Fix for a Broken GPT Image 2

Contents Introduction The Original Image How do I deal with this? Why is this happening? How can I tell if it's really happening? How do I make it stop? The Repaired Images (from the least change to the most change) The Repaired Image: GPT 2 - High The Repaired Image: GPT 2 - Low The Repaired Image: GPT 2 - Medium Final Notes Introduction Yes, sadly GPT Image 2 is broken. It leaves a bunch of yellow dots all over the image, and this is especially true when the images are very busy. I had no idea, and I didn't even really recognize that there was a problem. I was pleased with myself when I was creating some of the imagery for Rad & Brunhilde. I liked the look of them while they were in their battle scene with the Orcs. Those images were certainly quite busy. I got a comment on them from @thewhiskeyjack , pointing out the problem. twj-"When the image is too busy and not photorealistic, it has these tiny white and yellow dots (noise)." The Original Image I picked this image to study because it is a great example of the problem. It is very busy as there are a lot of objects in it (in the form of orcs, siege towers, etc.) Also, because this is one of the six images in the post that was commented on. The dots are pretty easy to see. Just look closely at anyone's skin. They're easiest to see there. Once you know what to look for, you'll see them literally everywhere in the image. Clothing, weapons, Orc bodies, etc. Once I saw them, I can't unsee them, and I see them on images from others as well. How do I deal with this? The way I see it, there are two problems. 1 - How do I stop making images with these dots in them? 2 -Can I fix these dots on drawings that already exist? To get both answers, I had a chat with "Chad" (ChatGPT) Why is this happening? Chad: Yeah—that is a thing. What you’re seeing is basically artifact noise from high-detail scenes. Those yellow/white dots are usually: tiny highlight artifacts overinterpreted sparks / dust / embers or just detail noise where the model is cramming too much into one image In busy scenes (battle, debris, particles), 2.0 tends to fill every gap with something visually interesting. Sometimes that turns into random speckling. You didn’t notice it because once you focus on the subject, your brain filters it out. But if someone looks at the background , it pops. It's caused by high detail prompts, lots of small elements (dust, motion, lighting effects), and/or strong contrast lighting. The model “hallucinates” micro highlights everywhere How can I tell if it's really happening? Zoom into one of your images and look at dark areas, out-of-focus background, and edges of motion blur. If you see random bright pixels with no structure, that’s the issue. How do I make it stop? Chad: Tell it to stop inventing micro-detail. Right now the model thinks “more detail = better.” You need to cap that. Add something like: "clean background, no speckled noise, no random particles, no visual artifacts" or tighter: "smooth gradients, no grain, no dust, no floating particles unless intentional" This works because those dots often come from dust, embers, or atmospheric particles, even if you never asked for them. This tells the model not to fill empty space with junk. If you do want atmosphere, control it: "subtle, soft atmospheric haze (not particle-based)" That keeps depth without the speckling. DOB: On a side note, these prompts add-ons work really well. I use the first prompt, but without "clean background" . My prompt is: "no speckled noise, random particles, or visual artifacts" . So that answers the first issue. The second one is now pretty easy. Put your bespeckled image into img2img (I put it into GPT 2), and ask for those things to be removed: "Clean up all speckled noise, random particles, and visual artifacts." . The Repaired Images I fixed my image with that prompt. I tried it with all three model qualities, and here are the results. These fixed images are presented from the least amount of change to the most amount of change. The Repaired Image: GPT 2 - High To be fair, this was the last one I tried. I tested Low first. It turns out that High does not do a great job of repairing the issue. You can tell it is better, but not by a lot. High does a much better job when it creates an image than when it tries to repair one. The Repaired Image: GPT 2 - Low You can definitely see the improvement in this one, even over High. Yes, there are significantly fewer dots, but there are still some left. The Repaired Image: GPT 2 - Medium This one is practically perfect. Just take a look at Red's thigh. Take a look at anyone's skin. There's nothing left. The image looks much more crisp and clear. Final Notes This might seem obvious, but it's worth saying all the same. I find that repairing an image does not work as well as making the image without the noise right from the start. I hope you enjoyed this blog, and that you found it useful.