metadata
license: apache-2.0
datasets:
- yuvalkirstain/pickapic_v2
base_model:
- stable-diffusion-v1-5/stable-diffusion-v1-5
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-generation-inference
Pretrained SD-1.5 weight for SePPO: Semi-Policy Preference Optimization for Diffusion Alignment
See Github Repo: SePPO
Paper Report: Daily Paper
Inference Code:
import os
import argparse
import numpy as np
import torch
from diffusers import StableDiffusionPipeline, UNet2DConditionModel
from PIL import Image
torch.set_grad_enabled(False)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Generate images and calculate scores.")
parser.add_argument('--unet_checkpoint', type=str, required=True, help="Path to the UNet model checkpoint")
parser.add_argument('--prompt', type=str, required=True, help="Prompt")
args = parser.parse_args()
unet = UNet2DConditionModel.from_pretrained(args.unet_checkpoint, torch_dtype=torch.float16).to('cuda')
pipe = StableDiffusionPipeline.from_pretrained("pt-sk/stable-diffusion-1.5", torch_dtype=torch.float16)
pipe = pipe.to('cuda')
pipe.safety_checker = None
pipe.unet = unet
generator = torch.Generator(device='cuda').manual_seed(0)
gs = 7.5
ims = pipe(prompt=args.prompt, generator=generator, guidance_scale=gs).images[0]
img_path = os.path.join('SePPO', "0.png")
if isinstance(ims, np.ndarray):
ims = Image.fromarray(ims)
ims.save(img_path, format='PNG')