introduction
If you’ve been following my blog, you’ll notice I have been playing with dalle 3.
Chatgpt got a update that rolled dalle 3 with the other beta plugins so I decided to play with it further
I really liked this from the last time I had it generate images I decided to play with it a bit more, using this as a starter.
and a prompt of
Uploading a image with a prompt seems to be more effective at setting the pose/art style. It took 2 retries, but this was a good start
I replied with
cool can you maintain the camera and jacket of the original?
I had it regenerate 4 times before I got to this one which was closer to what I was looking for. The lens were still a bit too shaded for my liking and weren’t quite the right shape. I decided to try
make the lens a bit more transparent like the uploaded image
2 regens got to this. The clothing and sunglasses were closer to what I was looking for. Now I wanted to get the pose more similar to the starting image.
I replied to the previous with the following
cool now make the character’s pose where the camera sees their shoulder similar to the original image
with 3 regens it was closer, but the glasses and cat ears were still off.
I fed the prompt
make it more towards the original, keep the cat ears
The output was closer to the pose I’m looking for, but the arm and hands didn’t make sense for where the glasses were positioned
Thus, I replied with
mirror it more
nice, this is the pose and camera angle I’m looking for, the hand is still a little off
I took four more retries with this prompt
make the arms anatomical, maintain the pose and angle
before I got this
the hand is much better, but I wasn’t quite satisfied with the angle or the hair style
so I replied with
cool can you make the hair style more similar to the original?
It was moving away from the pose, but the clothing was what I was looking for, I still wanted the pose and framing of the original so I retried 2 more times, with this prompt
cool can you have the pose and camera angle more like the original?
This is a contender if I stopped here, but I was really motivated to get the pose and framing, so I tried
can you make the hair and pose the same as the uploaded image? keep the cat ears
this is definitely close in the hairstyle and glasses, but the position of arm/hand and clothing was a little off
I tried some more attempts, but it didn’t quite yield what I was looking for. I was getting tired at this point so I decided to upload the original and include a prompt of
can adjust the output to be more similar to this?
Honestly I could stop here or the one before the previous, I still wasn’t quite happy with the pose so I tried another with the following prompt
make the pose more similar to the original
3 retries, but the other hand was more interesting.
conclusion
being persistent and uploading reference images seems to steer the model better than using natural language.