Files
easydiffusion/ui/easydiffusion/tasks/render_images.py

311 lines
10 KiB
Python

import json
import pprint
import queue
import time
from PIL import Image
from easydiffusion import model_manager, runtime
from easydiffusion.types import GenerateImageRequest, ModelsData, OutputFormatData, SaveToDiskData
from easydiffusion.types import Image as ResponseImage
from easydiffusion.types import GenerateImageResponse, RenderTaskData
from easydiffusion.utils import get_printable_request, log, save_images_to_disk, filter_nsfw
from sdkit.utils import (
img_to_base64_str,
base64_str_to_img,
img_to_buffer,
resize_img,
get_image,
log,
)
from .task import Task
class RenderTask(Task):
"For image generation"
def __init__(
self,
req: GenerateImageRequest,
task_data: RenderTaskData,
models_data: ModelsData,
output_format: OutputFormatData,
save_data: SaveToDiskData,
):
super().__init__(task_data.session_id)
task_data.request_id = self.id
self.render_request = req # Initial Request
self.task_data = task_data
self.models_data = models_data
self.output_format = output_format
self.save_data = save_data
self.temp_images: list = [None] * req.num_outputs * (1 if task_data.show_only_filtered_image else 2)
def run(self):
"Runs the image generation task on the assigned thread"
from easydiffusion import task_manager, app
from easydiffusion.backend_manager import backend
context = runtime.context
config = app.getConfig()
if config.get("block_nsfw", False): # override if set on the server
self.task_data.block_nsfw = True
def step_callback():
task_manager.keep_task_alive(self)
task_manager.current_state = task_manager.ServerStates.Rendering
if isinstance(task_manager.current_state_error, (SystemExit, StopAsyncIteration)) or isinstance(
self.error, StopAsyncIteration
):
backend.stop_rendering(context)
if isinstance(task_manager.current_state_error, StopAsyncIteration):
self.error = task_manager.current_state_error
task_manager.current_state_error = None
log.info(f"Session {self.session_id} sent cancel signal for task {self.id}")
task_manager.current_state = task_manager.ServerStates.LoadingModel
model_manager.resolve_model_paths(self.models_data)
models_to_force_reload = []
if runtime.set_vram_optimizations(context) or self.has_param_changed(context, "clip_skip"):
models_to_force_reload.append("stable-diffusion")
model_manager.reload_models_if_necessary(context, self.models_data, models_to_force_reload)
model_manager.fail_if_models_did_not_load(context)
task_manager.current_state = task_manager.ServerStates.Rendering
self.response = make_images(
context,
self.render_request,
self.task_data,
self.models_data,
self.output_format,
self.save_data,
self.buffer_queue,
self.temp_images,
step_callback,
self,
)
def has_param_changed(self, context, param_name):
if not getattr(context, "test_diffusers", False):
return False
if "stable-diffusion" not in context.models or "params" not in context.models["stable-diffusion"]:
return True
model = context.models["stable-diffusion"]
new_val = self.models_data.model_params.get("stable-diffusion", {}).get(param_name, False)
return model["params"].get(param_name) != new_val
def make_images(
context,
req: GenerateImageRequest,
task_data: RenderTaskData,
models_data: ModelsData,
output_format: OutputFormatData,
save_data: SaveToDiskData,
data_queue: queue.Queue,
task_temp_images: list,
step_callback,
task,
):
print_task_info(req, task_data, models_data, output_format, save_data)
images, seeds = make_images_internal(
context,
req,
task_data,
models_data,
output_format,
save_data,
data_queue,
task_temp_images,
step_callback,
task,
)
res = GenerateImageResponse(
req, task_data, models_data, output_format, save_data, images=construct_response(images, seeds, output_format)
)
res = res.json()
data_queue.put(json.dumps(res))
log.info("Task completed")
return res
def print_task_info(
req: GenerateImageRequest,
task_data: RenderTaskData,
models_data: ModelsData,
output_format: OutputFormatData,
save_data: SaveToDiskData,
):
req_str = pprint.pformat(get_printable_request(req, task_data, models_data, output_format, save_data)).replace(
"[", "\["
)
task_str = pprint.pformat(task_data.dict()).replace("[", "\[")
models_data = pprint.pformat(models_data.dict()).replace("[", "\[")
output_format = pprint.pformat(output_format.dict()).replace("[", "\[")
save_data = pprint.pformat(save_data.dict()).replace("[", "\[")
log.info(f"request: {req_str}")
log.info(f"task data: {task_str}")
log.info(f"models data: {models_data}")
log.info(f"output format: {output_format}")
log.info(f"save data: {save_data}")
def make_images_internal(
context,
req: GenerateImageRequest,
task_data: RenderTaskData,
models_data: ModelsData,
output_format: OutputFormatData,
save_data: SaveToDiskData,
data_queue: queue.Queue,
task_temp_images: list,
step_callback,
task,
):
from easydiffusion.backend_manager import backend
# prep the nsfw_filter
if task_data.block_nsfw:
filter_nsfw([Image.new("RGB", (1, 1))]) # hack - ensures that the model is available
images = generate_images_internal(
context,
req,
task_data,
models_data,
output_format,
data_queue,
task_temp_images,
step_callback,
task_data.stream_image_progress,
task_data.stream_image_progress_interval,
)
user_stopped = isinstance(task.error, StopAsyncIteration)
filters, filter_params = task_data.filters, task_data.filter_params
if len(filters) > 0 and not user_stopped:
filtered_images = backend.filter_images(context, images, filters, filter_params, input_type="base64")
else:
filtered_images = images
if task_data.block_nsfw:
filtered_images = filter_nsfw(filtered_images)
if save_data.save_to_disk_path is not None:
images_pil = [base64_str_to_img(img) for img in images]
filtered_images_pil = [base64_str_to_img(img) for img in filtered_images]
save_images_to_disk(images_pil, filtered_images_pil, req, task_data, models_data, output_format, save_data)
seeds = [*range(req.seed, req.seed + len(images))]
if task_data.show_only_filtered_image or filtered_images is images:
return filtered_images, seeds
else:
return images + filtered_images, seeds + seeds
def generate_images_internal(
context,
req: GenerateImageRequest,
task_data: RenderTaskData,
models_data: ModelsData,
output_format: OutputFormatData,
data_queue: queue.Queue,
task_temp_images: list,
step_callback,
stream_image_progress: bool,
stream_image_progress_interval: int,
):
from easydiffusion.backend_manager import backend
callback = make_step_callback(context, req, task_data, data_queue, task_temp_images, step_callback)
req.width, req.height = map(lambda x: x - x % 8, (req.width, req.height)) # clamp to 8
if req.control_image and task_data.control_filter_to_apply:
req.controlnet_filter = task_data.control_filter_to_apply
if req.init_image is not None and int(req.num_inference_steps * req.prompt_strength) == 0:
req.prompt_strength = 1 / req.num_inference_steps if req.num_inference_steps > 0 else 1
backend.set_options(
context,
output_format=output_format.output_format,
output_quality=output_format.output_quality,
output_lossless=output_format.output_lossless,
vae_tiling=task_data.enable_vae_tiling,
stream_image_progress=stream_image_progress,
stream_image_progress_interval=stream_image_progress_interval,
clip_skip=2 if task_data.clip_skip else 1,
)
images = backend.generate_images(context, callback=callback, output_type="base64", **req.dict())
return images
def construct_response(images: list, seeds: list, output_format: OutputFormatData):
return [ResponseImage(data=img, seed=seed) for img, seed in zip(images, seeds)]
def make_step_callback(
context,
req: GenerateImageRequest,
task_data: RenderTaskData,
data_queue: queue.Queue,
task_temp_images: list,
step_callback,
):
from easydiffusion.backend_manager import backend
n_steps = req.num_inference_steps if req.init_image is None else int(req.num_inference_steps * req.prompt_strength)
last_callback_time = -1
def update_temp_img(images, task_temp_images: list):
partial_images = []
if images is None:
return []
if task_data.block_nsfw:
images = filter_nsfw(images, print_log=False)
for i, img in enumerate(images):
img = img.convert("RGB")
img = resize_img(img, req.width, req.height)
buf = img_to_buffer(img, output_format="JPEG")
task_temp_images[i] = buf
partial_images.append({"path": f"/image/tmp/{task_data.request_id}/{i}"})
del images
return partial_images
def on_image_step(images, i, *args):
nonlocal last_callback_time
step_time = time.time() - last_callback_time if last_callback_time != -1 else -1
last_callback_time = time.time()
progress = {"step": i, "step_time": step_time, "total_steps": n_steps}
if images is not None:
progress["output"] = update_temp_img(images, task_temp_images)
data_queue.put(json.dumps(progress))
step_callback()
return on_image_step