from diffusers import (
    StableDiffusionPipeline,
    PNDMScheduler,
)
from diffusers.models import AutoencoderKL
import os
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

vae = AutoencoderKL.from_pretrained(
    "stabilityai/sd-vae-ft-ema", torch_dtype=torch.float32
)
common_config = {"beta_start": 0.00085, "beta_end": 0.012, "beta_schedule": "scaled_linear"}
SCHEDULER = PNDMScheduler(**common_config, skip_prk_steps=True, steps_offset=1,)
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
shared_pipe_kwargs = dict(
    vae=vae,
    torch_dtype=torch.float32,
    revision="fp16",
    use_auth_token=HF_API_TOKEN,
    scheduler=SCHEDULER,
)


base_sd_pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", **shared_pipe_kwargs
).to(device)

pai_model_dir = '.'

playground_v1_model_dir = os.path.join(
    pai_model_dir, "snapshots/36c9e19103f6b897886fb019ebc4d8e86b566032"
)

pgv1_shared_pipe_kwargs = dict(
    vae=vae,
    torch_dtype=torch.float32,
    tokenizer=base_sd_pipe.tokenizer,
    feature_extractor=base_sd_pipe.feature_extractor,
    text_encoder=base_sd_pipe.text_encoder,
    scheduler=SCHEDULER,
    local_files_only=True, 
)

pgv1_pipe = StableDiffusionPipeline.from_pretrained(
    playground_v1_model_dir,
    **pgv1_shared_pipe_kwargs,
).to(device)

img = pgv1_pipe("Frog", num_inference_steps=30)

img[0][0].save('frog.png')