masafy/masafee-lora

Masafee LoRA

Stable Diffusion LoRA for the original character Masafee — a red panda from a personal LINE sticker set. Trained entirely on a single consumer GPU (NVIDIA GeForce RTX 3060, 12GB) with kohya-ss/sd-scripts.

Model details

Base model	Counterfeit-V3.0
Method	LoRA (rank 32)
Learning rate	1e-4
Epochs	30 (sweet spot before overfitting)
Training images	13 hand-picked LINE stickers
Trigger word	masafee
Recommended weight	0.8
Hardware	RTX 3060 12GB (≈9 min / 16 epochs)

Usage

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "gsdf/Counterfeit-V3.0",
    torch_dtype=torch.float16,
).to("cuda")
pipe.load_lora_weights(
    "masafy/masafee-lora",
    weight_name="cf60-000030.safetensors",
)

prompt = "masafee, red panda character, sitting on a park bench, sunny day, masterpiece"
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
image.save("out.png")

Apply at network_mul=0.8 for the most faithful character likeness without rigidity.

Paper & full reproduction

The full experiment writeup (Japanese / English, two-column PDF), all training scripts, the 6-condition sweep, the epoch ablation, and side-by-side comparison grids are in the companion GitHub repo:

🔗 https://github.com/masafykun/masafee-lora

Key findings from the paper

1. Base model choice dominates — switching the base model (SD1.5 → Counterfeit-V3.0) had a far larger effect on output quality than tuning LR or LoRA rank.
2. Epoch ~30 is the sweet spot — extending to 40–60 epochs produced visible overfitting without improving character fidelity.
3. Output quality is bounded by training image count — with only 13 stickers, certain poses and lighting conditions remain out of distribution.

Citation

If you use this model in academic work, please cite it via the Zenodo archive:

@software{suzuki_masafee_lora_2026,
  author       = {Suzuki, Masato},
  title        = {{Masafee LoRA: Single-GPU LoRA Fine-tuning of an
                   Original Character on Stable Diffusion}},
  year         = {2026},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.20397946},
  url          = {https://doi.org/10.5281/zenodo.20397946},
  orcid        = {0009-0000-7977-2756}
}

Persistent archive (Zenodo DOI): 10.5281/zenodo.20397946

License

• LoRA weights (this file): released for personal and research use only. The weights encode a specific copyrighted character.
• Code, scripts, and paper text (in the GitHub repo): MIT.
• Character "Masafee" itself, the original sticker images, and any images generated using this LoRA: © 2026 masafykun / Masato Suzuki. Redistribution, commercial use, and use to imply official endorsement are not permitted.

By downloading this model you agree to use it only for personal experimentation or academic research.

---

Made with 🐾 by masafykun · masafy.org · ORCID