Lakshitha-MR-sd15/mystyle-lora-sd15

mystyle LoRA (Stable Diffusion 1.5)

A small style LoRA trained on 30 images, trained and tested entirely on free Kaggle GPUs.

Trigger word: mystyle

Samples

How to use

ComfyUI: put mystyle_lora.safetensors in models/loras, add a Load LoRA node after the checkpoint loader, and start your prompt with mystyle,. A ready workflow is included: mystyle_workflow_api.json.

diffusers:

import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", dtype=torch.float16).to("cuda")
pipe.load_lora_weights("Lakshitha-MR-sd15/mystyle-lora-sd15", weight_name="mystyle_lora.safetensors")
pipe("mystyle, a cat sitting on a chair").images[0]

Training details

• Base model: Stable Diffusion 1.5
• Method: LoRA, rank 16, via the Hugging Face diffusers training script
• Data: 30 images at 512x512, auto-captioned with BLIP, trigger word added to every caption
• Steps: 1000, learning rate 1e-4, fp16, batch size 1 with 4 gradient accumulation steps
• Hardware: Kaggle free T4 GPU

Limitations

Small dataset, so results vary by prompt. Works best with prompts similar to the training subject matter.