If you are like me, you would be scrolling through places like CivitAi in the hopes of finding a new model to try out or an image that catches your eye and want to try re-creating it.
While you often find the prompt in CivitAi or via something like PNGinfo, copying that prompt and trying to re-create that will not yield the image or even close to it. Why would that be and how can we solve it?
- The first possibility is that the image was not created using pure Text prompt, but using Image to image.
If you ran this attached image in PNG info, you will get below
While you can see that the prompt was “high reslution, shiba inu”, because I created this, I know that this was not created by text to image, but instead, it was created via image to image. There is nothing obvious in this metadata that shows that it was created via image to image, except of the reference to denoising strength. In such case, it is possible to suspect that it was created via I2I.
As a reference, this border collie is the starting point of the Shiba inu. - When the layout/setup/pose seems too unique or precise, it’s likely using ControlNet
When viewing some of the image samples in CivitAi, you sometimes see very similar images next to each other. In many cases, the poses are uncannily similar to the previous image. In such case, I make an assumption that they have locked the pose in ControlNet (either using depth, canny, open pose or line art) and created the artwork from that. In particular, SD or AI has trouble drawing things that has depth perception, such as something being in front of something else or selfie like angle.
If you are unsure about how to use ControlNet, there are several very good videos in YouTube.
Update: Since writing this, I’ve started alltools.solutions, where I write more generally about practical, free AI tools. If you’re curious where something like Stable Diffusion actually fits versus everything else people lump under “AI” these days, I’ve got a post on the different types of AI that might help make sense of it.
Hopefully, these will give you some guidance in your journey in the AI art.