Type | |
Stats | 319 13,637 |
Reviews | (163) |
Published | Jul 16, 2024 |
Base Model | |
Training | Steps: 600 Epochs: 10 |
Usage Tips | Clip Skip: 1 Strength: 0.8 |
Trigger Words | made of, tiny neon glass filaments, w0gf1l |
Hash | AutoV2 D10A5FCA1F |
A world of filaments, tiny neon filaments, tiny neon glass filaments! Now you're asking yourself, can't we just prompt that? And you would be correct but it usually ends up sloppy and uninteresting.
I attempted to put a little origination to the chaos. I think I managed it. I wouldn't call it perfect tho, it was a [expensive] struggle bus to get it to this point, with 6 total version being created with multiple training styles.
The two version I will upload are the best of those.
Now the technical stuff:
Use the triggers! The one's between the || triggerword || are optional however. Try to keep your prompts somewhat short putting the subject you want changed very close to the beginning. I found formatting your prompt like: "X made of, tiny neon || glass || filaments" (glass being optional) works the best. The rest of your prompt can be whatever but again, try to keep it short.
It does work with other loras but generally you want them to be less weighted than this one.
It can have issues with producing background on it own. If you provide the prompt and give it enough weight, it should make it. It also isn't what I would call great at making realistic people, however it does seem to do well with anime/cartoon styled people.
Doing a hi-res fix with a denoise around 0.2 - 0.26 is suggested and if you want the ultimate detail out of the filaments, upscale using 4x_NMKD-Siax_200k. I felt like DPM++ 3M SDE Karras sampler gave best results. Higher the steps usually the better. CFG, depends on the model using.
Note: Onsite generator users, I have no clue how well this will work for you. No promises!
And pony folks, if this gets used enough I will attempt to make you a version but this is it for now.