I have recently started using GPT 1.5 image to image generation on BudgetPixel for various tasks. In this short blog I will share a few way I have used this great tool. Everyone who knows me, knows GPT 1.5 is my favorite model. However, it's not always the best at every task or subject. I generated a fairy prompt aiming for a bit more realism than my usual style. I tried low and high, loving neither so I ran the same prompt in MidJourney V7. Results below: Low (left) looks pretty good, but I'm actually starting to get tired of the GPT girl look. High (middle) loses a lot of the ethereal misty quality. MidJourney (right) was pretty close to what I wanted but not quite there. At this point, I decided why not give GPT 1.5 the MidJourney image as reference. In this instance, I also included the full original prompt. The result below gave me the perfect balance I was looking for: ethereal beauty without the classic GPT 1.5 girl. Next up, I created an image for a Black and White challenge (shout out to @justrob since I used one of his style blocks). I then decided I wanted this image in color to use as a start image for video creation. I didn't want to rerun as it took a couple of tries to get a composition I liked. So instead, I turned to GPT 1.5 image to image with the following prompt: Make this image full color. Bold saturation. Color palette that coordinates with a natural color tiger. Ensure background does not overpower subjects. Change nothing else. I was amazed at how well this simple prompt worked. The resulting image: I did animate this and will share to the BudgetPixel feed when I share this blog post. Finally, using GPT with two reference images. In creating my next image, I wanted to elevate the image on left by changing the skyline to the buildings from my image on the right. I used the following prompt with both images as references: Use the woman from reference one but change the background buildings to match the ones from reference two. Hair a brighter almost glowing voluminous neon green. Eye catching. Change nothing else. Result below: Hopefully, this gave you some ideas on using GPT 1.5 with references for your projects. Please clap this blog and leave comments below!
Comments (17)
faerierealm
Yes, GPT 1.5 is incredible, and our primary tool for Img2Img work. But it will balk at nudity, so SD4.5 is a great alternative for those tasks that GPT 1.5 refuses to do. Your B&W image to color is very impressive; we've never done anything like that. Great article!
EternaSky
Oh i make character sheets to use with gpt img2img for my rift protocol creations. I spent thousands of creds on my blog already as not every generation is ideal. Its very useful to use ref images for various models.
Sealine
Thank you! Yes, GPT is anti-nudity and even prudish with some sexier references. I have used SD4.5 for some tasks in i2i but never generate in it because I have an unhealthy dislike for the seadream girl 😅
faerierealm
Haha in that case, stick with GPT! But the two actually mix pretty well. And by that, I mean SD4.5 can modify GPT (in i2i mode) while preserving style, and vice versa.
Sealine
I have on occasion used it to remove tops from the woman in a few images. I think that was the model that gave the most realistic nipples.
faerierealm
Yes, in our experience, GPT 1.5 and SD 4.5 are by far the two best models for i2i work. And only the latter can handle nipples haha.
panos
Img-to-img is a fantastic tool. Great blog, particularly as a how-to and also to get people who don't about it to learn all concerning it
Sealine
Thank you! I'm still learning the capabilities of the tool and was surprised by things it could do with ease!
DavidP
Great blog, sure helpful to people who've not discovered the joys of img2img.
Bobbie
That's a great idea. I will be upping my game now, lol. 😂