Th4dd3us/sd14-oldbook-lora

Stable Diffusion v1.4 · Old Book Illustrations LoRA

LoRA adapter for CompVis/stable-diffusion-v1-4, fine-tuned on gigant/oldbookillustrations to generate images in the style of 19th-century book engravings: ink outlines, line hatching and a sepia/monochrome palette.

Only the UNet attention layers were adapted; the VAE and the CLIP text encoder are unchanged.

Usage

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights("Th4dd3us/sd14-oldbook-lora")

image = pipe("An old woman spinning wool by the fireplace").images[0]

No trigger word is needed: plain scene descriptions (like the dataset captions) work best. A version with the LoRA already fused into the weights is available at Th4dd3us/sd14-oldbook-fused.

Training

Setting	Value
Training images	500 (first 500 rows of the train split)
Image column / caption column	1600px / info_alt
Preprocessing	resize shorter side to 512 → center crop 512×512
LoRA rank / alpha	16 / 16
Target modules	to_q, to_k, to_v, to_out.0 (UNet)
Trainable parameters	3,188,736 (0.37% of the UNet)
Optimizer / learning rate	AdamW / 1e-4
Batch size / epochs	3 / 5 (835 steps)
Precision	fp32
Hardware	1× NVIDIA T4 (Google Colab), ~55 min

This adapter was selected from three runs (rank 8 vs 16, 500 vs 1500 images); rank 16 produced the most consistent engraving texture on the evaluation prompts.

Limitations

• Style transfer is strongest on figures and interiors; landscapes and seascapes keep more of the base model's painterly colour.
• Evaluation is qualitative (3 prompts, fixed seed), not a quantitative benchmark.

License

The adapter inherits the CreativeML OpenRAIL-M license of the base model. The training data (gigant/oldbookillustrations) is licensed CC BY-NC 4.0, so use of this adapter should be limited to non-commercial purposes.