Newest version: 80 Epochs, 600 steps, 1024px, .0004 LR, 8 Network Dim, 16 Network alpha.
Good luck with my newest version! I hope you enjoy it.
Current title card images generated with 20 steps DEIS + ddim_uniform.
I use DEIS + DDIM Uniform. I am generating at a base resolution then upscaling once with Ultimate SD Upscaler by 1.6, using facedetailer, then another Ultimate SD Upscaler by 1.25. The upscaler model is RealEsrgan_x4plus_anime6B.
An 80 epoch lora of 60 captioned Nikke (in-game asset) character images.
Trained with the following settings:
| Engine | kohya | | Epochs | 80 | | Num Repeats | 1 | | Train Batch Size | 4 | | Steps | 600 |
| Resolution | 1024 | | LoRA Type | lora | | Enable Bucket | Enabled | | Shuffle Tags | Enabled |
| Keep Tokens | 1 | | Clip Skip | 1 | | Flip Augmentation | Enabled | | Unet LR | 0.00040 |
| Text Encoder LR | 0.00010 | | LR Scheduler | cosine_with_restarts | | LR Scheduler Cycles| 3 |
| Min SNR Gamma | 5 | | Network Dim | 8 | | Network Alpha | 16 | | Noise Offset | 0.10 |
| Optimizer | AdamW8Bit | | Optimizer Args | weight_decay=0.01,<br>eps=1e-8,<br>betas=(0.9, 0.999) |
Description
NEWEST Version, 80 epochs! This re-train retained the distinct look of the Nikke characters and Shift-Up studios/level-infinite art design even more effectively than the last version.
Please enjoy and share your work here to help spread this LoRA out and grow it's usage.
FAQ
Comments (3)
any list of Nikke characters traines? triggers word?
all characters and some costume names and sub-characters up until last summer. Just call them by name. or their modified unit name in the case of their outfits.
The triggering method is too vague, and several pictures are generated, which are basically irrelevant to the specified character. It is recommended that the producer provide a template for keywords, otherwise using this lora can only highlight the 2.5D drawing style and cannot generate effective character pictures.