2022-09-03 17:08:45 +08:00
|
|
|
import contextlib
|
|
|
|
import json
|
|
|
|
import math
|
|
|
|
import os
|
|
|
|
import sys
|
|
|
|
|
|
|
|
import torch
|
|
|
|
import numpy as np
|
|
|
|
from PIL import Image, ImageFilter, ImageOps
|
|
|
|
import random
|
2022-09-13 17:51:57 +08:00
|
|
|
import cv2
|
|
|
|
from skimage import exposure
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-05 08:25:37 +08:00
|
|
|
import modules.sd_hijack
|
2022-09-15 18:10:16 +08:00
|
|
|
from modules import devices, prompt_parser
|
2022-09-03 17:08:45 +08:00
|
|
|
from modules.sd_hijack import model_hijack
|
|
|
|
from modules.sd_samplers import samplers, samplers_for_img2img
|
|
|
|
from modules.shared import opts, cmd_opts, state
|
|
|
|
import modules.shared as shared
|
2022-09-07 17:32:28 +08:00
|
|
|
import modules.face_restoration
|
2022-09-03 17:08:45 +08:00
|
|
|
import modules.images as images
|
2022-09-10 04:16:02 +08:00
|
|
|
import modules.styles
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-13 17:51:57 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
# some of those options should not be changed at all because they would break the model, so I removed them from options.
|
|
|
|
opt_C = 4
|
|
|
|
opt_f = 8
|
|
|
|
|
|
|
|
|
2022-09-13 17:51:57 +08:00
|
|
|
def setup_color_correction(image):
|
|
|
|
correction_target = cv2.cvtColor(np.asarray(image.copy()), cv2.COLOR_RGB2LAB)
|
|
|
|
return correction_target
|
|
|
|
|
|
|
|
|
|
|
|
def apply_color_correction(correction, image):
|
|
|
|
image = Image.fromarray(cv2.cvtColor(exposure.match_histograms(
|
|
|
|
cv2.cvtColor(
|
|
|
|
np.asarray(image),
|
|
|
|
cv2.COLOR_RGB2LAB
|
|
|
|
),
|
|
|
|
correction,
|
|
|
|
channel_axis=2
|
|
|
|
), cv2.COLOR_LAB2RGB).astype("uint8"))
|
|
|
|
|
|
|
|
return image
|
|
|
|
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
class StableDiffusionProcessing:
|
2022-09-14 22:56:21 +08:00
|
|
|
def __init__(self, sd_model=None, outpath_samples=None, outpath_grids=None, prompt="", styles=None, seed=-1, subseed=-1, subseed_strength=0, seed_resize_from_h=-1, seed_resize_from_w=-1, sampler_index=0, batch_size=1, n_iter=1, steps=50, cfg_scale=7.0, width=512, height=512, restore_faces=False, tiling=False, do_not_save_samples=False, do_not_save_grid=False, extra_generation_params=None, overlay_images=None, negative_prompt=None):
|
2022-09-03 17:08:45 +08:00
|
|
|
self.sd_model = sd_model
|
|
|
|
self.outpath_samples: str = outpath_samples
|
|
|
|
self.outpath_grids: str = outpath_grids
|
|
|
|
self.prompt: str = prompt
|
2022-09-03 22:21:15 +08:00
|
|
|
self.prompt_for_display: str = None
|
2022-09-03 17:08:45 +08:00
|
|
|
self.negative_prompt: str = (negative_prompt or "")
|
2022-09-14 22:56:21 +08:00
|
|
|
self.styles: str = styles
|
2022-09-03 17:08:45 +08:00
|
|
|
self.seed: int = seed
|
2022-09-09 22:54:04 +08:00
|
|
|
self.subseed: int = subseed
|
|
|
|
self.subseed_strength: float = subseed_strength
|
|
|
|
self.seed_resize_from_h: int = seed_resize_from_h
|
|
|
|
self.seed_resize_from_w: int = seed_resize_from_w
|
2022-09-03 17:08:45 +08:00
|
|
|
self.sampler_index: int = sampler_index
|
|
|
|
self.batch_size: int = batch_size
|
|
|
|
self.n_iter: int = n_iter
|
|
|
|
self.steps: int = steps
|
|
|
|
self.cfg_scale: float = cfg_scale
|
|
|
|
self.width: int = width
|
|
|
|
self.height: int = height
|
2022-09-07 17:32:28 +08:00
|
|
|
self.restore_faces: bool = restore_faces
|
2022-09-05 08:25:37 +08:00
|
|
|
self.tiling: bool = tiling
|
2022-09-03 17:08:45 +08:00
|
|
|
self.do_not_save_samples: bool = do_not_save_samples
|
|
|
|
self.do_not_save_grid: bool = do_not_save_grid
|
|
|
|
self.extra_generation_params: dict = extra_generation_params
|
|
|
|
self.overlay_images = overlay_images
|
|
|
|
self.paste_to = None
|
2022-09-13 17:51:57 +08:00
|
|
|
self.color_corrections = None
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-07 22:00:51 +08:00
|
|
|
def init(self, seed):
|
2022-09-03 17:08:45 +08:00
|
|
|
pass
|
|
|
|
|
|
|
|
def sample(self, x, conditioning, unconditional_conditioning):
|
|
|
|
raise NotImplementedError()
|
|
|
|
|
|
|
|
|
|
|
|
class Processed:
|
|
|
|
def __init__(self, p: StableDiffusionProcessing, images_list, seed, info):
|
|
|
|
self.images = images_list
|
|
|
|
self.prompt = p.prompt
|
2022-09-13 00:57:31 +08:00
|
|
|
self.negative_prompt = p.negative_prompt
|
2022-09-03 17:08:45 +08:00
|
|
|
self.seed = seed
|
|
|
|
self.info = info
|
|
|
|
self.width = p.width
|
|
|
|
self.height = p.height
|
|
|
|
self.sampler = samplers[p.sampler_index].name
|
|
|
|
self.cfg_scale = p.cfg_scale
|
|
|
|
self.steps = p.steps
|
|
|
|
|
|
|
|
def js(self):
|
|
|
|
obj = {
|
2022-09-03 22:21:15 +08:00
|
|
|
"prompt": self.prompt if type(self.prompt) != list else self.prompt[0],
|
2022-09-13 00:57:31 +08:00
|
|
|
"negative_prompt": self.negative_prompt if type(self.negative_prompt) != list else self.negative_prompt[0],
|
2022-09-03 22:21:15 +08:00
|
|
|
"seed": int(self.seed if type(self.seed) != list else self.seed[0]),
|
2022-09-03 17:08:45 +08:00
|
|
|
"width": self.width,
|
|
|
|
"height": self.height,
|
|
|
|
"sampler": self.sampler,
|
|
|
|
"cfg_scale": self.cfg_scale,
|
|
|
|
"steps": self.steps,
|
|
|
|
}
|
|
|
|
|
|
|
|
return json.dumps(obj)
|
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
# from https://discuss.pytorch.org/t/help-regarding-slerp-function-for-generative-model-sampling/32475/3
|
|
|
|
def slerp(val, low, high):
|
|
|
|
low_norm = low/torch.norm(low, dim=1, keepdim=True)
|
|
|
|
high_norm = high/torch.norm(high, dim=1, keepdim=True)
|
|
|
|
omega = torch.acos((low_norm*high_norm).sum(1))
|
|
|
|
so = torch.sin(omega)
|
|
|
|
res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high
|
|
|
|
return res
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
|
2022-09-14 02:49:58 +08:00
|
|
|
def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, seed_resize_from_h=0, seed_resize_from_w=0, p=None):
|
2022-09-03 17:08:45 +08:00
|
|
|
xs = []
|
2022-09-14 02:49:58 +08:00
|
|
|
|
2022-09-16 15:04:07 +08:00
|
|
|
# if we have multiple seeds, this means we are working with batch size>1; this then
|
|
|
|
# enables the generation of additional tensors with noise that the sampler will use during its processing.
|
|
|
|
# Using those pre-genrated tensors instead of siimple torch.randn allows a batch with seeds [100, 101] to
|
|
|
|
# produce the same images as with two batches [100], [101].
|
2022-09-16 14:47:03 +08:00
|
|
|
if p is not None and p.sampler is not None and len(seeds) > 1 and opts.enable_batch_seeds:
|
2022-09-14 02:49:58 +08:00
|
|
|
sampler_noises = [[] for _ in range(p.sampler.number_of_needed_noises(p))]
|
|
|
|
else:
|
|
|
|
sampler_noises = None
|
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
for i, seed in enumerate(seeds):
|
|
|
|
noise_shape = shape if seed_resize_from_h <= 0 or seed_resize_from_w <= 0 else (shape[0], seed_resize_from_h//8, seed_resize_from_w//8)
|
|
|
|
|
|
|
|
subnoise = None
|
|
|
|
if subseeds is not None:
|
|
|
|
subseed = 0 if i >= len(subseeds) else subseeds[i]
|
2022-09-12 22:32:44 +08:00
|
|
|
|
2022-09-13 01:09:32 +08:00
|
|
|
subnoise = devices.randn(subseed, noise_shape)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
# randn results depend on device; gpu and cpu get different results for same seed;
|
|
|
|
# the way I see it, it's better to do this on CPU, so that everyone gets same result;
|
2022-09-09 22:54:04 +08:00
|
|
|
# but the original script had it like this, so I do not dare change it for now because
|
2022-09-03 17:08:45 +08:00
|
|
|
# it will break everyone's seeds.
|
2022-09-13 01:09:32 +08:00
|
|
|
noise = devices.randn(seed, noise_shape)
|
2022-09-09 22:54:04 +08:00
|
|
|
|
|
|
|
if subnoise is not None:
|
|
|
|
#noise = subnoise * subseed_strength + noise * (1 - subseed_strength)
|
|
|
|
noise = slerp(subseed_strength, noise, subnoise)
|
|
|
|
|
|
|
|
if noise_shape != shape:
|
|
|
|
#noise = torch.nn.functional.interpolate(noise.unsqueeze(1), size=shape[1:], mode="bilinear").squeeze()
|
2022-09-13 01:09:32 +08:00
|
|
|
x = devices.randn(seed, shape)
|
|
|
|
dx = (shape[2] - noise_shape[2]) // 2
|
2022-09-09 22:54:04 +08:00
|
|
|
dy = (shape[1] - noise_shape[1]) // 2
|
|
|
|
w = noise_shape[2] if dx >= 0 else noise_shape[2] + 2 * dx
|
|
|
|
h = noise_shape[1] if dy >= 0 else noise_shape[1] + 2 * dy
|
|
|
|
tx = 0 if dx < 0 else dx
|
|
|
|
ty = 0 if dy < 0 else dy
|
|
|
|
dx = max(-dx, 0)
|
|
|
|
dy = max(-dy, 0)
|
|
|
|
|
|
|
|
x[:, ty:ty+h, tx:tx+w] = noise[:, dy:dy+h, dx:dx+w]
|
|
|
|
noise = x
|
|
|
|
|
2022-09-14 02:49:58 +08:00
|
|
|
if sampler_noises is not None:
|
|
|
|
cnt = p.sampler.number_of_needed_noises(p)
|
2022-09-09 22:54:04 +08:00
|
|
|
|
2022-09-14 02:49:58 +08:00
|
|
|
for j in range(cnt):
|
|
|
|
sampler_noises[j].append(devices.randn_without_seed(tuple(noise_shape)))
|
2022-09-09 22:54:04 +08:00
|
|
|
|
|
|
|
xs.append(noise)
|
2022-09-14 02:49:58 +08:00
|
|
|
|
|
|
|
if sampler_noises is not None:
|
|
|
|
p.sampler.sampler_noises = [torch.stack(n).to(shared.device) for n in sampler_noises]
|
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
x = torch.stack(xs).to(shared.device)
|
2022-09-03 17:08:45 +08:00
|
|
|
return x
|
|
|
|
|
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
def fix_seed(p):
|
|
|
|
p.seed = int(random.randrange(4294967294)) if p.seed is None or p.seed == -1 else p.seed
|
|
|
|
p.subseed = int(random.randrange(4294967294)) if p.subseed is None or p.subseed == -1 else p.subseed
|
2022-09-07 06:44:44 +08:00
|
|
|
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
def process_images(p: StableDiffusionProcessing) -> Processed:
|
|
|
|
"""this is the main loop that both txt2img and img2img use; it calls func_init once inside all the scopes and func_sample once per batch"""
|
|
|
|
|
2022-09-17 16:34:33 +08:00
|
|
|
if type(p.prompt) == list:
|
|
|
|
assert(len(p.prompt) > 0)
|
|
|
|
else:
|
|
|
|
assert p.prompt is not None
|
|
|
|
|
2022-09-12 04:24:24 +08:00
|
|
|
devices.torch_gc()
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
fix_seed(p)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
os.makedirs(p.outpath_samples, exist_ok=True)
|
|
|
|
os.makedirs(p.outpath_grids, exist_ok=True)
|
|
|
|
|
2022-09-05 08:25:37 +08:00
|
|
|
modules.sd_hijack.model_hijack.apply_circular(p.tiling)
|
|
|
|
|
2022-09-15 13:57:03 +08:00
|
|
|
comments = {}
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-14 22:56:21 +08:00
|
|
|
shared.prompt_styles.apply_styles(p)
|
2022-09-10 04:16:02 +08:00
|
|
|
|
|
|
|
if type(p.prompt) == list:
|
|
|
|
all_prompts = p.prompt
|
2022-09-03 17:08:45 +08:00
|
|
|
else:
|
2022-09-10 04:16:02 +08:00
|
|
|
all_prompts = p.batch_size * p.n_iter * [p.prompt]
|
2022-09-03 22:21:15 +08:00
|
|
|
|
2022-09-09 22:54:04 +08:00
|
|
|
if type(p.seed) == list:
|
2022-09-10 03:55:54 +08:00
|
|
|
all_seeds = p.seed
|
2022-09-03 22:21:15 +08:00
|
|
|
else:
|
2022-09-12 21:00:46 +08:00
|
|
|
all_seeds = [int(p.seed + (x if p.subseed_strength == 0 else 0)) for x in range(len(all_prompts))]
|
2022-09-09 22:54:04 +08:00
|
|
|
|
|
|
|
if type(p.subseed) == list:
|
|
|
|
all_subseeds = p.subseed
|
|
|
|
else:
|
|
|
|
all_subseeds = [int(p.subseed + x) for x in range(len(all_prompts))]
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
def infotext(iteration=0, position_in_batch=0):
|
2022-09-09 22:54:04 +08:00
|
|
|
index = position_in_batch + iteration * p.batch_size
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
generation_params = {
|
|
|
|
"Steps": p.steps,
|
|
|
|
"Sampler": samplers[p.sampler_index].name,
|
|
|
|
"CFG scale": p.cfg_scale,
|
2022-09-09 22:54:04 +08:00
|
|
|
"Seed": all_seeds[index],
|
2022-09-07 17:32:28 +08:00
|
|
|
"Face restoration": (opts.face_restoration_model if p.restore_faces else None),
|
2022-09-09 22:54:04 +08:00
|
|
|
"Size": f"{p.width}x{p.height}",
|
2022-09-17 17:05:04 +08:00
|
|
|
"Model hash": (None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash),
|
2022-09-06 02:22:30 +08:00
|
|
|
"Batch size": (None if p.batch_size < 2 else p.batch_size),
|
|
|
|
"Batch pos": (None if p.batch_size < 2 else position_in_batch),
|
2022-09-09 22:54:04 +08:00
|
|
|
"Variation seed": (None if p.subseed_strength == 0 else all_subseeds[index]),
|
|
|
|
"Variation seed strength": (None if p.subseed_strength == 0 else p.subseed_strength),
|
|
|
|
"Seed resize from": (None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"),
|
2022-09-13 04:44:36 +08:00
|
|
|
"Denoising strength": getattr(p, 'denoising_strength', None),
|
2022-09-03 17:08:45 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if p.extra_generation_params is not None:
|
|
|
|
generation_params.update(p.extra_generation_params)
|
|
|
|
|
|
|
|
generation_params_text = ", ".join([k if k == v else f'{k}: {v}' for k, v in generation_params.items() if v is not None])
|
2022-09-09 22:16:51 +08:00
|
|
|
|
|
|
|
negative_prompt_text = "\nNegative prompt: " + p.negative_prompt if p.negative_prompt else ""
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-10 04:16:02 +08:00
|
|
|
return f"{all_prompts[index]}{negative_prompt_text}\n{generation_params_text}".strip() + "".join(["\n\n" + x for x in comments])
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if os.path.exists(cmd_opts.embeddings_dir):
|
|
|
|
model_hijack.load_textual_inversion_embeddings(cmd_opts.embeddings_dir, p.sd_model)
|
|
|
|
|
|
|
|
output_images = []
|
|
|
|
precision_scope = torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext
|
|
|
|
ema_scope = (contextlib.nullcontext if cmd_opts.lowvram else p.sd_model.ema_scope)
|
|
|
|
with torch.no_grad(), precision_scope("cuda"), ema_scope():
|
2022-09-07 22:00:51 +08:00
|
|
|
p.init(seed=all_seeds[0])
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-06 15:11:25 +08:00
|
|
|
if state.job_count == -1:
|
|
|
|
state.job_count = p.n_iter
|
2022-09-06 07:09:01 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
for n in range(p.n_iter):
|
|
|
|
if state.interrupted:
|
|
|
|
break
|
|
|
|
|
|
|
|
prompts = all_prompts[n * p.batch_size:(n + 1) * p.batch_size]
|
|
|
|
seeds = all_seeds[n * p.batch_size:(n + 1) * p.batch_size]
|
2022-09-10 21:16:18 +08:00
|
|
|
subseeds = all_subseeds[n * p.batch_size:(n + 1) * p.batch_size]
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-17 16:34:33 +08:00
|
|
|
if (len(prompts) == 0):
|
|
|
|
break
|
|
|
|
|
2022-09-15 18:10:16 +08:00
|
|
|
#uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt])
|
|
|
|
#c = p.sd_model.get_learned_conditioning(prompts)
|
|
|
|
uc = prompt_parser.get_learned_conditioning(len(prompts) * [p.negative_prompt], p.steps)
|
|
|
|
c = prompt_parser.get_learned_conditioning(prompts, p.steps)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if len(model_hijack.comments) > 0:
|
2022-09-15 13:57:03 +08:00
|
|
|
for comment in model_hijack.comments:
|
|
|
|
comments[comment] = 1
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
# we manually generate all input noises because each one should have a specific seed
|
2022-09-14 02:49:58 +08:00
|
|
|
x = create_random_tensors([opt_C, p.height // opt_f, p.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, seed_resize_from_h=p.seed_resize_from_h, seed_resize_from_w=p.seed_resize_from_w, p=p)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if p.n_iter > 1:
|
|
|
|
shared.state.job = f"Batch {n+1} out of {p.n_iter}"
|
|
|
|
|
|
|
|
samples_ddim = p.sample(x=x, conditioning=c, unconditional_conditioning=uc)
|
2022-09-07 00:33:51 +08:00
|
|
|
if state.interrupted:
|
|
|
|
|
|
|
|
# if we are interruped, sample returns just noise
|
|
|
|
# use the image collected previously in sampler loop
|
|
|
|
samples_ddim = shared.state.current_latent
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
x_samples_ddim = p.sd_model.decode_first_stage(samples_ddim)
|
|
|
|
x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
|
|
|
|
|
2022-09-13 08:15:35 +08:00
|
|
|
if opts.filter_nsfw:
|
2022-09-13 13:34:41 +08:00
|
|
|
import modules.safety as safety
|
|
|
|
x_samples_ddim = modules.safety.censor_batch(x_samples_ddim)
|
2022-09-13 08:15:35 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
for i, x_sample in enumerate(x_samples_ddim):
|
|
|
|
x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2)
|
|
|
|
x_sample = x_sample.astype(np.uint8)
|
|
|
|
|
2022-09-07 17:32:28 +08:00
|
|
|
if p.restore_faces:
|
2022-09-12 22:47:36 +08:00
|
|
|
if opts.save and not p.do_not_save_samples and opts.save_images_before_face_restoration:
|
|
|
|
images.save_image(Image.fromarray(x_sample), p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p)
|
|
|
|
|
2022-09-12 04:24:24 +08:00
|
|
|
devices.torch_gc()
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-07 17:32:28 +08:00
|
|
|
x_sample = modules.face_restoration.restore_faces(x_sample)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
image = Image.fromarray(x_sample)
|
|
|
|
|
2022-09-13 17:51:57 +08:00
|
|
|
if p.color_corrections is not None and i < len(p.color_corrections):
|
|
|
|
image = apply_color_correction(p.color_corrections[i], image)
|
2022-09-12 22:47:36 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
if p.overlay_images is not None and i < len(p.overlay_images):
|
|
|
|
overlay = p.overlay_images[i]
|
|
|
|
|
|
|
|
if p.paste_to is not None:
|
|
|
|
x, y, w, h = p.paste_to
|
|
|
|
base_image = Image.new('RGBA', (overlay.width, overlay.height))
|
|
|
|
image = images.resize_image(1, image, w, h)
|
|
|
|
base_image.paste(image, (x, y))
|
|
|
|
image = base_image
|
|
|
|
|
|
|
|
image = image.convert('RGBA')
|
|
|
|
image.alpha_composite(overlay)
|
|
|
|
image = image.convert('RGB')
|
|
|
|
|
|
|
|
if opts.samples_save and not p.do_not_save_samples:
|
2022-09-12 20:41:30 +08:00
|
|
|
images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
output_images.append(image)
|
|
|
|
|
2022-09-06 07:09:01 +08:00
|
|
|
state.nextjob()
|
|
|
|
|
2022-09-18 06:18:30 +08:00
|
|
|
p.color_corrections = None
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
unwanted_grid_because_of_img_count = len(output_images) < 2 and opts.grid_only_if_multiple
|
2022-09-14 15:34:44 +08:00
|
|
|
if (opts.return_grid or opts.grid_save) and not p.do_not_save_grid and not unwanted_grid_because_of_img_count:
|
2022-09-03 22:21:15 +08:00
|
|
|
grid = images.image_grid(output_images, p.batch_size)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-14 15:34:44 +08:00
|
|
|
if opts.return_grid:
|
2022-09-03 17:08:45 +08:00
|
|
|
output_images.insert(0, grid)
|
|
|
|
|
|
|
|
if opts.grid_save:
|
2022-09-14 15:34:44 +08:00
|
|
|
images.save_image(grid, p.outpath_grids, "grid", all_seeds[0], all_prompts[0], opts.grid_format, info=infotext(), short_filename=not opts.grid_extended_filename, p=p)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-12 04:24:24 +08:00
|
|
|
devices.torch_gc()
|
2022-09-09 22:54:04 +08:00
|
|
|
return Processed(p, output_images, all_seeds[0], infotext())
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
|
|
|
|
class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
|
|
|
|
sampler = None
|
|
|
|
|
2022-09-07 22:00:51 +08:00
|
|
|
def init(self, seed):
|
2022-09-03 17:08:45 +08:00
|
|
|
self.sampler = samplers[self.sampler_index].constructor(self.sd_model)
|
|
|
|
|
|
|
|
def sample(self, x, conditioning, unconditional_conditioning):
|
|
|
|
samples_ddim = self.sampler.sample(self, x, conditioning, unconditional_conditioning)
|
|
|
|
return samples_ddim
|
|
|
|
|
|
|
|
|
|
|
|
def get_crop_region(mask, pad=0):
|
|
|
|
h, w = mask.shape
|
|
|
|
|
|
|
|
crop_left = 0
|
|
|
|
for i in range(w):
|
|
|
|
if not (mask[:, i] == 0).all():
|
|
|
|
break
|
|
|
|
crop_left += 1
|
|
|
|
|
|
|
|
crop_right = 0
|
|
|
|
for i in reversed(range(w)):
|
|
|
|
if not (mask[:, i] == 0).all():
|
|
|
|
break
|
|
|
|
crop_right += 1
|
|
|
|
|
|
|
|
crop_top = 0
|
|
|
|
for i in range(h):
|
|
|
|
if not (mask[i] == 0).all():
|
|
|
|
break
|
|
|
|
crop_top += 1
|
|
|
|
|
|
|
|
crop_bottom = 0
|
|
|
|
for i in reversed(range(h)):
|
|
|
|
if not (mask[i] == 0).all():
|
|
|
|
break
|
|
|
|
crop_bottom += 1
|
|
|
|
|
|
|
|
return (
|
|
|
|
int(max(crop_left-pad, 0)),
|
|
|
|
int(max(crop_top-pad, 0)),
|
|
|
|
int(min(w - crop_right + pad, w)),
|
|
|
|
int(min(h - crop_bottom + pad, h))
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
def fill(image, mask):
|
|
|
|
image_mod = Image.new('RGBA', (image.width, image.height))
|
|
|
|
|
|
|
|
image_masked = Image.new('RGBa', (image.width, image.height))
|
|
|
|
image_masked.paste(image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(mask.convert('L')))
|
|
|
|
|
|
|
|
image_masked = image_masked.convert('RGBa')
|
|
|
|
|
2022-09-04 06:29:43 +08:00
|
|
|
for radius, repeats in [(256, 1), (64, 1), (16, 2), (4, 4), (2, 2), (0, 1)]:
|
2022-09-03 17:08:45 +08:00
|
|
|
blurred = image_masked.filter(ImageFilter.GaussianBlur(radius)).convert('RGBA')
|
|
|
|
for _ in range(repeats):
|
|
|
|
image_mod.alpha_composite(blurred)
|
|
|
|
|
|
|
|
return image_mod.convert("RGB")
|
|
|
|
|
|
|
|
|
|
|
|
class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
|
|
|
|
sampler = None
|
|
|
|
|
2022-09-04 02:02:38 +08:00
|
|
|
def __init__(self, init_images=None, resize_mode=0, denoising_strength=0.75, mask=None, mask_blur=4, inpainting_fill=0, inpaint_full_res=True, inpainting_mask_invert=0, **kwargs):
|
2022-09-03 17:08:45 +08:00
|
|
|
super().__init__(**kwargs)
|
|
|
|
|
|
|
|
self.init_images = init_images
|
|
|
|
self.resize_mode: int = resize_mode
|
|
|
|
self.denoising_strength: float = denoising_strength
|
|
|
|
self.init_latent = None
|
|
|
|
self.image_mask = mask
|
2022-09-04 06:29:43 +08:00
|
|
|
#self.image_unblurred_mask = None
|
|
|
|
self.latent_mask = None
|
2022-09-03 17:08:45 +08:00
|
|
|
self.mask_for_overlay = None
|
|
|
|
self.mask_blur = mask_blur
|
|
|
|
self.inpainting_fill = inpainting_fill
|
|
|
|
self.inpaint_full_res = inpaint_full_res
|
2022-09-04 02:02:38 +08:00
|
|
|
self.inpainting_mask_invert = inpainting_mask_invert
|
2022-09-03 17:08:45 +08:00
|
|
|
self.mask = None
|
|
|
|
self.nmask = None
|
|
|
|
|
2022-09-07 22:00:51 +08:00
|
|
|
def init(self, seed):
|
2022-09-03 17:08:45 +08:00
|
|
|
self.sampler = samplers_for_img2img[self.sampler_index].constructor(self.sd_model)
|
|
|
|
crop_region = None
|
|
|
|
|
|
|
|
if self.image_mask is not None:
|
2022-09-04 02:02:38 +08:00
|
|
|
self.image_mask = self.image_mask.convert('L')
|
|
|
|
|
|
|
|
if self.inpainting_mask_invert:
|
|
|
|
self.image_mask = ImageOps.invert(self.image_mask)
|
|
|
|
|
2022-09-04 06:29:43 +08:00
|
|
|
#self.image_unblurred_mask = self.image_mask
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
if self.mask_blur > 0:
|
2022-09-04 02:02:38 +08:00
|
|
|
self.image_mask = self.image_mask.filter(ImageFilter.GaussianBlur(self.mask_blur))
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if self.inpaint_full_res:
|
|
|
|
self.mask_for_overlay = self.image_mask
|
|
|
|
mask = self.image_mask.convert('L')
|
2022-09-06 04:08:06 +08:00
|
|
|
crop_region = get_crop_region(np.array(mask), opts.upscale_at_full_resolution_padding)
|
2022-09-03 17:08:45 +08:00
|
|
|
x1, y1, x2, y2 = crop_region
|
|
|
|
|
|
|
|
mask = mask.crop(crop_region)
|
|
|
|
self.image_mask = images.resize_image(2, mask, self.width, self.height)
|
|
|
|
self.paste_to = (x1, y1, x2-x1, y2-y1)
|
|
|
|
else:
|
|
|
|
self.image_mask = images.resize_image(self.resize_mode, self.image_mask, self.width, self.height)
|
2022-09-07 05:58:01 +08:00
|
|
|
np_mask = np.array(self.image_mask)
|
2022-09-13 22:14:40 +08:00
|
|
|
np_mask = np.clip((np_mask.astype(np.float32)) * 2, 0, 255).astype(np.uint8)
|
2022-09-07 05:58:01 +08:00
|
|
|
self.mask_for_overlay = Image.fromarray(np_mask)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
self.overlay_images = []
|
|
|
|
|
2022-09-07 22:00:51 +08:00
|
|
|
latent_mask = self.latent_mask if self.latent_mask is not None else self.image_mask
|
|
|
|
|
2022-09-16 13:33:47 +08:00
|
|
|
add_color_corrections = opts.img2img_color_correction and self.color_corrections is None
|
|
|
|
if add_color_corrections:
|
|
|
|
self.color_corrections = []
|
2022-09-03 17:08:45 +08:00
|
|
|
imgs = []
|
|
|
|
for img in self.init_images:
|
|
|
|
image = img.convert("RGB")
|
|
|
|
|
|
|
|
if crop_region is None:
|
|
|
|
image = images.resize_image(self.resize_mode, image, self.width, self.height)
|
|
|
|
|
|
|
|
if self.image_mask is not None:
|
|
|
|
image_masked = Image.new('RGBa', (image.width, image.height))
|
|
|
|
image_masked.paste(image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(self.mask_for_overlay.convert('L')))
|
|
|
|
|
|
|
|
self.overlay_images.append(image_masked.convert('RGBA'))
|
|
|
|
|
|
|
|
if crop_region is not None:
|
|
|
|
image = image.crop(crop_region)
|
|
|
|
image = images.resize_image(2, image, self.width, self.height)
|
|
|
|
|
2022-09-08 15:03:21 +08:00
|
|
|
if self.image_mask is not None:
|
|
|
|
if self.inpainting_fill != 1:
|
|
|
|
image = fill(image, latent_mask)
|
|
|
|
|
2022-09-16 13:33:47 +08:00
|
|
|
if add_color_corrections:
|
2022-09-13 17:51:57 +08:00
|
|
|
self.color_corrections.append(setup_color_correction(image))
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
image = np.array(image).astype(np.float32) / 255.0
|
|
|
|
image = np.moveaxis(image, 2, 0)
|
|
|
|
|
|
|
|
imgs.append(image)
|
|
|
|
|
|
|
|
if len(imgs) == 1:
|
|
|
|
batch_images = np.expand_dims(imgs[0], axis=0).repeat(self.batch_size, axis=0)
|
|
|
|
if self.overlay_images is not None:
|
|
|
|
self.overlay_images = self.overlay_images * self.batch_size
|
|
|
|
elif len(imgs) <= self.batch_size:
|
|
|
|
self.batch_size = len(imgs)
|
|
|
|
batch_images = np.array(imgs)
|
|
|
|
else:
|
|
|
|
raise RuntimeError(f"bad number of images passed: {len(imgs)}; expecting {self.batch_size} or less")
|
|
|
|
|
|
|
|
image = torch.from_numpy(batch_images)
|
|
|
|
image = 2. * image - 1.
|
|
|
|
image = image.to(shared.device)
|
|
|
|
|
|
|
|
self.init_latent = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image))
|
|
|
|
|
|
|
|
if self.image_mask is not None:
|
2022-09-07 22:00:51 +08:00
|
|
|
init_mask = latent_mask
|
2022-09-04 06:29:43 +08:00
|
|
|
latmask = init_mask.convert('RGB').resize((self.init_latent.shape[3], self.init_latent.shape[2]))
|
2022-09-13 01:09:32 +08:00
|
|
|
latmask = np.moveaxis(np.array(latmask, dtype=np.float32), 2, 0) / 255
|
2022-09-03 17:08:45 +08:00
|
|
|
latmask = latmask[0]
|
2022-09-07 22:00:51 +08:00
|
|
|
latmask = np.around(latmask)
|
2022-09-03 17:08:45 +08:00
|
|
|
latmask = np.tile(latmask[None], (4, 1, 1))
|
|
|
|
|
|
|
|
self.mask = torch.asarray(1.0 - latmask).to(shared.device).type(self.sd_model.dtype)
|
|
|
|
self.nmask = torch.asarray(latmask).to(shared.device).type(self.sd_model.dtype)
|
|
|
|
|
|
|
|
if self.inpainting_fill == 2:
|
2022-09-07 22:00:51 +08:00
|
|
|
self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], [seed + x + 1 for x in range(self.init_latent.shape[0])]) * self.nmask
|
2022-09-03 17:08:45 +08:00
|
|
|
elif self.inpainting_fill == 3:
|
|
|
|
self.init_latent = self.init_latent * self.mask
|
|
|
|
|
|
|
|
def sample(self, x, conditioning, unconditional_conditioning):
|
|
|
|
samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning)
|
|
|
|
|
|
|
|
if self.mask is not None:
|
|
|
|
samples = samples * self.nmask + self.init_latent * self.mask
|
|
|
|
|
|
|
|
return samples
|