SachinHatikankan/warli-sdxl-lora

Warli Finetuning LoRA

Prompt

wrli style, forest with animals

Prompt

wrli style, mumbai city with big buildings

Prompt

wrli style, villagers fishing from the side of the river

Prompt

wrli style, passengers in a train

---

Model

Base Model	stabilityai/stable-diffusion-xl-base-1.0
VAE	madebyollin/sdxl-vae-fp16-fix
Fine-tuning Method	LoRA (rank 16)
Task	Text-to-image · style transfer
Trigger Word	wrli style
Training Steps	1500
Resolution	768
License	MIT

---

Dataset

21 handpicked images of authentic Warli paintings, sourced respectfully and curated for consistency, white-on-brown palette, traditional stick-figure grammar, and varied subjects (dancing, farming, the tarpa dance circle). The set was kept deliberately tight and trained with a single instance prompt to bind the style cleanly to the trigger.

---

Training Target

The LoRA is optimised to reproduce:

• White pigment figures on an earthen brown background
• Geometric stick-figure anatomy (triangular torsos, circular heads)
• Dense circular and concentric narrative compositions
• Sun/wheel motifs and decorative borders
• Clean high-contrast line-art

---

Usage

Prepend wrli style to your prompt. Short prompts work best — long, detailed prompts can override the style binding (see Limitations).

!pip install -q -U "torchao>=0.16.0"

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights("SachinHatikankan/warli-sdxl-lora")

image = pipe("wrli style, a village festival", guidance_scale=9.0).images[0]
image.save("warli.png")

---

Limitations

• Trained with a single instance prompt, so the model has strong style binding but weak subject control. Short prompts keep the trigger dominant; long, semantically heavy prompts (e.g. named landmarks) can override the style and revert toward base-model realism.
• Best suited to generic Warli-friendly subjects (people, animals, villages, activities) rather than specific named objects.
• Optimised for flat line-art and not photorealism. Trained at 768px due to source image resolution constraints.