- Jul 21, 2017
- 167
- 176
I use AI reguarly, it's not that hard to make realistic images these days especially with NovelAI. It requires no training now on their models, just good prompt engineering. I have to wait a couple more days until I can show you that and I say this as an artist myself (pixel art/vector art/etc), the results can be pretty clean and require minimal to no clean up if done well. The main investment is time to refine the prompts and then finding good results. Self-hosted models are easier to control, however, using the first main result as the img2img guide to ensure facial similarity helps a lot.I haven't messed around with NovelAI at any point, but the way that AI image generation works means that any public service can't really be suitable for creating realistic characters. The only ways to get consistent facial features is to train the model on a specific person, or to use an individual already known to the model. (which most services disallow due to legal concerns) I know novelAI allowed training for its text generation, but even if that is also true for its image generation I doubt that training unique models for each character would be a worthwhile use of Myscra's time.
I'm not trying to discourage you from creating some examples. I just feel like a major drawback of the suggestion to try using generated images is that it will actively detract from the overall quality of the game if there isn't a pretty high level of consistency, and consistency take a huge amount of effort compared to getting one or two good looking images.