aion21Background info:
- These images are text2img results or inpainted results (no manual fixing), so they contain the full metadata made with NovelAI (including models, prompts, UCs, samplers, steps, etc.).
- Pixiv doesn't strip metadata of the images when uploading; so if NovelAI's metadata stays intact, you can download images from pixiv then upload them to NovelAI's image generator interface and play around with the prompts right away (instead of copy-pasting). You have to click on an image to open it then download, so the file has a name like "109931627_p0.png" and will have the metadata. Otherwise, you can also use Notepad or another program to check the metadata.
Testing info: (please note that I could be wrong)
- I have tested the vector "Shirow Miwa" before on V2 (please see my previous posts). I'm still testing on V3 regarding the syntaxes; for this prompt, [[[Shirow]]] {{{{Miwa}}}} seems to work better.
- DPM++ 2M is really good for details and has a neat drawing/painting effect, but it can also lead to blemishes. I have also tested some other samplers in this collection.
- For this prompt, I followed NovelAI's guide about making a more consisten character:
https://blog.novelai.net/538b4b678a4e
https://blog.novelai.net/2bee02bad60f (the same guide but in Japanese)
In my opinion, doing this also has the benefit of improving quality for characters, since the AI won't have to guess what to choose then make weird depictions because it's confused on what we want (just as a figure of speech, because AI isn't sentient). Of course, doing this means you need to know what words to use for hair, clothes, and such. I'm not very good at this lol.
Disclaimer and resources: These pics are made with NovelAI's image gen, aka NovelAI Diffusion or NAI Diffusion.
Website: https://novelai.net/
Documentation about functions: https://docs.novelai.net/
Useful website: https://zele.st/NovelAI/
2023-11-24 10:58:38 +0000