austin-k-wang/DiffusionLAIR-SD1.5

DiffusionLAIR SD1.5

DiffusionLAIR SD1.5 is a preference-optimized fine-tune of Stable Diffusion v1.5. The UNet was fine-tuned using the Diffusion LAIR objective, while the text encoder, VAE, and other pipeline components were kept frozen.

Model details

• Base model: runwayml/stable-diffusion-v1-5
• Current base-model mirror: stable-diffusion-v1-5/stable-diffusion-v1-5
• Base-model revision: 451f4fe16113bff5a5d2269ed5ad43b0592e9a14
• Architecture: Stable Diffusion v1.5 / StableDiffusionPipeline
• Resolution: 512 × 512
• Fine-tuned component: UNet
• Weight format: Safetensors
• Weight precision: float32
• Training objective: Diffusion LAIR
• Reward/preference model: PickScore
• Training dataset: Pick-a-pic v2

Usage

import torch
from diffusers import StableDiffusionPipeline

model_id = "austin-k-wang/DiffusionLAIR-SD1.5"

pipe = StableDiffusionPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    use_safetensors=True,
).to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)

image = pipe(
    "A queen with red hair and a green and black dress stands veiled in a highly detailed and elegant digital painting.",
    generator=generator,
    guidance_scale=7.5,
    num_inference_steps=30,
).images[0]

image.save("output.png")