Having been a bit disappointed by ControlNet, but also knowing I can feed it poses via images, I decided to plough ahead with creating a new piece in Blender to some semi-finished state and then transferring it over. I reused by rather scrappy human model and spent quite a while posing a set:
Next I need to start adding some environment - probably a sort of gangway. I also need to make it clearer that the chap with the idea is not a religious moment, and to change each of the blokes so it's clearer they are not all the same person in some animation sequence, but rather a shuffling line...
I was re-reading some of the material on ControlNet, and it occurred to me that I can actually trying using the image in lots of different ways - pose, but also depth, edge detection etc to see which gives me interesting style transfer while keeping close to the image above... I can also:
- Try adjusting the rendering to spit out some controlnet-friendly images directly e.g. no need to extract a depth map when I can generate an accurate one directly
- Try cutting the image up by removing elements that clash and then recombing them later
No comments:
Post a Comment