aetherart-ukiyo-sd21

AetherArt — SD 2.1 Ukiyo-e LoRA

A rank-8 LoRA adapter that steers Stable Diffusion 2.1 toward Japanese ukiyo-e woodblock print style. Trained on 80 WikiArt Ukiyo-e images against the sd2-community/stable-diffusion-2-1 base model. Activate the style with the trigger token ukyowood anywhere in the prompt.

Usage

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "sd2-community/stable-diffusion-2-1",
    torch_dtype=torch.float16,
).to("cuda")

pipe.load_lora_weights("gauravgandhi2411/aetherart-ukiyo-sd21")

img = pipe(
    "a ukyowood mountain landscape at sunset, traditional woodblock print",
    negative_prompt="text, watermark, calligraphy, signature, words, letters",
    num_inference_steps=30,
    guidance_scale=7.5,
).images[0]
img.save("output.png")

Training details

Checkpoint selection

Selected checkpoint-1000 over the other checkpoints trained during the run:

• checkpoint-500: underfit — style signal present but not saturated, more like a mild filter than a transformation.
• checkpoint-1000: selected — consistent warm amber palette and characteristic flatness of traditional woodblock prints across test prompts.
• checkpoint-1500: overfit — validation loss rose from 0.268 to 0.495. Outputs showed over-saturated colors and partial prompt-alignment breakdown.

Checkpoint selection was made by visual evaluation only — no quantitative held-out set was used.

Default negative prompt

The following negative prompt is applied automatically by the AetherArt application whenever this adapter is active:

text, watermark, calligraphy, signature, words, letters

Known limitations

• Calligraphy artifact (partially mitigated, not fixed): WikiArt Ukiyo-e source images contain metadata captions with artist signatures and script text embedded in the image margins. The adapter learned these as part of "ukiyo-e style." The default negative prompt suppresses most instances but does not eliminate the artifact entirely — the style signal and text signal are entangled in the adapter weights. The correct fix is retraining on a curated dataset with no text annotations, which would require approximately 5 hours of curation work.
• The adapter was trained and evaluated on 512 × 512 resolution. Results at other resolutions are untested.
• CLIP scoring does not capture the quality improvements from this adapter. See the CLIP-blindness finding linked below.

Links

• AetherArt repository: https://github.com/gaurav-gandhi-2411/AetherArt
• CLIP-blindness finding: see reports/clip_blindness.md — nine Phase 6b experiments showing CLIP delta <1 SE while LPIPS ranged 0.40–0.73; underfitting paradox; why CLIP cannot guide LoRA training decisions.
• Companion SDXL adapter (1024×1024): gauravgandhi2411/aetherart-ukiyo-sdxl — same rank-8, same dataset, trained on GCP L4. Both runs independently select checkpoint-1000.