2023-01-29 02:15:39 +00:00
|
|
|
from diffusers import (
|
|
|
|
DiffusionPipeline,
|
|
|
|
)
|
|
|
|
from logging import getLogger
|
2023-01-29 16:31:22 +00:00
|
|
|
from typing import Any, Optional, Tuple
|
2023-01-29 02:15:39 +00:00
|
|
|
|
|
|
|
from ..params import (
|
|
|
|
Size,
|
|
|
|
)
|
2023-02-02 03:20:48 +00:00
|
|
|
from ..utils import (
|
|
|
|
run_gc,
|
|
|
|
)
|
2023-01-29 02:15:39 +00:00
|
|
|
|
|
|
|
import numpy as np
|
|
|
|
|
|
|
|
logger = getLogger(__name__)
|
|
|
|
|
|
|
|
last_pipeline_instance = None
|
|
|
|
last_pipeline_options = (None, None, None)
|
|
|
|
last_pipeline_scheduler = None
|
|
|
|
|
2023-02-02 04:21:22 +00:00
|
|
|
latent_channels = 4
|
|
|
|
latent_factor = 8
|
2023-01-29 02:15:39 +00:00
|
|
|
|
2023-02-02 04:21:22 +00:00
|
|
|
|
|
|
|
def get_latents_from_seed(seed: int, size: Size, batch: int = 1) -> np.ndarray:
|
2023-01-29 02:15:39 +00:00
|
|
|
'''
|
|
|
|
From https://www.travelneil.com/stable-diffusion-updates.html
|
|
|
|
'''
|
2023-02-02 04:21:22 +00:00
|
|
|
latents_shape = (batch, latent_channels, size.height // latent_factor,
|
|
|
|
size.width // latent_factor)
|
2023-01-29 02:15:39 +00:00
|
|
|
rng = np.random.default_rng(seed)
|
|
|
|
image_latents = rng.standard_normal(latents_shape).astype(np.float32)
|
|
|
|
return image_latents
|
|
|
|
|
|
|
|
|
2023-01-29 16:31:22 +00:00
|
|
|
def get_tile_latents(full_latents: np.ndarray, dims: Tuple[int, int, int]) -> np.ndarray:
|
|
|
|
x, y, tile = dims
|
2023-02-02 04:21:22 +00:00
|
|
|
t = tile // latent_factor
|
|
|
|
x = x // latent_factor
|
|
|
|
y = y // latent_factor
|
2023-01-29 16:31:22 +00:00
|
|
|
xt = x + t
|
|
|
|
yt = y + t
|
|
|
|
|
2023-02-02 03:20:48 +00:00
|
|
|
return full_latents[:, :, y:yt, x:xt]
|
2023-01-29 16:31:22 +00:00
|
|
|
|
|
|
|
|
2023-01-29 02:15:39 +00:00
|
|
|
def load_pipeline(pipeline: DiffusionPipeline, model: str, provider: str, scheduler: Any, device: Optional[str] = None):
|
|
|
|
global last_pipeline_instance
|
|
|
|
global last_pipeline_scheduler
|
|
|
|
global last_pipeline_options
|
|
|
|
|
|
|
|
options = (pipeline, model, provider)
|
|
|
|
if last_pipeline_instance != None and last_pipeline_options == options:
|
2023-02-03 05:34:02 +00:00
|
|
|
logger.debug('reusing existing diffusion pipeline')
|
2023-01-29 02:15:39 +00:00
|
|
|
pipe = last_pipeline_instance
|
|
|
|
else:
|
2023-02-03 05:34:02 +00:00
|
|
|
logger.debug('unloading previous diffusion pipeline')
|
2023-01-29 02:15:39 +00:00
|
|
|
last_pipeline_instance = None
|
|
|
|
last_pipeline_scheduler = None
|
2023-02-02 03:20:48 +00:00
|
|
|
run_gc()
|
2023-01-29 02:15:39 +00:00
|
|
|
|
2023-02-03 05:34:02 +00:00
|
|
|
logger.debug('loading new diffusion pipeline from %s', model)
|
2023-01-29 02:15:39 +00:00
|
|
|
pipe = pipeline.from_pretrained(
|
|
|
|
model,
|
|
|
|
provider=provider,
|
|
|
|
safety_checker=None,
|
|
|
|
scheduler=scheduler.from_pretrained(model, subfolder='scheduler')
|
|
|
|
)
|
|
|
|
|
|
|
|
if device is not None:
|
|
|
|
pipe = pipe.to(device)
|
|
|
|
|
|
|
|
last_pipeline_instance = pipe
|
|
|
|
last_pipeline_options = options
|
|
|
|
last_pipeline_scheduler = scheduler
|
|
|
|
|
|
|
|
if last_pipeline_scheduler != scheduler:
|
2023-02-03 05:34:02 +00:00
|
|
|
logger.debug('loading new diffusion scheduler')
|
2023-01-29 02:15:39 +00:00
|
|
|
scheduler = scheduler.from_pretrained(
|
|
|
|
model, subfolder='scheduler')
|
|
|
|
|
|
|
|
if device is not None:
|
|
|
|
scheduler = scheduler.to(device)
|
|
|
|
|
|
|
|
pipe.scheduler = scheduler
|
|
|
|
last_pipeline_scheduler = scheduler
|
2023-02-02 03:20:48 +00:00
|
|
|
run_gc()
|
2023-01-29 02:15:39 +00:00
|
|
|
|
|
|
|
return pipe
|