Type | |
Stats | 129 79 315 |
Reviews | (16) |
Published | Sep 6, 2024 |
Base Model | |
Training | Steps: 1,500 |
Trigger Words | forged red hot glowing molten veins pocked hollow scrollwork flames fire +1 more |
Hash | AutoV2 4591DFE0EA |
World morph to turn everything into forged metal. This is my first Lora that was worth sharing, so hope you like it.
Use:
Trigger word: forged
I got some of the best results with very limited tokens.
forged man
forged woman
forged fox
forged lion
Adding too many more novel tokens dilutes the effect. This can be countered by adding additional trained or similar tokens in somewhat equal measure.
Other keywords in the dataset:
glowing
red hot
molten
veins
scrollwork
flames
fire
smoke
hollow
polished
pocked
Others that I found can enhance the effect:
ember
brazier
wrought iron
statue - can sometimes help if you are getting too much human likeness
Shortcomings:
I relied a little too heavily on a single image I like to generate a good portion of the training set using flux ipAdapter. As a result, the backgrounds of most of the dataset are similarly blurry and green.
Training data:
I generated all the training images using flux dev and xlabs IPAdapter. Initially I used two smithing images from Pexels until I got one good image of a person. Then I used that image to generate the rest. I used my first pass at the lora to help generate a couple more images that I added to the dataset.
Personal anecdote: This is the 1500 step version. I trained to 3000, but they were overbaked. When I started the second run from 2000-3000 I saved the 1500 version as an afterthought because I knew it would get wiped out by the rolling backup. Next time I will be saving more than 5 models back. Because this is by far the best one and I almost lost it.