import torch
from diffusers import AutoencoderKL, DiffusionPipeline
from safetensors.torch import load_file

# Load pipeline
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    vae=vae,
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda")

# Load LoRA weights
pipe.load_lora_weights("aaronw122/prtkl-sdxl-lora", weight_name="pytorch_lora_weights.safetensors")

# Load textual inversion embeddings
from huggingface_hub import hf_hub_download
ti_path = hf_hub_download("aaronw122/prtkl-sdxl-lora", "results_emb.safetensors")
state_dict = load_file(ti_path)
pipe.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer)
pipe.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2)

# Generate
image = pipe(
    "<s0><s1>, a figure dancing with arms raised, white background",
    negative_prompt="photorealistic, detailed, shading, gradient, gray, color, dense",
    num_inference_steps=30,
    guidance_scale=5.0,
).images[0]
image.save("particle_art.png")

Training Details

Base model: SDXL 1.0
Method: Pivotal Tuning (LoRA + Textual Inversion)
Optimizer: AdamW 8-bit
Checkpoint: Step 1800
Rank: 32
Trigger tokens: <s0><s1> (mapped from prtkl)
TI frozen after: Step 500

Prompt Tips

Always prefix prompts with <s0><s1>,
Add "white background" for cleanest results
Use negative prompt: "photorealistic, detailed, shading, gradient, gray, color, dense, beige, tan, sepia"
Guidance scale 5.0 works well