434 lines
20 KiB
Python
434 lines
20 KiB
Python
# These nodes were made using code from the Deforum extension for A1111 webui
|
|
# You can find the project here: https://github.com/deforum-art/sd-webui-deforum
|
|
|
|
import numexpr
|
|
import torch
|
|
import numpy as np
|
|
import pandas as pd
|
|
import re
|
|
|
|
from .ScheduleFuncs import *
|
|
|
|
def prepare_batch_prompt(prompt_series, max_frames, frame_idx, prompt_weight_1=0, prompt_weight_2=0, prompt_weight_3=0,
|
|
prompt_weight_4=0): # calculate expressions from the text input and return a string
|
|
max_f = max_frames - 1
|
|
pattern = r'`.*?`' # set so the expression will be read between two backticks (``)
|
|
regex = re.compile(pattern)
|
|
prompt_parsed = str(prompt_series)
|
|
|
|
for match in regex.finditer(prompt_parsed):
|
|
matched_string = match.group(0)
|
|
parsed_string = matched_string.replace('t', f'{frame_idx}').replace("pw_a", f"{prompt_weight_1}").replace("pw_b",
|
|
f"{prompt_weight_2}").replace("pw_c", f"{prompt_weight_3}").replace("pw_d",
|
|
f"{prompt_weight_4}").replace("max_f",
|
|
f"{max_f}").replace('`', '') # replace t, max_f and `` respectively
|
|
parsed_value = numexpr.evaluate(parsed_string)
|
|
prompt_parsed = prompt_parsed.replace(matched_string, str(parsed_value))
|
|
return prompt_parsed.strip()
|
|
|
|
def batch_split_weighted_subprompts(text, pre_text, app_text):
|
|
pos = {}
|
|
neg = {}
|
|
pre_text = str(pre_text)
|
|
app_text = str(app_text)
|
|
|
|
if "--neg" in pre_text:
|
|
pre_pos, pre_neg = pre_text.split("--neg")
|
|
else:
|
|
pre_pos, pre_neg = pre_text, ""
|
|
|
|
if "--neg" in app_text:
|
|
app_pos, app_neg = app_text.split("--neg")
|
|
else:
|
|
app_pos, app_neg = app_text, ""
|
|
|
|
for frame, prompt in text.items():
|
|
negative_prompts = ""
|
|
positive_prompts = ""
|
|
prompt_split = prompt.split("--neg")
|
|
|
|
if len(prompt_split) > 1:
|
|
positive_prompts, negative_prompts = prompt_split[0], prompt_split[1]
|
|
else:
|
|
positive_prompts = prompt_split[0]
|
|
|
|
pos[frame] = ""
|
|
neg[frame] = ""
|
|
pos[frame] += (str(pre_pos) + " " + positive_prompts + " " + str(app_pos))
|
|
neg[frame] += (str(pre_neg) + " " + negative_prompts + " " + str(app_neg))
|
|
if pos[frame].endswith('0'):
|
|
pos[frame] = pos[frame][:-1]
|
|
if neg[frame].endswith('0'):
|
|
neg[frame] = neg[frame][:-1]
|
|
return pos, neg
|
|
|
|
def interpolate_prompt_series(animation_prompts, max_frames, start_frame, pre_text, app_text, prompt_weight_1=[],
|
|
prompt_weight_2=[], prompt_weight_3=[], prompt_weight_4=[], Is_print = False):
|
|
|
|
max_f = max_frames # needed for numexpr even though it doesn't look like it's in use.
|
|
parsed_animation_prompts = {}
|
|
for key, value in animation_prompts.items():
|
|
if check_is_number(key): # default case 0:(1 + t %5), 30:(5-t%2)
|
|
parsed_animation_prompts[key] = value
|
|
else: # math on the left hand side case 0:(1 + t %5), maxKeyframes/2:(5-t%2)
|
|
parsed_animation_prompts[int(numexpr.evaluate(key))] = value
|
|
|
|
sorted_prompts = sorted(parsed_animation_prompts.items(), key=lambda item: int(item[0]))
|
|
|
|
# Automatically set the first keyframe to 0 if it's missing
|
|
if sorted_prompts[0][0] != "0":
|
|
sorted_prompts.insert(0, ("0", sorted_prompts[0][1]))
|
|
|
|
# Automatically set the last keyframe to the maximum number of frames
|
|
if sorted_prompts[-1][0] != str(max_frames):
|
|
sorted_prompts.append((str(max_frames), sorted_prompts[-1][1]))
|
|
# Setup containers for interpolated prompts
|
|
cur_prompt_series = pd.Series([np.nan for a in range(max_frames)])
|
|
nxt_prompt_series = pd.Series([np.nan for a in range(max_frames)])
|
|
|
|
# simple array for strength values
|
|
weight_series = [np.nan] * max_frames
|
|
|
|
# in case there is only one keyed promt, set all prompts to that prompt
|
|
if len(sorted_prompts) == 1:
|
|
for i in range(0, len(cur_prompt_series) - 1):
|
|
current_prompt = sorted_prompts[0][1]
|
|
cur_prompt_series[i] = str(current_prompt)
|
|
nxt_prompt_series[i] = str(current_prompt)
|
|
|
|
# Initialized outside of loop for nan check
|
|
current_key = 0
|
|
next_key = 0
|
|
|
|
if type(prompt_weight_1) in {int, float}:
|
|
prompt_weight_1 = tuple([prompt_weight_1] * max_frames)
|
|
|
|
if type(prompt_weight_2) in {int, float}:
|
|
prompt_weight_2 = tuple([prompt_weight_2] * max_frames)
|
|
|
|
if type(prompt_weight_3) in {int, float}:
|
|
prompt_weight_3 = tuple([prompt_weight_3] * max_frames)
|
|
|
|
if type(prompt_weight_4) in {int, float}:
|
|
prompt_weight_4 = tuple([prompt_weight_4] * max_frames)
|
|
|
|
# For every keyframe prompt except the last
|
|
for i in range(0, len(sorted_prompts) - 1):
|
|
# Get current and next keyframe
|
|
current_key = int(sorted_prompts[i][0])
|
|
next_key = int(sorted_prompts[i + 1][0])
|
|
|
|
# Ensure there's no weird ordering issues or duplication in the animation prompts
|
|
# (unlikely because we sort above, and the json parser will strip dupes)
|
|
if current_key >= next_key:
|
|
print(
|
|
f"WARNING: Sequential prompt keyframes {i}:{current_key} and {i + 1}:{next_key} are not monotonously increasing; skipping interpolation.")
|
|
continue
|
|
|
|
# Get current and next keyframes' positive and negative prompts (if any)
|
|
current_prompt = sorted_prompts[i][1]
|
|
next_prompt = sorted_prompts[i + 1][1]
|
|
|
|
# Calculate how much to shift the weight from current to next prompt at each frame.
|
|
weight_step = 1 / (next_key - current_key)
|
|
|
|
for f in range(max(current_key, 0), min(next_key, len(cur_prompt_series))):
|
|
next_weight = weight_step * (f - current_key)
|
|
current_weight = 1 - next_weight
|
|
|
|
# add the appropriate prompts and weights to their respective containers.
|
|
weight_series[f] = 0.0
|
|
cur_prompt_series[f] = str(current_prompt)
|
|
nxt_prompt_series[f] = str(next_prompt)
|
|
|
|
weight_series[f] += current_weight
|
|
|
|
current_key = next_key
|
|
next_key = max_frames
|
|
current_weight = 0.0
|
|
|
|
index_offset = 0
|
|
# Evaluate the current and next prompt's expressions
|
|
for i in range(start_frame, len(cur_prompt_series)):
|
|
cur_prompt_series[i] = prepare_batch_prompt(cur_prompt_series[i], max_frames, i, prompt_weight_1.iloc[i],
|
|
prompt_weight_2.iloc[i], prompt_weight_3.iloc[i], prompt_weight_4.iloc[i])
|
|
nxt_prompt_series[i] = prepare_batch_prompt(nxt_prompt_series[i], max_frames, i, prompt_weight_1.iloc[i],
|
|
prompt_weight_2.iloc[i], prompt_weight_3.iloc[i], prompt_weight_4.iloc[i])
|
|
if Is_print == True:
|
|
# Show the to/from prompts with evaluated expressions for transparency.
|
|
print("\n", "Max Frames: ", max_frames, "\n", "frame index: ", (start_frame + i), "\n", "Current Prompt: ",
|
|
cur_prompt_series[i], "\n", "Next Prompt: ", nxt_prompt_series[i], "\n", "Strength : ",
|
|
weight_series[i], "\n")
|
|
index_offset = index_offset + 1
|
|
|
|
|
|
|
|
# Output methods depending if the prompts are the same or if the current frame is a keyframe.
|
|
# if it is an in-between frame and the prompts differ, composable diffusion will be performed.
|
|
return (cur_prompt_series, nxt_prompt_series, weight_series)
|
|
|
|
|
|
def BatchPoolAnimConditioning(cur_prompt_series, nxt_prompt_series, weight_series, clip):
|
|
pooled_out = []
|
|
cond_out = []
|
|
|
|
for i in range(len(cur_prompt_series)):
|
|
tokens = clip.tokenize(str(cur_prompt_series[i]))
|
|
cond_to, pooled_to = clip.encode_from_tokens(tokens, return_pooled=True)
|
|
|
|
if i < len(nxt_prompt_series):
|
|
tokens = clip.tokenize(str(nxt_prompt_series[i]))
|
|
cond_from, pooled_from = clip.encode_from_tokens(tokens, return_pooled=True)
|
|
else:
|
|
cond_from, pooled_from = torch.zeros_like(cond_to), torch.zeros_like(pooled_to)
|
|
|
|
interpolated_conditioning = addWeighted([[cond_to, {"pooled_output": pooled_to}]],
|
|
[[cond_from, {"pooled_output": pooled_from}]],
|
|
weight_series[i])
|
|
|
|
interpolated_cond = interpolated_conditioning[0][0]
|
|
interpolated_pooled = interpolated_conditioning[0][1].get("pooled_output", pooled_from)
|
|
|
|
cond_out.append(interpolated_cond)
|
|
pooled_out.append(interpolated_pooled)
|
|
|
|
final_pooled_output = torch.cat(pooled_out, dim=0)
|
|
final_conditioning = torch.cat(cond_out, dim=0)
|
|
|
|
return [[final_conditioning, {"pooled_output": final_pooled_output}]]
|
|
|
|
|
|
|
|
|
|
|
|
def BatchGLIGENConditioning(cur_prompt_series, nxt_prompt_series, weight_series, clip):
|
|
pooled_out = []
|
|
cond_out = []
|
|
|
|
for i in range(len(cur_prompt_series)):
|
|
tokens = clip.tokenize(str(cur_prompt_series[i]))
|
|
cond_to, pooled_to = clip.encode_from_tokens(tokens, return_pooled=True)
|
|
|
|
tokens = clip.tokenize(str(nxt_prompt_series[i]))
|
|
cond_from, pooled_from = clip.encode_from_tokens(tokens, return_pooled=True)
|
|
|
|
interpolated_conditioning = addWeighted([[cond_to, {"pooled_output": pooled_to}]],
|
|
[[cond_from, {"pooled_output": pooled_from}]],
|
|
weight_series[i])
|
|
|
|
interpolated_cond = interpolated_conditioning[0][0]
|
|
interpolated_pooled = interpolated_conditioning[0][1].get("pooled_output", pooled_from)
|
|
|
|
pooled_out.append(interpolated_pooled)
|
|
cond_out.append(interpolated_cond)
|
|
|
|
final_pooled_output = torch.cat(pooled_out, dim=0)
|
|
final_conditioning = torch.cat(cond_out, dim=0)
|
|
|
|
return cond_out, pooled_out
|
|
|
|
def BatchPoolAnimConditioningSDXL(cur_prompt_series, nxt_prompt_series, weight_series, clip):
|
|
pooled_out = []
|
|
cond_out = []
|
|
|
|
for i in range(len(cur_prompt_series)):
|
|
interpolated_conditioning = addWeighted(cur_prompt_series[i],
|
|
nxt_prompt_series[i],
|
|
weight_series[i])
|
|
|
|
interpolated_cond = interpolated_conditioning[0][0]
|
|
interpolated_pooled = interpolated_conditioning[0][1].get("pooled_output")
|
|
|
|
pooled_out.append(interpolated_pooled)
|
|
cond_out.append(interpolated_cond)
|
|
|
|
final_pooled_output = torch.cat(pooled_out, dim=0)
|
|
final_conditioning = torch.cat(cond_out, dim=0)
|
|
|
|
return [[final_conditioning, {"pooled_output": final_pooled_output}]]
|
|
|
|
|
|
def BatchInterpolatePromptsSDXL(animation_promptsG, animation_promptsL, max_frames, clip, app_text_G,
|
|
app_text_L, pre_text_G, pre_text_L, pw_a, pw_b, pw_c, pw_d, width, height, crop_w,
|
|
crop_h, target_width, target_height, Is_print = False):
|
|
|
|
# parse the conditioning strength and determine in-betweens.
|
|
# Get prompts sorted by keyframe
|
|
max_f = max_frames # needed for numexpr even though it doesn't look like it's in use.
|
|
parsed_animation_promptsG = {}
|
|
parsed_animation_promptsL = {}
|
|
for key, value in animation_promptsG.items():
|
|
if check_is_number(key): # default case 0:(1 + t %5), 30:(5-t%2)
|
|
parsed_animation_promptsG[key] = value
|
|
else: # math on the left hand side case 0:(1 + t %5), maxKeyframes/2:(5-t%2)
|
|
parsed_animation_promptsG[int(numexpr.evaluate(key))] = value
|
|
|
|
sorted_prompts_G = sorted(parsed_animation_promptsG.items(), key=lambda item: int(item[0]))
|
|
|
|
for key, value in animation_promptsL.items():
|
|
if check_is_number(key): # default case 0:(1 + t %5), 30:(5-t%2)
|
|
parsed_animation_promptsL[key] = value
|
|
else: # math on the left hand side case 0:(1 + t %5), maxKeyframes/2:(5-t%2)
|
|
parsed_animation_promptsL[int(numexpr.evaluate(key))] = value
|
|
|
|
sorted_prompts_L = sorted(parsed_animation_promptsL.items(), key=lambda item: int(item[0]))
|
|
|
|
# Setup containers for interpolated prompts
|
|
cur_prompt_series_G = pd.Series([np.nan for a in range(max_frames)])
|
|
nxt_prompt_series_G = pd.Series([np.nan for a in range(max_frames)])
|
|
|
|
cur_prompt_series_L = pd.Series([np.nan for a in range(max_frames)])
|
|
nxt_prompt_series_L = pd.Series([np.nan for a in range(max_frames)])
|
|
|
|
# simple array for strength values
|
|
weight_series = [np.nan] * max_frames
|
|
|
|
# in case there is only one keyed promt, set all prompts to that prompt
|
|
if len(sorted_prompts_G) - 1 == 0:
|
|
for i in range(0, len(cur_prompt_series_G) - 1):
|
|
current_prompt_G = sorted_prompts_G[0][1]
|
|
cur_prompt_series_G[i] = str(pre_text_G) + " " + str(current_prompt_G) + " " + str(app_text_G)
|
|
nxt_prompt_series_G[i] = str(pre_text_G) + " " + str(current_prompt_G) + " " + str(app_text_G)
|
|
|
|
if len(sorted_prompts_L) - 1 == 0:
|
|
for i in range(0, len(cur_prompt_series_L) - 1):
|
|
current_prompt_L = sorted_prompts_L[0][1]
|
|
cur_prompt_series_L[i] = str(pre_text_L) + " " + str(current_prompt_L) + " " + str(app_text_L)
|
|
nxt_prompt_series_L[i] = str(pre_text_L) + " " + str(current_prompt_L) + " " + str(app_text_L)
|
|
|
|
# Initialized outside of loop for nan check
|
|
current_key = 0
|
|
next_key = 0
|
|
|
|
# For every keyframe prompt except the last
|
|
for i in range(0, len(sorted_prompts_G) - 1):
|
|
# Get current and next keyframe
|
|
current_key = int(sorted_prompts_G[i][0])
|
|
next_key = int(sorted_prompts_G[i + 1][0])
|
|
|
|
# Ensure there's no weird ordering issues or duplication in the animation prompts
|
|
# (unlikely because we sort above, and the json parser will strip dupes)
|
|
if current_key >= next_key:
|
|
print(
|
|
f"WARNING: Sequential prompt keyframes {i}:{current_key} and {i + 1}:{next_key} are not monotonously increasing; skipping interpolation.")
|
|
continue
|
|
|
|
# Get current and next keyframes' positive and negative prompts (if any)
|
|
current_prompt_G = sorted_prompts_G[i][1]
|
|
next_prompt_G = sorted_prompts_G[i + 1][1]
|
|
|
|
# Calculate how much to shift the weight from current to next prompt at each frame.
|
|
weight_step = 1 / (next_key - current_key)
|
|
|
|
for f in range(current_key, next_key):
|
|
next_weight = weight_step * (f - current_key)
|
|
current_weight = 1 - next_weight
|
|
|
|
# add the appropriate prompts and weights to their respective containers.
|
|
if f < max_frames:
|
|
cur_prompt_series_G[f] = ''
|
|
nxt_prompt_series_G[f] = ''
|
|
weight_series[f] = 0.0
|
|
|
|
cur_prompt_series_G[f] += (str(pre_text_G) + " " + str(current_prompt_G) + " " + str(app_text_G))
|
|
nxt_prompt_series_G[f] += (str(pre_text_G) + " " + str(next_prompt_G) + " " + str(app_text_G))
|
|
|
|
weight_series[f] += current_weight
|
|
|
|
current_key = next_key
|
|
next_key = max_frames
|
|
current_weight = 0.0
|
|
# second loop to catch any nan runoff
|
|
for f in range(current_key, next_key):
|
|
next_weight = weight_step * (f - current_key)
|
|
|
|
# add the appropriate prompts and weights to their respective containers.
|
|
cur_prompt_series_G[f] = ''
|
|
nxt_prompt_series_G[f] = ''
|
|
weight_series[f] = current_weight
|
|
|
|
cur_prompt_series_G[f] += (str(pre_text_G) + " " + str(current_prompt_G) + " " + str(app_text_G))
|
|
nxt_prompt_series_G[f] += (str(pre_text_G) + " " + str(next_prompt_G) + " " + str(app_text_G))
|
|
|
|
# Reset outside of loop for nan check
|
|
current_key = 0
|
|
next_key = 0
|
|
|
|
# For every keyframe prompt except the last
|
|
for i in range(0, len(sorted_prompts_L) - 1):
|
|
# Get current and next keyframe
|
|
current_key = int(sorted_prompts_L[i][0])
|
|
next_key = int(sorted_prompts_L[i + 1][0])
|
|
|
|
# Ensure there's no weird ordering issues or duplication in the animation prompts
|
|
# (unlikely because we sort above, and the json parser will strip dupes)
|
|
if current_key >= next_key:
|
|
print(
|
|
f"WARNING: Sequential prompt keyframes {i}:{current_key} and {i + 1}:{next_key} are not monotonously increasing; skipping interpolation.")
|
|
continue
|
|
|
|
# Get current and next keyframes' positive and negative prompts (if any)
|
|
current_prompt_L = sorted_prompts_L[i][1]
|
|
next_prompt_L = sorted_prompts_L[i + 1][1]
|
|
|
|
# Calculate how much to shift the weight from current to next prompt at each frame.
|
|
weight_step = 1 / (next_key - current_key)
|
|
|
|
for f in range(current_key, next_key):
|
|
next_weight = weight_step * (f - current_key)
|
|
current_weight = 1 - next_weight
|
|
|
|
# add the appropriate prompts and weights to their respective containers.
|
|
if f < max_frames:
|
|
cur_prompt_series_L[f] = ''
|
|
nxt_prompt_series_L[f] = ''
|
|
weight_series[f] = 0.0
|
|
|
|
cur_prompt_series_L[f] += (str(pre_text_L) + " " + str(current_prompt_L) + " " + str(app_text_L))
|
|
nxt_prompt_series_L[f] += (str(pre_text_L) + " " + str(next_prompt_L) + " " + str(app_text_L))
|
|
|
|
weight_series[f] += current_weight
|
|
|
|
current_key = next_key
|
|
next_key = max_frames
|
|
current_weight = 0.0
|
|
# second loop to catch any nan runoff
|
|
for f in range(current_key, next_key):
|
|
next_weight = weight_step * (f - current_key)
|
|
|
|
# add the appropriate prompts and weights to their respective containers.
|
|
cur_prompt_series_L[f] = ''
|
|
nxt_prompt_series_L[f] = ''
|
|
weight_series[f] = current_weight
|
|
|
|
cur_prompt_series_L[f] += (str(pre_text_L) + " " + str(current_prompt_L) + " " + str(app_text_L))
|
|
nxt_prompt_series_L[f] += (str(pre_text_L) + " " + str(next_prompt_L) + " " + str(app_text_L))
|
|
|
|
# Evaluate the current and next prompt's expressions
|
|
for i in range(0, max_frames):
|
|
cur_prompt_series_G[i] = prepare_batch_prompt(cur_prompt_series_G[i], max_frames, i,
|
|
pw_a, pw_b, pw_c, pw_d)
|
|
nxt_prompt_series_G[i] = prepare_batch_prompt(nxt_prompt_series_G[i], max_frames, i,
|
|
pw_a, pw_b, pw_c, pw_d)
|
|
cur_prompt_series_L[i] = prepare_batch_prompt(cur_prompt_series_L[i], max_frames, i,
|
|
pw_a, pw_b, pw_c, pw_d)
|
|
nxt_prompt_series_L[i] = prepare_batch_prompt(nxt_prompt_series_L[i], max_frames, i,
|
|
pw_a, pw_b, pw_c, pw_d)
|
|
|
|
current_conds = []
|
|
next_conds = []
|
|
for i in range(0, max_frames):
|
|
current_conds.append(SDXLencode(clip, width, height, crop_w, crop_h, target_width, target_height,
|
|
cur_prompt_series_G[i], cur_prompt_series_L[i]))
|
|
next_conds.append(SDXLencode(clip, width, height, crop_w, crop_h, target_width, target_height,
|
|
nxt_prompt_series_G[i], nxt_prompt_series_L[i]))
|
|
|
|
if Is_print == True:
|
|
# Show the to/from prompts with evaluated expressions for transparency.
|
|
for i in range(0, max_frames):
|
|
print("\n", "Max Frames: ", max_frames, "\n", "Current Prompt G: ", cur_prompt_series_G[i],
|
|
"\n", "Current Prompt L: ", cur_prompt_series_L[i], "\n", "Next Prompt G: ", nxt_prompt_series_G[i],
|
|
"\n", "Next Prompt L : ", nxt_prompt_series_L[i], "\n"), "\n", "Current weight: ", weight_series[i]
|
|
|
|
return BatchPoolAnimConditioningSDXL(current_conds, next_conds, weight_series, clip)
|