Nanite-Labs/nanites-thangka-sd-1.5

nanites-thangka-1.5 (alpha)

[image]

LoRA adapter for Stable Diffusion 1.5, fine-tuned on 46 Tibetan thangka-style reference images curated from public-domain museum and Wikimedia Commons sources.

This is an alpha release: the style transfer (red field, side banners, multi-figure composition, mineral-pigment palette) is working, but the central figure is still soft relative to the reference set. Known limitation of the small (SD 1.5) base model + modest dataset size; subsequent training runs and larger datasets are the planned path to v0.2.

Trigger token

Use the phrase thangka painting in your prompt to invoke the style.

Example prompt

thangka painting, wrathful deity, red background, skull crown, gold leaf, mandorla, frontal, intricate detail

Recommended negative

blurry, lowres, text, watermark, signature, frame, cartoon, 3d, photo, photograph, modern```

## Training

basemodel: stable-diffusion-v1-5/stable-diffusion-v1-5 steps: 1200 lorarank: 16 loraalpha: 16 resolution: 512 lr: 0.0001 datasetsize: 46 lossavglast50: 0.179 elapsedsec: 2382.7218890190125 (40 min) hardware: RTX 4070 8GB, CPU offload

## License

The LoRA adapter inherits the underlying **Stable Diffusion 1.5 License** (CreativeML Open RAIL-M): the adapter is a derivative of SD 1.5's UNet weights, so downstream use is subject to the same Open RAIL-M terms. Reference images retain their original licenses (museum / Wikimedia Commons, typically CC-BY-SA 4.0); provenance is captured in the source repository (`experts/thangka/data/manifest.csv`).