cutoff / scripts /cutoff.py
ddoc's picture
Upload 13 files
d2ae5b0
from collections import defaultdict
from typing import Union, List, Tuple
import numpy as np
import torch
from torch import Tensor, nn
import gradio as gr
from modules.processing import StableDiffusionProcessing
from modules import scripts
from scripts.cutofflib.sdhook import SDHook
from scripts.cutofflib.embedding import CLIP, generate_prompts, token_to_block
from scripts.cutofflib.utils import log, set_debug
from scripts.cutofflib.xyz import init_xyz
NAME = 'Cutoff'
PAD = '_</w>'
def check_neg(s: str, negative_prompt: str, all_negative_prompts: Union[List[str],None]):
if s == negative_prompt:
return True
if all_negative_prompts is not None:
return s in all_negative_prompts
return False
def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
# cf. https://memo.sugyan.com/entry/2022/09/09/230645
inputs_are_torch = False
input_device = v0.device
if not isinstance(v0, np.ndarray):
inputs_are_torch = True
v0 = v0.cpu().numpy()
v1 = v1.cpu().numpy()
dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
if np.abs(dot) > DOT_THRESHOLD:
v2 = (1 - t) * v0 + t * v1
else:
theta_0 = np.arccos(dot)
sin_theta_0 = np.sin(theta_0)
theta_t = theta_0 * t
sin_theta_t = np.sin(theta_t)
s0 = np.sin(theta_0 - theta_t) / sin_theta_0
s1 = sin_theta_t / sin_theta_0
v2 = s0 * v0 + s1 * v1
if inputs_are_torch:
v2 = torch.from_numpy(v2).to(input_device)
return v2
class Hook(SDHook):
def __init__(
self,
enabled: bool,
targets: List[str],
padding: Union[str,int],
weight: float,
disable_neg: bool,
strong: bool,
interpolate: str,
):
super().__init__(enabled)
self.targets = targets
self.padding = padding
self.weight = float(weight)
self.disable_neg = disable_neg
self.strong = strong
self.intp = interpolate
def interpolate(self, t1: Tensor, t2: Tensor, w):
if self.intp == 'lerp':
return torch.lerp(t1, t2, w)
else:
return slerp(w, t1, t2)
def hook_clip(self, p: StableDiffusionProcessing, clip: nn.Module):
skip = False
def hook(mod: nn.Module, inputs: Tuple[List[str]], output: Tensor):
nonlocal skip
if skip:
# called from <A> below
return
assert isinstance(mod, CLIP)
prompts, *rest = inputs
assert len(prompts) == output.shape[0]
# Check wether we are processing Negative prompt or not.
# I firmly believe there is no one who uses a negative prompt
# exactly identical to a prompt.
if self.disable_neg:
if all(check_neg(x, p.negative_prompt, p.all_negative_prompts) for x in prompts):
# Now we are processing Negative prompt and skip it.
return
output = output.clone()
for pidx, prompt in enumerate(prompts):
tt = token_to_block(mod, prompt)
cutoff = generate_prompts(mod, prompt, self.targets, self.padding)
switch_base = np.full_like(cutoff.sw, self.strong)
switch = np.full_like(cutoff.sw, True)
active = cutoff.active_blocks()
prompt_to_tokens = defaultdict(lambda: [])
for tidx, (token, block_index) in enumerate(tt):
if block_index in active:
sw = switch.copy()
sw[block_index] = False
prompt = cutoff.text(sw)
else:
prompt = cutoff.text(switch_base)
prompt_to_tokens[prompt].append((tidx, token))
#log(prompt_to_tokens)
ks = list(prompt_to_tokens.keys())
if len(ks) == 0:
# without any (negative) prompts
ks.append('')
try:
# <A>
skip = True
vs = mod(ks)
finally:
skip = False
tensor = output[pidx, :, :] # e.g. (77, 768)
for k, t in zip(ks, vs):
assert tensor.shape == t.shape
for tidx, token in prompt_to_tokens[k]:
log(f'{tidx:03} {token.token:<16} {k}')
tensor[tidx, :] = self.interpolate(tensor[tidx,:], t[tidx,:], self.weight)
return output
self.hook_layer(clip, hook)
class Script(scripts.Script):
def __init__(self):
super().__init__()
self.last_hooker: Union[SDHook,None] = None
def title(self):
return NAME
def show(self, is_img2img):
return scripts.AlwaysVisible
def ui(self, is_img2img):
with gr.Accordion(NAME, open=False):
enabled = gr.Checkbox(label='Enabled', value=False)
targets = gr.Textbox(label='Target tokens (comma separated)', placeholder='red, blue')
weight = gr.Slider(minimum=-1.0, maximum=2.0, step=0.01, value=0.5, label='Weight')
with gr.Accordion('Details', open=False):
disable_neg = gr.Checkbox(value=True, label='Disable for Negative prompt.')
strong = gr.Checkbox(value=False, label='Cutoff strongly.')
padding = gr.Textbox(label='Padding token (ID or single token)')
lerp = gr.Radio(choices=['Lerp', 'SLerp'], value='Lerp', label='Interpolation method')
debug = gr.Checkbox(value=False, label='Debug log')
debug.change(fn=set_debug, inputs=[debug], outputs=[])
return [
enabled,
targets,
weight,
disable_neg,
strong,
padding,
lerp,
debug,
]
def process(
self,
p: StableDiffusionProcessing,
enabled: bool,
targets_: str,
weight: Union[float,int],
disable_neg: bool,
strong: bool,
padding: Union[str,int],
intp: str,
debug: bool,
):
set_debug(debug)
if self.last_hooker is not None:
self.last_hooker.__exit__(None, None, None)
self.last_hooker = None
if not enabled:
return
if targets_ is None or len(targets_) == 0:
return
targets = [x.strip() for x in targets_.split(',')]
targets = [x for x in targets if len(x) != 0]
if len(targets) == 0:
return
if padding is None:
padding = PAD
elif isinstance(padding, str):
if len(padding) == 0:
padding = PAD
else:
try:
padding = int(padding)
except:
if not padding.endswith('</w>'):
padding += '</w>'
weight = float(weight)
intp = intp.lower()
self.last_hooker = Hook(
enabled=True,
targets=targets,
padding=padding,
weight=weight,
disable_neg=disable_neg,
strong=strong,
interpolate=intp,
)
self.last_hooker.setup(p)
self.last_hooker.__enter__()
p.extra_generation_params.update({
f'{NAME} enabled': enabled,
f'{NAME} targets': targets,
f'{NAME} padding': padding,
f'{NAME} weight': weight,
f'{NAME} disable_for_neg': disable_neg,
f'{NAME} strong': strong,
f'{NAME} interpolation': intp,
})
init_xyz(Script, NAME)