled on almost the same detailed drawing. By 40 steps dpmpp_2m and heun are almost identical.

Even though I asked for shaky lines and scribbles, none of them showed up once the image was finished. Every extra step made the drawing more careful, not less. The only "mistakes" in the grid are the 4-step images, and those come from the machine running out of steps, not from my prompt.
Steps × Sampler grid
Prompt: interestingness rock, celestial shape
Settings: same as the house grid

I ran the same steps × sampler plot with a rock prompt to check that the house wasn't a one-off. It showed the same thing. At 4 steps euler was a blur, dpmpp_2m was grainy, and heun was random color. By 16 steps they all agreed on the same flat stone surrounded by rocks. It's almost the same composition as seed 42 in the first grid, even though the prompt is worded a little differently, which shows how much the seed decides the layout. There was one small difference: dpmpp_2m and heun carved ripple rings into the middle of the rock that euler never did.

After that I wanted to see what happens when the model starts from an image instead of from noise, so I tried img2img with a seed × scheduler plot on three different input images.
Input 1 (orange circuit image)

The first input was this orange abstract image of glowing circuit-like rings, with one bright flare on the left.
Seed × Scheduler grid from Input 1

The model kept the warm color and the bright spot, but everything else changed. The rings turned into deep red shapes, and in the first and third columns they became eye sockets and part of a face. There was no face in the original, so the model found one in an abstract pattern because a face is something familiar. The three scheduler rows look almost exactly the same.
Input 2 (magenta lens image)

The second input was the magenta lens image from my first assignment, which already looked strange and hard to read.
Seed × Scheduler grid from Input 2 (magenta)

The magenta color and the dark background were kept, but the lens became smooth blobs. In the third column eyes showed up again, this time in a face made of a fur-like texture. Again, the scheduler rows are almost identical, and only the seed changed what came out.
Input 3 (landscape painting)

The third input was a muted landscape painting of gray cliffs, orange trees and a small waterfall.
Seed × Scheduler grid from Input 3

This one changed the most. The painting came back as three bright postcard landscapes: a blue lake with clouds, a forest, and a sunny meadow with a stream and flowers. Almost nothing of the original was kept except that they are all landscapes. The only place the scheduler made a visible difference was the exponential row in the first column, where the lake and its islands changed shape a little.

Reflection
Overall, the seed did a lot to decide what the image was, and most other settings only decide how polished it looks. Schedulers barely mattered in any of my tests, and samplers only mattered when the image was cut short at a low step count. Higher CFG didn't make the image more accurate, just more similar to things that are known, and it turned words like "celestial" into their most standard version like a starry sky.

The most interesting images in the set were the unfinished 4-step ones, where every sampler took its own approach. That fits the interestingness prompt. They have some structure you can tell but can't fully figure out, while the 40-step versions are so resolved there's nothing left to discover. The img2img results were interesting for a different reason, because they show what the model expects to see. Abstract shapes turned into faces, and the painting turned into a postcard.

For my research question, few of the settings got the model to make the human mistakes I asked for. The house got cleaner with every step, and img2img turned anything unusual into something more familiar. The only imperfections came from the model itself, not from the prompt. Given enough steps, the model was able to fix these smaller errors.