Hi... I'm still really frustrated and super upset. I haven't been able to get the results I want with style LoRAs and I honestly don't know what I'm doing wrong—maybe my dataset? IDK. This one at least feels like a step up, but it still doesn't feel quite right. Any help would be appreciated! Here are my training values:
Anima v1.1
Training Settings — Megrim Style LoRA (epoch 15)
Base model: circlestone-labs/Anima (Anima Base v1.0) — 448 layers
Type: Style LoRA · Trigger word: @megrim
Optimizer: AdamW8bit
Learning rate: 1e-4
Text encoder LR: 0 (text encoder not trained)
Weight decay: 0.1
Betas: (0.9, 0.99)
LR scheduler: cosine_with_restarts (3 cycles)
LR warmup: 100 steps
Network dim: 32
Network alpha: 32
Network module: networks.lora_anima
Train UNet only: Yes
Batch size: 1
Gradient accumulation steps: 1
Resolution: 1024, 1024 (buckets 512–1536)
Mixed precision: bf16
Gradient checkpointing: Yes
Loss type: l2
Timestep sampling: logit_normal · sigmoid_scale 1.0 · shift 1.0
Seed: 42
Epochs (total run): 18
Published epoch: 15 (best in testing)
Steps at epoch 15: 4,875
Total steps (run): 5,850
Training images: 325
Save every: 3 epochs
Dataset: 325 images + 325 captions, WD14 tags, every caption starts with trigger @megrim plus style tags flat color, sketch, simple shading (recommended for style LoRAs). shuffle_caption=true, keep_tokens=2, no caption dropout.
Trainer: gazingstars123/Anima-Standalone-Trainer (Kohya sd-scripts fork, commit 79c30b3)
Anima v1.0
New version for ANIMA, as always, don't forget to check out the artist's social media.
Hi, it's been a while since I've posted anything. I wanted to try my luck with the Anima model, but honestly, I'm not completely satisfied with the result. For an image closer to the original artist's style, don't include the quality tags—it works fine without them. I'm not sure if what doesn't quite work for me is due to the checkpoints I'm using or the LORA training. As always, I appreciate any suggestions or comments.
Hello once again! This is my new attempt at creating this LoRA with Illustrious.
This LoRA is designed to capture the art style of Megrim. While it's still in development and there's room for refinement, it's showing promising results. The trigger word is "megrim," so feel free to experiment!
Please remember to visit the original artist's websites and support their work. This project is purely for amateur purposes and not intended for commercial use.
Description
Hello again! This is my second attempt at creating this LoRA, representing the second epoch from a larger dataset. The fifth epoch seems a bit overtrained, but this one strikes a good balance. Give it a try!
This LoRA is designed to capture the art style of Megrim. While it's still in development and there's room for refinement, it's showing promising results. The trigger word is "Megrim," so feel free to experiment!
Please remember to visit the original artist's websites and support their work. This project is purely for amateur purposes and not intended for commercial use.