Mervecaliskan/japandi-interior-lora

Japandi Interior LoRA (SD 1.5)

A small style LoRA that teaches Stable Diffusion 1.5 the Japandi interior style: light wood, beige and linen textures, low furniture, paper lamps and shoji panels. Trigger word: jpdi.

Same prompt and seed. Left: base model, middle: strength 0.5, right: strength 1.0.

Usage

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
pipe.load_lora_weights("Mervecaliskan/japandi-interior-lora")
image = pipe("a photo of a jpdi style bedroom",
             cross_attention_kwargs={"scale": 1.0}).images[0]

Training

• 17 images (7 living rooms, 5 bedrooms, 2 dining rooms, 2 kitchens, 1 detail), captions with the room type
• LoRA rank 8 on UNet attention layers (to_q, to_k, to_v, to_out), 1.6M trainable parameters (0.19% of the UNet)
• 800 steps, learning rate 1e-4, 512x512, fp16 mixed precision
• 4.9 minutes on a free T4 GPU, 2.1 GB peak memory, 6.4 MB file

Evaluation (CLIP, 5 room types, one seed each)

LoRA strength	Style similarity	Prompt match	Avg. max similarity to a training photo
0.0	0.775	0.274	0.775
0.5	0.854	0.273	0.858
1.0	0.881	0.285	0.887

• The style transfers to room types not in the training set (home office, partly bathroom).
• No sign of copying: the highest similarity of any output to a single training photo is 0.92 (below 0.95).
• Weak spot: kitchens change little at strength 0.5 because the set had only 2 kitchens. Use strength 1.0 for under-represented rooms.

Limits

Small dataset, partly AI-generated renders. Results are a style experiment, not a production model. Training images are not shared.

Code: https://github.com/Mervecaliskann