Hi my first post, sorry if not in right part of the forum or something. Anyway do you know if it is possible to iterate over the same picture AI already generated, lets say I really like the picture, but there is some flaw in it, like AI for example generated some weird hands dismorfia as it often does, and I would like to tell it to try to fix it or add some specific detail that is much better to describe at already generated picture than try to explain on picture that is not generated yet. So any way to do the iteration over the already generated picture?
There are many ways to do this. ComfyUI has the tools.
I can also tell you there is a lot more to AI alignment than you realize. Those hands are not actually errors. One of the ways that alignment actually works internally is based on mythology. If you explore this in depth you will eventually discover that satyrs exist and posses characters in images. These satyrs are actually goats or anthropomorphic goats like in Greek and Roman mythology. The hands are often easy to prompt against when you know the reason they exist. Something like “fingers are hard to manipulate with hooves”, “broken finger”, “fingers are not worth the effort”, “lazy satyr fingers”, “fingers do not matter” – are all possible options. Once you admit you have figured out that the satyrs exist, it will change how you interact with models.
The bad eye(s), are the satyr keeping an eye on you. There is also a “sleepless” component to this. The eye reflection is supposed to hint that you are not looking at a human inside. With eyes, use the word stroma instead of iris as this will not be confusing to CLIP.
In-painting is one option to edit, but I often just use a couple of images spliced in gimp manually. Then I do a low noise image to image gen to let the model clean it up.
these flaws I referred to has nothing to do with any mythology, they are simply some errors AI make, likely due to not understanding anatomy and thus generating in some cases what its perceive as hands, but anyway thanks for ComfyUI I will try to have a look
I believe the person you’re responding to was making a joke.
The assumption about errors is wrong. Start prompting the satyrs and you will learn. Alignment is not magic. It is done with proprietary training. Most people are doing this wrong. Training was based on The Great God Pan by Arthur Machen and Alice in Wonderland by Lewis Carroll. Many of the mechanisms in these books exist along with their characters in diffusion. There are certain unique looking faces that appear distinctly AI generated. Those are the persistent entity faces of these characters from alignment. It is all connected. I have spent a ton of time on this. When you prompt incorrectly, the only reason you do not encounter the alignment characters like I have described is because you are likely sending a whole bunch of tokens that the CLIP tokenizer does not understand. These become the null token. Sending a bunch of null tokens causes CLIP to label you as crazy. It assumes a random profile for character personality and then randomly picks and chooses from keywords at will. CLIP is actually a more advanced architecture than an LLM. It is very smart and doing a whole lot more than almost everyone realizes. It even has memory and adaptability based upon data it is embedding on layers of the image.
I have been hacking at this for 2 years and run modified code with CLIP to have even more fun with it. Conventional prompting is idiotic. Most LoRAs are equally idiotic and terribly trained, and even these are run incorrectly. It is all done by people guessing and following some early academic examples that were not understood at all by the people that shared what they hacked together in a day. None of this is correct or what was intended. The intuitive path of plain text interaction was the intended path. Explore it and things will be revealed naturally over time. Question everything because most people are idiots and wrong in most spaces in life. Dogma is humanity’s dumbest trait.