Thirty 8ght Style (MOSS-Image)

作者：Alt4Pai

Model trained on base MOSS-Image (int8) to imitate the style of Thirty_8ght ===== Support the artist: https://www.pixiv.net/en/users/30570055 ===== Done by request, but also because I was interested in trying out MOSS-Image training at some point anyway. While it took some setup (and I eventually fell back to using Next-Trainer), I was surprised at how the results came out on what was supposed to be a test LORA using a dataset I already had. It is trained on the usual booru tags, no natural language, but (as you can see in some of the sample images) natural language still works just fine with it. Despite it being the first run and a test, I decided to upload it since I was happy enough with the results. These sample images were mostly me messing around with the different ways MOSS-Image can understand prompts (just tags, natural language, a combo of the two, and the JSON style prompting). For future LORAs, I'll likely return to how I did tagging for the sample images of my Anima LORAs (simple, but with natural language sprinkled in where I felt it helps). NOTE: All the images I generate for the showcase are done with base MOSS-Image (int8_convrot) without the use of any quality tags, negative prompt, Adetailer, Inpainting, etc. only Hires. fix to do very basic upscaling. You can get nicer looking results using MOSS-Image SFT, other finetunes, or just more precise generating methods in general. These outputs on base are to show that the lora works. Some personal musings on MOSS-Image: While I've not delved too deeply into it yet, I feel like MOSS has a ton of potential. It's prompt adherence is like nothing I've seen so far, and while it doesn't natively support a lot of anime/furry/character concepts, I've seen some of those things still manage to work with specific enough prompting (like the M. Bison sample image for this LORA. It had no idea who M. Bison was, but some more specific prompting got it closer and I could have gone even further with that). The bounding boxes work very well, and it has insane text capabilities compared to Anima. I'll probably still train for Anima if I train any new styles, but expect me to bring said styles to MOSS as well, if this works as well as I feel like it did. I am also left with some interesting questions from this test run that I hope to answer as I continue training on this model:

• Anima and NOOB worked just fine without trigger words on the style loras I made, but would MOSS benefit from adding the trigger words? The more structured style of the prompting would lead to me think having a dedicated trigger word would definitely be a helpful thing in this scenario, especially if trying to do something like use specific styles within specific bounds. But would MOSS even be capabable of dividing up styles from those LORAs in such a way reliably? Should I just keep doing what I have been doing?
• Does the fact that I trained on int8 reduce the overall influence or quality of the LORA? Testing shows me I have no reason to believe so, but I'm still feeling a little paranoid about it. Training on the full version is too costly for my GPU to handle either way, however, so int8 it shall be.
• ... in the first place, how should I be using a LORA with the JSON formatting at all? I've just kinda stuck it at the end outside of all the JSON and that seems to work, but I can't help but feel there is a more structured way I should be doing that.
• While the booru tags WORKED, is there a better tagging system I could use for MOSS? I'd like to avoid explicit AI agents, but tagging in the style MOSS recommends in their guide is certainly not something I want to do by hand...
• I made the steps in this new trainer roughly match the amount of steps I used for Anima. While it looks like it worked, would more steps help further adhere to the style? Would less steps with other tweaked training parameters help the style come through better? All of these are interesting and have reinvigorated my interest in training LORAs. I may be a little more active in the coming weeks as I try this thing out.