Update sd-webui-progressive-growing/scripts/progressive_growing_always.py
Browse files
sd-webui-progressive-growing/scripts/progressive_growing_always.py
CHANGED
|
@@ -1,139 +1,1024 @@
|
|
| 1 |
-
"""sd-webui-progressive-growing
|
|
|
|
| 2 |
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
|
|
|
| 8 |
|
| 9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
"""
|
| 11 |
|
| 12 |
from __future__ import annotations
|
| 13 |
|
|
|
|
|
|
|
| 14 |
import gradio as gr
|
|
|
|
|
|
|
| 15 |
|
| 16 |
from modules import scripts, sd_samplers, devices
|
| 17 |
from modules import processing as processing_mod
|
| 18 |
-
from modules.processing import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
#
|
| 22 |
-
#
|
| 23 |
-
# -----------------------------
|
| 24 |
|
| 25 |
-
|
| 26 |
-
|
| 27 |
|
| 28 |
-
|
| 29 |
-
import torch
|
| 30 |
|
| 31 |
-
|
|
|
|
| 32 |
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
max_scale = float(self.progressive_growing_max_scale)
|
| 36 |
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
v_int = int(v)
|
| 46 |
v_int = max(opt_f, v_int)
|
| 47 |
v_int = (v_int // opt_f) * opt_f
|
| 48 |
return max(opt_f, v_int)
|
| 49 |
|
| 50 |
-
#
|
| 51 |
-
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
-
# 3) ΠΠ°ΡΠ°Π»ΡΠ½ΡΠΉ Π»Π°ΡΠ΅Π½Ρ (noise)
|
| 55 |
x = create_random_tensors(
|
| 56 |
(opt_C, initial_height // opt_f, initial_width // opt_f),
|
| 57 |
-
seeds,
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
seed_resize_from_h=self.seed_resize_from_h,
|
| 61 |
-
seed_resize_from_w=self.seed_resize_from_w,
|
| 62 |
-
p=self
|
| 63 |
)
|
| 64 |
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
x,
|
| 69 |
-
|
| 70 |
-
unconditional_conditioning,
|
| 71 |
-
image_conditioning=self.txt2img_image_conditioning(x)
|
| 72 |
)
|
| 73 |
|
| 74 |
total_stages = len(resolution_steps)
|
| 75 |
|
| 76 |
-
# 5) ΠΡΠΎΠ³ΡΠ΅ΡΡΠΈΠ²Π½ΡΠΉ ΡΠΎΡΡ
|
| 77 |
for i in range(1, total_stages):
|
| 78 |
-
target_width =
|
| 79 |
-
target_height = _snap(self.height * resolution_steps[i])
|
| 80 |
|
| 81 |
-
# upscale latent
|
| 82 |
samples = torch.nn.functional.interpolate(
|
| 83 |
samples,
|
| 84 |
size=(target_height // opt_f, target_width // opt_f),
|
| 85 |
-
mode='bicubic',
|
| 86 |
-
align_corners=False
|
| 87 |
)
|
| 88 |
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
subseed_strength=subseed_strength,
|
| 98 |
-
seed_resize_from_h=self.seed_resize_from_h,
|
| 99 |
-
seed_resize_from_w=self.seed_resize_from_w,
|
| 100 |
-
p=self
|
| 101 |
)
|
| 102 |
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
)
|
|
|
|
| 118 |
|
| 119 |
return samples
|
| 120 |
|
| 121 |
|
| 122 |
-
|
| 123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
}
|
| 125 |
|
|
|
|
|
|
|
| 126 |
|
| 127 |
-
#
|
| 128 |
-
#
|
| 129 |
-
#
|
| 130 |
|
| 131 |
-
_PATCHED
|
| 132 |
_ORIG_SAMPLE = None
|
| 133 |
|
| 134 |
|
| 135 |
def _apply_patch_once() -> None:
|
| 136 |
-
"""Patch StableDiffusionProcessingTxt2Img.sample
|
| 137 |
|
| 138 |
global _PATCHED, _ORIG_SAMPLE
|
| 139 |
if _PATCHED:
|
|
@@ -143,83 +1028,148 @@ def _apply_patch_once() -> None:
|
|
| 143 |
if cls is None:
|
| 144 |
return
|
| 145 |
|
| 146 |
-
|
| 147 |
-
if getattr(cls, '_progressive_growing_ext_patched', False):
|
| 148 |
_PATCHED = True
|
| 149 |
return
|
| 150 |
|
| 151 |
_ORIG_SAMPLE = cls.sample
|
| 152 |
|
| 153 |
-
def _sample_wrapper(self,
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 163 |
|
| 164 |
-
|
| 165 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 166 |
|
| 167 |
-
|
| 168 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
_PATCHED = True
|
| 170 |
|
| 171 |
|
| 172 |
-
#
|
| 173 |
# Always-visible UI script
|
| 174 |
-
#
|
| 175 |
-
|
| 176 |
|
| 177 |
class ProgressiveGrowingAlwaysVisible(scripts.Script):
|
|
|
|
| 178 |
def title(self):
|
| 179 |
return "Progressive Growing"
|
| 180 |
|
| 181 |
def show(self, is_img2img):
|
| 182 |
-
# Only for txt2img, always visible
|
| 183 |
return scripts.AlwaysVisible if not is_img2img else False
|
| 184 |
|
| 185 |
def ui(self, is_img2img):
|
| 186 |
with gr.Accordion("Progressive Growing", open=False):
|
| 187 |
enabled = gr.Checkbox(value=False, label="Enable")
|
| 188 |
-
version = gr.Dropdown(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 194 |
|
| 195 |
gr.Markdown(
|
| 196 |
-
"
|
| 197 |
-
"-
|
| 198 |
-
"-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 199 |
)
|
| 200 |
|
| 201 |
-
return [enabled, version, min_scale, max_scale,
|
|
|
|
|
|
|
| 202 |
|
| 203 |
-
def process(self, p,
|
| 204 |
-
|
|
|
|
|
|
|
| 205 |
|
| 206 |
-
|
| 207 |
-
p.enable_progressive_growing = bool(enabled)
|
| 208 |
-
p.progressive_growing_version = str(version)
|
| 209 |
-
p.progressive_growing_min_scale = float(min_scale)
|
| 210 |
-
p.progressive_growing_max_scale = float(max_scale)
|
| 211 |
-
p.progressive_growing_steps = int(steps)
|
| 212 |
-
p.progressive_growing_refinement = bool(refinement)
|
| 213 |
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
pass
|
|
|
|
| 1 |
+
"""sd-webui-progressive-growing Β· extension-only rewrite
|
| 2 |
+
===========================================================
|
| 3 |
|
| 4 |
+
Architecture contract
|
| 5 |
+
---------------------
|
| 6 |
+
UI β process() β patch wrapper β version fn(p, plan, ...) β samples
|
| 7 |
+
β on any conflict/disable
|
| 8 |
+
_ORIG_SAMPLE(...)
|
| 9 |
|
| 10 |
+
Files
|
| 11 |
+
-----
|
| 12 |
+
This single file is self-contained for easy drop-in.
|
| 13 |
+
Split into lib_progressive/ when the project grows beyond ~600 lines.
|
| 14 |
|
| 15 |
+
Versions
|
| 16 |
+
--------
|
| 17 |
+
v1 (exact) β original code 1:1, kept as reference / regression baseline
|
| 18 |
+
v2 (safe) β validation, dedup stages, conflict guards, predictable output
|
| 19 |
+
v3 (fast) β minimal refinement, no VAE decode on intermediate stages
|
| 20 |
+
v4 (balanced) β refinement only on stages >= BALANCED_REFINE_THRESHOLD (linear scale, not area)
|
| 21 |
+
v5 (latent) β pure latent upscale, zero refinement
|
| 22 |
"""
|
| 23 |
|
| 24 |
from __future__ import annotations
|
| 25 |
|
| 26 |
+
import contextlib
|
| 27 |
+
import math
|
| 28 |
import gradio as gr
|
| 29 |
+
import numpy as np
|
| 30 |
+
import torch
|
| 31 |
|
| 32 |
from modules import scripts, sd_samplers, devices
|
| 33 |
from modules import processing as processing_mod
|
| 34 |
+
from modules.processing import (
|
| 35 |
+
create_random_tensors,
|
| 36 |
+
decode_latent_batch,
|
| 37 |
+
opt_C,
|
| 38 |
+
opt_f,
|
| 39 |
+
)
|
| 40 |
|
| 41 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 42 |
+
# Constants
|
| 43 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 44 |
|
| 45 |
+
BALANCED_REFINE_THRESHOLD = 0.70 # v4: only refine stages whose linear scale >= this
|
| 46 |
+
ADAPTIVE_REFINE_THRESHOLD = 0.60 # v6: threshold for long-plan (5+ stages) adaptive refinement
|
|
|
|
| 47 |
|
| 48 |
+
LATENT_INTERP_MODES = ["bicubic", "bilinear", "nearest", "area"]
|
| 49 |
+
LATENT_INTERP_DEFAULT = "bicubic" # best general-purpose default; nearest is fastest
|
| 50 |
|
| 51 |
+
MIN_STAGES_AFTER_DEDUP = 1 # if dedup collapses plan to 1 size, skip progressive entirely
|
|
|
|
| 52 |
|
| 53 |
+
AUTO_STAGE_JUMP_DEFAULT = 1.35 # target linear growth factor per stage
|
| 54 |
+
AUTO_STAGE_MAX_DEFAULT = 6 # hard cap on auto-computed stage count
|
| 55 |
|
| 56 |
+
REFINE_STEP_MODES = ["uniform", "late-heavy", "final-heavy"]
|
| 57 |
+
REFINE_STEP_DEFAULT = "uniform" # matches legacy behaviour; switch to late-heavy once comfortable
|
|
|
|
| 58 |
|
| 59 |
+
# ββ Sampler compatibility profiles βββββββββββββββββββββββββββββββββββββββββββ
|
| 60 |
+
#
|
| 61 |
+
# Progressive growing runs each stage sampler on a *smaller* latent than the
|
| 62 |
+
# final target. Two classes of problem arise:
|
| 63 |
+
#
|
| 64 |
+
# 1. Custom sigma schedulers (e.g. "CosineExponential blend") compute sigmas
|
| 65 |
+
# that are spatially shaped β [steps+1, H, W] β by reading p.width/p.height
|
| 66 |
+
# at sampler.sample() time. With stage dims still at final size the sigmas
|
| 67 |
+
# mismatch the stage latent β RuntimeError on the first sigma operation.
|
| 68 |
+
#
|
| 69 |
+
# 2. img2img contexts: p.init_latent / p.mask / p.nmask are full-size tensors.
|
| 70 |
+
# KDiffusionSampler.sample() copies them into model_wrap_cfg before
|
| 71 |
+
# launching the sampler loop, causing shape mismatches inside the loop.
|
| 72 |
+
#
|
| 73 |
+
# The "smea" profile wraps every sampler.sample() / sample_img2img() call with
|
| 74 |
+
# _stage_sampler_context(), which temporarily sets p.width/p.height to stage
|
| 75 |
+
# dims and rescales init_latent/mask/nmask. This directly mirrors the
|
| 76 |
+
# _Rescaler pattern already used inside sd-webui-smea itself.
|
| 77 |
+
#
|
| 78 |
+
# SMEA_SAMPLER_PATTERNS β families that need _stage_sampler_context.
|
| 79 |
+
# Each entry is a lowercase substring of p.sampler_name.
|
| 80 |
+
SMEA_SAMPLER_PATTERNS: list[str] = [
|
| 81 |
+
"euler dy", # Euler Dy, Euler Dy koishi-star
|
| 82 |
+
"euler smea", # Euler Smea, Euler Smea Dy, Euler Smea Max, all multi-*
|
| 83 |
+
"euler h max", # Euler h max a/b/c/β¦
|
| 84 |
+
"euler max", # Euler Max, Max1b β¦ Max4f (catches all Max* variants)
|
| 85 |
+
"kohaku_lonyu", # Kohaku_LoNyu_Yog
|
| 86 |
+
"tcd", # TCD / TCD Euler a
|
| 87 |
+
]
|
| 88 |
|
| 89 |
+
# INCOMPATIBLE_SAMPLER_PATTERNS β families with no viable compatibility path yet.
|
| 90 |
+
# Unlike SMEA, these are blocked entirely until a specific adapter is built.
|
| 91 |
+
# Empty for now; add entries here when a truly irreconcilable sampler is found.
|
| 92 |
+
INCOMPATIBLE_SAMPLER_PATTERNS: list[str] = []
|
| 93 |
|
| 94 |
+
|
| 95 |
+
def _get_sampler_profile(p) -> tuple[str, str]:
|
| 96 |
+
"""
|
| 97 |
+
Classify the active sampler into one of three profiles:
|
| 98 |
+
|
| 99 |
+
'standard' β normal k-diffusion; no special handling required.
|
| 100 |
+
'smea' β sd-webui-smea family; needs _stage_sampler_context.
|
| 101 |
+
'unsupported' β no compatibility path yet; progressive will be skipped.
|
| 102 |
+
|
| 103 |
+
Returns (profile_name, reason). reason is '' unless profile == 'unsupported'.
|
| 104 |
+
"""
|
| 105 |
+
name = (getattr(p, 'sampler_name', '') or '').lower()
|
| 106 |
+
|
| 107 |
+
for pattern in INCOMPATIBLE_SAMPLER_PATTERNS:
|
| 108 |
+
if pattern in name:
|
| 109 |
+
return 'unsupported', (
|
| 110 |
+
f'disabled β sampler "{p.sampler_name}" has no compatibility '
|
| 111 |
+
f'profile yet. Use a standard k-diffusion sampler or the SMEA family.'
|
| 112 |
+
)
|
| 113 |
+
|
| 114 |
+
for pattern in SMEA_SAMPLER_PATTERNS:
|
| 115 |
+
if pattern in name:
|
| 116 |
+
return 'smea', ''
|
| 117 |
+
|
| 118 |
+
return 'standard', ''
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
@contextlib.contextmanager
|
| 122 |
+
def _stage_sampler_context(p, stage_w: int, stage_h: int):
|
| 123 |
+
"""
|
| 124 |
+
Temporarily adapt p's size-dependent state to stage dimensions.
|
| 125 |
+
|
| 126 |
+
Three adaptations, all restored in finally:
|
| 127 |
+
|
| 128 |
+
1. p.width / p.height -> stage dims.
|
| 129 |
+
|
| 130 |
+
2. p.init_latent / p.mask / p.nmask -> rescaled to stage latent dims.
|
| 131 |
+
KDiffusionSampler.sample() copies these into model_wrap_cfg before
|
| 132 |
+
the sampler loop.
|
| 133 |
+
|
| 134 |
+
3. p.sampler_noise_scheduler_override -> wrapped to resize spatial sigmas.
|
| 135 |
+
SMEA's process() hook sets this override with full-size dims captured
|
| 136 |
+
in a closure. KDiffusionSampler.sample() calls the override DIRECTLY
|
| 137 |
+
(bypassing get_sigmas entirely) when it is set, so any spatial sigma
|
| 138 |
+
tensor [T,H,W] or [T,C,H,W] it returns will have full-size H/W dims.
|
| 139 |
+
We wrap the override here so its output is resized to stage_hw before
|
| 140 |
+
the sampler loop sees it.
|
| 141 |
+
"""
|
| 142 |
+
import torch.nn.functional as _F
|
| 143 |
+
|
| 144 |
+
stage_hw = (stage_h // opt_f, stage_w // opt_f)
|
| 145 |
+
|
| 146 |
+
# save
|
| 147 |
+
orig_w, orig_h = p.width, p.height
|
| 148 |
+
orig_init = getattr(p, 'init_latent', None)
|
| 149 |
+
orig_mask = getattr(p, 'mask', None)
|
| 150 |
+
orig_nmask = getattr(p, 'nmask', None)
|
| 151 |
+
orig_override = getattr(p, 'sampler_noise_scheduler_override', None)
|
| 152 |
+
|
| 153 |
+
# adapt dims
|
| 154 |
+
p.width, p.height = stage_w, stage_h
|
| 155 |
+
|
| 156 |
+
if orig_init is not None:
|
| 157 |
+
p.init_latent = _F.interpolate(orig_init, size=stage_hw, mode='nearest-exact')
|
| 158 |
+
if orig_mask is not None:
|
| 159 |
+
p.mask = _F.interpolate(
|
| 160 |
+
orig_mask.unsqueeze(0), size=stage_hw, mode='nearest-exact'
|
| 161 |
+
).squeeze(0)
|
| 162 |
+
if orig_nmask is not None:
|
| 163 |
+
p.nmask = _F.interpolate(
|
| 164 |
+
orig_nmask.unsqueeze(0), size=stage_hw, mode='nearest-exact'
|
| 165 |
+
).squeeze(0)
|
| 166 |
+
|
| 167 |
+
# wrap noise scheduler override to resize spatial sigmas to stage dims
|
| 168 |
+
if orig_override is not None:
|
| 169 |
+
def _stage_override(steps, _orig=orig_override, _hw=stage_hw):
|
| 170 |
+
sigs = _orig(steps)
|
| 171 |
+
if not isinstance(sigs, torch.Tensor) or sigs.ndim < 3:
|
| 172 |
+
return sigs
|
| 173 |
+
try:
|
| 174 |
+
if sigs.ndim == 3: # [T, H, W]
|
| 175 |
+
sigs = _F.interpolate(
|
| 176 |
+
sigs.unsqueeze(1), size=_hw, mode='nearest-exact'
|
| 177 |
+
).squeeze(1)
|
| 178 |
+
elif sigs.ndim == 4: # [T, C, H, W]
|
| 179 |
+
sigs = _F.interpolate(sigs, size=_hw, mode='nearest-exact')
|
| 180 |
+
except Exception:
|
| 181 |
+
pass
|
| 182 |
+
return sigs
|
| 183 |
+
p.sampler_noise_scheduler_override = _stage_override
|
| 184 |
+
|
| 185 |
+
try:
|
| 186 |
+
yield
|
| 187 |
+
finally:
|
| 188 |
+
p.width, p.height = orig_w, orig_h
|
| 189 |
+
if orig_init is not None: p.init_latent = orig_init
|
| 190 |
+
if orig_mask is not None: p.mask = orig_mask
|
| 191 |
+
if orig_nmask is not None: p.nmask = orig_nmask
|
| 192 |
+
p.sampler_noise_scheduler_override = orig_override
|
| 193 |
+
|
| 194 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββοΏ½οΏ½ββββββββββ
|
| 195 |
+
# Stage planner
|
| 196 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 197 |
+
|
| 198 |
+
def _snap(v: float, factor: int = opt_f) -> int:
|
| 199 |
+
"""Round v down to nearest multiple of factor (min = factor)."""
|
| 200 |
+
v_int = max(factor, int(v))
|
| 201 |
+
return max(factor, (v_int // factor) * factor)
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
class StagePlan:
|
| 205 |
+
"""Validated, deduplicated list of (width, height) latent sizes."""
|
| 206 |
+
|
| 207 |
+
def __init__(self, sizes: list[tuple[int, int]], final_w: int, final_h: int):
|
| 208 |
+
self.sizes = sizes # [(w, h), ...], last entry == (final_w, final_h)
|
| 209 |
+
self.final_w = final_w
|
| 210 |
+
self.final_h = final_h
|
| 211 |
+
self.n_stages = len(sizes)
|
| 212 |
+
|
| 213 |
+
def __repr__(self) -> str:
|
| 214 |
+
parts = [f"{w}Γ{h}" for w, h in self.sizes]
|
| 215 |
+
return " β ".join(parts)
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
def _auto_n_steps(min_s: float, max_s: float,
|
| 219 |
+
target_jump: float, max_stages: int) -> int:
|
| 220 |
+
"""
|
| 221 |
+
Compute stage count so each step grows the linear scale by ~target_jump.
|
| 222 |
+
|
| 223 |
+
Formula: n = 1 + ceil( log(max_s / min_s) / log(target_jump) )
|
| 224 |
+
Clamped to [2, max_stages].
|
| 225 |
+
|
| 226 |
+
Examples with target_jump=1.35:
|
| 227 |
+
0.25 β 1.0 : β 6 stages
|
| 228 |
+
0.50 β 1.0 : β 3 stages
|
| 229 |
+
0.75 β 1.0 : β 2 stages
|
| 230 |
+
"""
|
| 231 |
+
growth = max_s / max(min_s, 1e-6)
|
| 232 |
+
n_steps = 1 + math.ceil(math.log(max(growth, 1.0)) / math.log(max(target_jump, 1.001)))
|
| 233 |
+
return max(2, min(max_stages, n_steps))
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
def build_plan(p) -> StagePlan:
|
| 237 |
+
"""
|
| 238 |
+
Build a deduplicated stage plan from p's progressive-growing attributes.
|
| 239 |
+
|
| 240 |
+
When p.progressive_growing_auto_stages is True the stage count is computed
|
| 241 |
+
automatically from min/max scale and p.progressive_growing_auto_jump
|
| 242 |
+
(target linear growth per step), capped at p.progressive_growing_auto_max.
|
| 243 |
+
The manual Stages slider is ignored in auto mode.
|
| 244 |
+
|
| 245 |
+
Guarantees:
|
| 246 |
+
- min_scale <= max_scale (swapped silently)
|
| 247 |
+
- last stage == (p.width, p.height) snapped to opt_f
|
| 248 |
+
- duplicate sizes removed
|
| 249 |
+
- if only one unique size remains: plan.n_stages == 1
|
| 250 |
+
(caller should fall back to normal sampling)
|
| 251 |
+
"""
|
| 252 |
+
min_s = float(getattr(p, 'progressive_growing_min_scale', 0.25))
|
| 253 |
+
max_s = float(getattr(p, 'progressive_growing_max_scale', 1.0))
|
| 254 |
+
|
| 255 |
+
if min_s > max_s:
|
| 256 |
+
min_s, max_s = max_s, min_s
|
| 257 |
+
|
| 258 |
+
auto = bool(getattr(p, 'progressive_growing_auto_stages', False))
|
| 259 |
+
if auto:
|
| 260 |
+
target_jump = float(getattr(p, 'progressive_growing_auto_jump', AUTO_STAGE_JUMP_DEFAULT))
|
| 261 |
+
max_stages = int(getattr(p, 'progressive_growing_auto_max', AUTO_STAGE_MAX_DEFAULT))
|
| 262 |
+
n_steps = _auto_n_steps(min_s, max_s, target_jump, max_stages)
|
| 263 |
+
else:
|
| 264 |
+
n_steps = max(2, int(getattr(p, 'progressive_growing_steps', 4)))
|
| 265 |
+
|
| 266 |
+
scales = np.linspace(min_s, max_s, n_steps)
|
| 267 |
+
final_w = _snap(p.width)
|
| 268 |
+
final_h = _snap(p.height)
|
| 269 |
+
|
| 270 |
+
raw_sizes: list[tuple[int, int]] = []
|
| 271 |
+
for s in scales:
|
| 272 |
+
raw_sizes.append((_snap(p.width * s), _snap(p.height * s)))
|
| 273 |
+
|
| 274 |
+
# deduplicate while preserving order
|
| 275 |
+
seen: set[tuple[int, int]] = set()
|
| 276 |
+
unique: list[tuple[int, int]] = []
|
| 277 |
+
for sz in raw_sizes:
|
| 278 |
+
if sz not in seen:
|
| 279 |
+
seen.add(sz)
|
| 280 |
+
unique.append(sz)
|
| 281 |
+
|
| 282 |
+
# force the last entry to be the true final size
|
| 283 |
+
if not unique or unique[-1] != (final_w, final_h):
|
| 284 |
+
if (final_w, final_h) in seen:
|
| 285 |
+
unique = [sz for sz in unique if sz != (final_w, final_h)]
|
| 286 |
+
unique.append((final_w, final_h))
|
| 287 |
+
|
| 288 |
+
plan = StagePlan(unique, final_w, final_h)
|
| 289 |
+
plan.auto_mode = auto # carry metadata for infotext
|
| 290 |
+
plan.requested = n_steps # stages before dedup
|
| 291 |
+
return plan
|
| 292 |
+
|
| 293 |
+
|
| 294 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 295 |
+
# Conflict / capability guards
|
| 296 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 297 |
+
|
| 298 |
+
def _should_use_progressive(p):
|
| 299 |
+
"""
|
| 300 |
+
Return (True, '', plan) if progressive should run, else (False, reason, None).
|
| 301 |
+
Building the plan here avoids a second build_plan() call inside the version fn.
|
| 302 |
+
Reasons appear in infotext via extra_generation_params.
|
| 303 |
+
"""
|
| 304 |
+
if not getattr(p, 'enable_progressive_growing', False):
|
| 305 |
+
return False, '', None
|
| 306 |
+
|
| 307 |
+
if getattr(p, 'enable_hr', False):
|
| 308 |
+
return False, 'disabled β incompatible with Hires. fix', None
|
| 309 |
+
|
| 310 |
+
profile, reason = _get_sampler_profile(p)
|
| 311 |
+
if profile == 'unsupported':
|
| 312 |
+
return False, reason, None
|
| 313 |
+
|
| 314 |
+
plan = build_plan(p)
|
| 315 |
+
if plan.n_stages <= MIN_STAGES_AFTER_DEDUP:
|
| 316 |
+
return False, f'disabled β all stages collapsed to one size ({plan.final_w}Γ{plan.final_h})', None
|
| 317 |
+
|
| 318 |
+
return True, '', plan
|
| 319 |
+
|
| 320 |
+
|
| 321 |
+
def _write_params(p, plan: StagePlan, refine_policy: str, refine_step_policy: str) -> None:
|
| 322 |
+
"""
|
| 323 |
+
Write run parameters into infotext.
|
| 324 |
+
|
| 325 |
+
refine_policy β effective policy for *which* stages refine
|
| 326 |
+
e.g. 'all stages', 'final stage only', 'none'
|
| 327 |
+
refine_step_policy β effective step budget policy actually used
|
| 328 |
+
e.g. 'uniform', 'late-heavy (v1 fixed)', 'none'
|
| 329 |
+
Callers must pass the *actual* behaviour, not the
|
| 330 |
+
UI dropdown value, so v1 / v5 / checkbox-off are honest.
|
| 331 |
+
|
| 332 |
+
If the refinement checkbox is off, both refine fields collapse to 'off (checkbox)'.
|
| 333 |
+
"""
|
| 334 |
+
refinement_on = getattr(p, 'progressive_growing_refinement', True)
|
| 335 |
+
try:
|
| 336 |
+
ep = p.extra_generation_params
|
| 337 |
+
ep['PG'] = getattr(p, 'progressive_growing_version', 'v2 (safe)')
|
| 338 |
+
ep['PG plan'] = str(plan)
|
| 339 |
+
ep['PG stages'] = plan.n_stages
|
| 340 |
+
ep['PG refine'] = refine_policy if refinement_on else 'off (checkbox)'
|
| 341 |
+
ep['PG refine steps'] = refine_step_policy if refinement_on else 'off (checkbox)'
|
| 342 |
+
ep['PG interp'] = getattr(p, 'progressive_growing_interp_mode', LATENT_INTERP_DEFAULT)
|
| 343 |
+
# auto-stage metadata
|
| 344 |
+
if getattr(plan, 'auto_mode', False):
|
| 345 |
+
ep['PG stages mode'] = 'auto'
|
| 346 |
+
ep['PG target jump'] = getattr(p, 'progressive_growing_auto_jump', AUTO_STAGE_JUMP_DEFAULT)
|
| 347 |
+
ep['PG requested'] = getattr(plan, 'requested', plan.n_stages)
|
| 348 |
+
ep['PG actual'] = plan.n_stages
|
| 349 |
+
else:
|
| 350 |
+
ep['PG stages mode'] = 'manual'
|
| 351 |
+
except Exception:
|
| 352 |
+
pass
|
| 353 |
+
|
| 354 |
+
|
| 355 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 356 |
+
# Shared helpers
|
| 357 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 358 |
+
|
| 359 |
+
def _initial_latent(p, w: int, h: int, seeds, subseeds, subseed_strength):
|
| 360 |
+
return create_random_tensors(
|
| 361 |
+
(opt_C, h // opt_f, w // opt_f),
|
| 362 |
+
seeds,
|
| 363 |
+
subseeds=subseeds,
|
| 364 |
+
subseed_strength=subseed_strength,
|
| 365 |
+
seed_resize_from_h=p.seed_resize_from_h,
|
| 366 |
+
seed_resize_from_w=p.seed_resize_from_w,
|
| 367 |
+
p=p,
|
| 368 |
+
)
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
def _upscale_latent(samples: torch.Tensor, w: int, h: int, interp_mode: str = LATENT_INTERP_DEFAULT) -> torch.Tensor:
|
| 372 |
+
"""
|
| 373 |
+
Upscale latent to (h // opt_f, w // opt_f) using the requested interpolation.
|
| 374 |
+
|
| 375 |
+
Supported modes mirror torch.nn.functional.interpolate:
|
| 376 |
+
bicubic β smooth, best for photographic content (default)
|
| 377 |
+
bilinear β slightly faster, softer result
|
| 378 |
+
nearest β fastest, hard edges; useful for pixel art / very structured images
|
| 379 |
+
area β anti-aliased average pooling; suited for large downscales if
|
| 380 |
+
you manually set min_scale > max_scale outside build_plan()
|
| 381 |
+
(build_plan() itself always swaps them, so area rarely fires
|
| 382 |
+
in normal use β kept for completeness)
|
| 383 |
+
|
| 384 |
+
align_corners=False for bicubic/bilinear matches PyTorch convention and
|
| 385 |
+
the behaviour of v1 (exact), avoiding edge-pixel drift on upscale.
|
| 386 |
+
"""
|
| 387 |
+
align = interp_mode in ("bicubic", "bilinear")
|
| 388 |
+
return torch.nn.functional.interpolate(
|
| 389 |
+
samples,
|
| 390 |
+
size=(h // opt_f, w // opt_f),
|
| 391 |
+
mode=interp_mode,
|
| 392 |
+
align_corners=False if align else None,
|
| 393 |
+
)
|
| 394 |
+
|
| 395 |
+
|
| 396 |
+
def _make_noise(p, shape, seeds, subseeds, subseed_strength):
|
| 397 |
+
return create_random_tensors(
|
| 398 |
+
shape,
|
| 399 |
+
seeds,
|
| 400 |
+
subseeds=subseeds,
|
| 401 |
+
subseed_strength=subseed_strength,
|
| 402 |
+
seed_resize_from_h=p.seed_resize_from_h,
|
| 403 |
+
seed_resize_from_w=p.seed_resize_from_w,
|
| 404 |
+
p=p,
|
| 405 |
+
)
|
| 406 |
+
|
| 407 |
+
|
| 408 |
+
def _call_script_hooks(p, samples) -> None:
|
| 409 |
+
"""
|
| 410 |
+
Fire process_before_every_sampling on all registered scripts, if available.
|
| 411 |
+
|
| 412 |
+
In stock A1111 this hook lets ControlNet, ADetailer, and other extensions
|
| 413 |
+
inject per-sampling adjustments (e.g. attention maps, masks). Calling it
|
| 414 |
+
before every sampler.sample / sampler.sample_img2img invocation keeps
|
| 415 |
+
progressive growing compatible with those extensions.
|
| 416 |
+
|
| 417 |
+
Guarded so the extension degrades gracefully on forks that lack the hook.
|
| 418 |
+
"""
|
| 419 |
+
scripts_obj = getattr(p, 'scripts', None)
|
| 420 |
+
if scripts_obj is None:
|
| 421 |
+
return
|
| 422 |
+
hook = getattr(scripts_obj, 'process_before_every_sampling', None)
|
| 423 |
+
if hook is None:
|
| 424 |
+
return
|
| 425 |
+
try:
|
| 426 |
+
hook(p, samples)
|
| 427 |
+
except Exception:
|
| 428 |
+
pass
|
| 429 |
+
|
| 430 |
+
|
| 431 |
+
def _get_refine_steps(p, refine_idx: int, n_refine: int, mode: str) -> int:
|
| 432 |
+
"""
|
| 433 |
+
Compute the step budget for a single refinement pass.
|
| 434 |
+
|
| 435 |
+
Parameters
|
| 436 |
+
----------
|
| 437 |
+
p : processing object, provides p.steps (total generation steps)
|
| 438 |
+
refine_idx : 0-based index among refine passes that will actually run
|
| 439 |
+
(0 = earliest/smallest stage being refined)
|
| 440 |
+
n_refine : total number of refine passes that will run this generation
|
| 441 |
+
mode : one of REFINE_STEP_MODES
|
| 442 |
+
|
| 443 |
+
Modes
|
| 444 |
+
-----
|
| 445 |
+
uniform current behaviour: p.steps // total_refine_stages, equal for all
|
| 446 |
+
late-heavy weights grow towards the final pass. Weights for n passes:
|
| 447 |
+
n=1 β [1.0]
|
| 448 |
+
n=2 β [0.4, 0.6]
|
| 449 |
+
n=3 β [0.2, 0.3, 0.5]
|
| 450 |
+
n=4+ β geometric series r=1.5 normalised to 1.0
|
| 451 |
+
final-heavy almost all budget to the last pass:
|
| 452 |
+
non-final passes each get floor(p.steps * 0.08)
|
| 453 |
+
final pass gets whatever remains, minimum 1
|
| 454 |
+
"""
|
| 455 |
+
total = max(1, p.steps)
|
| 456 |
+
n = max(1, n_refine)
|
| 457 |
+
idx = max(0, min(refine_idx, n - 1))
|
| 458 |
+
|
| 459 |
+
if mode == 'final-heavy':
|
| 460 |
+
if idx < n - 1:
|
| 461 |
+
return max(1, int(total * 0.08))
|
| 462 |
+
non_final_total = max(1, int(total * 0.08)) * (n - 1)
|
| 463 |
+
return max(1, total - non_final_total)
|
| 464 |
+
|
| 465 |
+
if mode == 'late-heavy':
|
| 466 |
+
if n == 1:
|
| 467 |
+
weights = [1.0]
|
| 468 |
+
elif n == 2:
|
| 469 |
+
weights = [0.4, 0.6]
|
| 470 |
+
elif n == 3:
|
| 471 |
+
weights = [0.2, 0.3, 0.5]
|
| 472 |
+
else:
|
| 473 |
+
# geometric series with ratio 1.5
|
| 474 |
+
r = 1.5
|
| 475 |
+
raw = [r ** i for i in range(n)]
|
| 476 |
+
s = sum(raw)
|
| 477 |
+
weights = [v / s for v in raw]
|
| 478 |
+
steps = max(1, round(total * weights[idx]))
|
| 479 |
+
return steps
|
| 480 |
+
|
| 481 |
+
# uniform (default / legacy)
|
| 482 |
+
return max(1, total // n)
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
def _resize_model_wrap_cfg(p, stage_w: int, stage_h: int) -> None:
|
| 486 |
+
"""
|
| 487 |
+
Resize spatial state inside p.sampler.model_wrap_cfg to match stage dims.
|
| 488 |
+
|
| 489 |
+
KDiffusionSampler.sample() calls get_scalings() and copies model_wrap_cfg.*
|
| 490 |
+
into the sampler loop *before* the first step. Any full-size spatial tensor
|
| 491 |
+
still living there will mismatch the stage-sized latent `x` on the first
|
| 492 |
+
sigma operation (e.g. `x - eps * (sigma_hat**2 - sigmas[i]**2)**0.5`).
|
| 493 |
+
|
| 494 |
+
This must be called *after* create_sampler() and *inside* _stage_sampler_context
|
| 495 |
+
so the resize is coherent with the p.width/p.height override.
|
| 496 |
+
|
| 497 |
+
Only non-None tensors are touched; all errors are silently swallowed so a
|
| 498 |
+
fork with a different model_wrap_cfg structure cannot crash the whole pass.
|
| 499 |
+
"""
|
| 500 |
+
import torch.nn.functional as _F
|
| 501 |
+
stage_hw = (stage_h // opt_f, stage_w // opt_f)
|
| 502 |
+
mw = getattr(getattr(p, 'sampler', None), 'model_wrap_cfg', None)
|
| 503 |
+
if mw is None:
|
| 504 |
+
return
|
| 505 |
+
for name in ('init_latent', 'mask', 'nmask'):
|
| 506 |
+
t = getattr(mw, name, None)
|
| 507 |
+
if t is None:
|
| 508 |
+
continue
|
| 509 |
+
try:
|
| 510 |
+
if name == 'init_latent':
|
| 511 |
+
setattr(mw, name, _F.interpolate(t, size=stage_hw, mode='nearest-exact'))
|
| 512 |
+
else:
|
| 513 |
+
setattr(mw, name,
|
| 514 |
+
_F.interpolate(t.unsqueeze(0), size=stage_hw,
|
| 515 |
+
mode='nearest-exact').squeeze(0))
|
| 516 |
+
except Exception:
|
| 517 |
+
pass
|
| 518 |
+
|
| 519 |
+
|
| 520 |
+
@contextlib.contextmanager
|
| 521 |
+
def _force_safe_sigmas_for_smea(p):
|
| 522 |
+
"""
|
| 523 |
+
For SMEA-family progressive stage passes, temporarily force plain 1D sigmas.
|
| 524 |
+
|
| 525 |
+
SMEA's custom spatial sigma schedulers (e.g. CosineExponential blend,
|
| 526 |
+
Karras Exponential v3) produce [T, H, W] or [T, C, H, W] tensors keyed
|
| 527 |
+
to full-size H/W. Wrapping p.sampler_noise_scheduler_override (as done in
|
| 528 |
+
_stage_sampler_context) is not always enough because some scheduler paths
|
| 529 |
+
bypass the override entirely and cache sigmas elsewhere (e.g.
|
| 530 |
+
sampler_extra_args['sigmas'] populated during create_sampler).
|
| 531 |
+
|
| 532 |
+
The safest fix for PG + SMEA: drop the override entirely for the duration
|
| 533 |
+
of this stage pass so KDiffusionSampler falls back to its own get_sigmas()
|
| 534 |
+
with the already stage-sized p.width/p.height. The SMEA sampler function
|
| 535 |
+
(sample_euler_max*) receives plain 1D sigmas and works fine.
|
| 536 |
+
|
| 537 |
+
Restores original override unconditionally in finally.
|
| 538 |
+
"""
|
| 539 |
+
orig_override = getattr(p, 'sampler_noise_scheduler_override', None)
|
| 540 |
+
orig_scheduler = getattr(p, 'scheduler', None)
|
| 541 |
+
try:
|
| 542 |
+
p.sampler_noise_scheduler_override = None
|
| 543 |
+
try:
|
| 544 |
+
p.scheduler = 'Use sampler default'
|
| 545 |
+
except Exception:
|
| 546 |
+
pass
|
| 547 |
+
yield
|
| 548 |
+
finally:
|
| 549 |
+
p.sampler_noise_scheduler_override = orig_override
|
| 550 |
+
try:
|
| 551 |
+
if orig_scheduler is not None:
|
| 552 |
+
p.scheduler = orig_scheduler
|
| 553 |
+
except Exception:
|
| 554 |
+
pass
|
| 555 |
+
|
| 556 |
+
|
| 557 |
+
|
| 558 |
+
def _patch_sampler_func_sigmas_to_1d(p) -> None:
|
| 559 |
+
"""
|
| 560 |
+
Wrap p.sampler.func so that the `sigmas` positional argument passed to
|
| 561 |
+
the SMEA sampler function is always 1D (scalar per step).
|
| 562 |
+
|
| 563 |
+
KDiffusionSampler.sample() computes sigmas (via override OR get_sigmas OR
|
| 564 |
+
sd_schedulers.apply_scheduler) and captures them in a lambda closure:
|
| 565 |
+
|
| 566 |
+
lambda: self.func(model_wrap_cfg, x, sigmas, extra_args=..., ...)
|
| 567 |
+
|
| 568 |
+
Patching get_sigmas or override cannot intercept all paths β the sigmas
|
| 569 |
+
variable lives in sample()'s local scope and is passed positionally.
|
| 570 |
+
Wrapping self.func on the instance is the only guaranteed intercept point.
|
| 571 |
+
|
| 572 |
+
Spatial sigma shapes:
|
| 573 |
+
[T, H, W] -> [T] (mean over H, W)
|
| 574 |
+
[T, C, H, W] -> [T] (mean over C, H, W)
|
| 575 |
+
|
| 576 |
+
SMEA / Euler Max* samplers only use sigmas[i] as a per-step scalar, so
|
| 577 |
+
collapsing spatial dims does not affect sampling quality.
|
| 578 |
+
"""
|
| 579 |
+
sampler = getattr(p, 'sampler', None)
|
| 580 |
+
if sampler is None:
|
| 581 |
+
return
|
| 582 |
+
orig_func = getattr(sampler, 'func', None)
|
| 583 |
+
if orig_func is None:
|
| 584 |
+
return
|
| 585 |
+
|
| 586 |
+
def _collapse(sigs):
|
| 587 |
+
if not isinstance(sigs, torch.Tensor) or sigs.ndim < 2:
|
| 588 |
+
return sigs
|
| 589 |
+
try:
|
| 590 |
+
dims = tuple(range(1, sigs.ndim))
|
| 591 |
+
return sigs.mean(dim=dims)
|
| 592 |
+
except Exception:
|
| 593 |
+
return sigs
|
| 594 |
+
|
| 595 |
+
def _safe_func(model, x, sigmas, *args, **kwargs):
|
| 596 |
+
return orig_func(model, x, _collapse(sigmas), *args, **kwargs)
|
| 597 |
+
|
| 598 |
+
sampler.func = _safe_func
|
| 599 |
+
|
| 600 |
+
|
| 601 |
+
def _stage_sample_txt2img(p, x, conditioning, unconditional_conditioning,
|
| 602 |
+
image_cond, stage_w: int, stage_h: int,
|
| 603 |
+
sampler_profile: str) -> torch.Tensor:
|
| 604 |
+
"""
|
| 605 |
+
Run a single txt2img sampler pass on a stage-sized latent.
|
| 606 |
+
|
| 607 |
+
For 'smea' profile:
|
| 608 |
+
- _stage_sampler_context sets stage-sized p.width/p.height, rescales
|
| 609 |
+
init_latent/mask/nmask, and wraps sampler_noise_scheduler_override.
|
| 610 |
+
- _force_safe_sigmas_for_smea then sets override=None so
|
| 611 |
+
KDiffusionSampler falls back to plain 1D get_sigmas(). This is the
|
| 612 |
+
definitive fix for spatial sigma mismatches in Euler Max* samplers.
|
| 613 |
+
- sampler is created fresh inside the context so all size-dependent
|
| 614 |
+
state (sigma schedules, cached spatial tensors) matches stage dims.
|
| 615 |
+
- sampler_extra_args['sigmas'] is evicted in case create_sampler
|
| 616 |
+
cached a full-size spatial schedule during construction.
|
| 617 |
+
|
| 618 |
+
For 'standard' profile: uses p.sampler as-is (already created in wrapper).
|
| 619 |
+
"""
|
| 620 |
+
if sampler_profile == 'smea':
|
| 621 |
+
with _stage_sampler_context(p, stage_w, stage_h):
|
| 622 |
+
with _force_safe_sigmas_for_smea(p):
|
| 623 |
+
p.sampler = sd_samplers.create_sampler(p.sampler_name, p.sd_model)
|
| 624 |
+
_patch_sampler_func_sigmas_to_1d(p)
|
| 625 |
+
_resize_model_wrap_cfg(p, stage_w, stage_h)
|
| 626 |
+
_call_script_hooks(p, x)
|
| 627 |
+
return p.sampler.sample(
|
| 628 |
+
p, x, conditioning, unconditional_conditioning,
|
| 629 |
+
image_conditioning=image_cond,
|
| 630 |
+
)
|
| 631 |
+
else:
|
| 632 |
+
_call_script_hooks(p, x)
|
| 633 |
+
return p.sampler.sample(
|
| 634 |
+
p, x, conditioning, unconditional_conditioning,
|
| 635 |
+
image_conditioning=image_cond,
|
| 636 |
+
)
|
| 637 |
+
|
| 638 |
+
|
| 639 |
+
def _stage_sample_img2img(p, samples, noise, conditioning, unconditional_conditioning,
|
| 640 |
+
stage_w: int, stage_h: int,
|
| 641 |
+
steps: int, sampler_profile: str) -> torch.Tensor:
|
| 642 |
+
"""
|
| 643 |
+
Run a single img2img sampler pass on a stage-sized latent.
|
| 644 |
+
|
| 645 |
+
Same contract as _stage_sample_txt2img: for 'smea' profile a fresh sampler
|
| 646 |
+
is created inside _stage_sampler_context and model_wrap_cfg is patched to
|
| 647 |
+
stage dims before the sampler loop starts.
|
| 648 |
+
"""
|
| 649 |
+
if sampler_profile == 'smea':
|
| 650 |
+
with _stage_sampler_context(p, stage_w, stage_h):
|
| 651 |
+
with _force_safe_sigmas_for_smea(p):
|
| 652 |
+
p.sampler = sd_samplers.create_sampler(p.sampler_name, p.sd_model)
|
| 653 |
+
_patch_sampler_func_sigmas_to_1d(p)
|
| 654 |
+
_resize_model_wrap_cfg(p, stage_w, stage_h)
|
| 655 |
+
_call_script_hooks(p, samples)
|
| 656 |
+
return p.sampler.sample_img2img(
|
| 657 |
+
p, samples, noise,
|
| 658 |
+
conditioning, unconditional_conditioning,
|
| 659 |
+
steps=steps,
|
| 660 |
+
image_conditioning=p.image_conditioning,
|
| 661 |
+
)
|
| 662 |
+
else:
|
| 663 |
+
_call_script_hooks(p, samples)
|
| 664 |
+
return p.sampler.sample_img2img(
|
| 665 |
+
p, samples, noise,
|
| 666 |
+
conditioning, unconditional_conditioning,
|
| 667 |
+
steps=steps,
|
| 668 |
+
image_conditioning=p.image_conditioning,
|
| 669 |
+
)
|
| 670 |
+
|
| 671 |
+
|
| 672 |
+
def _run_img2img_refinement(p, samples, seeds, subseeds, subseed_strength,
|
| 673 |
+
conditioning, unconditional_conditioning,
|
| 674 |
+
refine_idx: int, n_refine: int,
|
| 675 |
+
refine_step_mode: str,
|
| 676 |
+
sampler_profile: str = 'standard') -> torch.Tensor:
|
| 677 |
+
"""
|
| 678 |
+
Full VAE-decode refinement.
|
| 679 |
+
Decodes latents, re-encodes as img2img conditioning, then delegates to
|
| 680 |
+
_stage_sample_img2img which handles sampler lifecycle per profile.
|
| 681 |
+
|
| 682 |
+
refine_idx / n_refine / refine_step_mode feed into _get_refine_steps so
|
| 683 |
+
the step budget reflects the chosen policy rather than always uniform.
|
| 684 |
+
"""
|
| 685 |
+
steps_for_refinement = _get_refine_steps(p, refine_idx, n_refine, refine_step_mode)
|
| 686 |
+
noise = _make_noise(p, samples.shape[1:], seeds, subseeds, subseed_strength)
|
| 687 |
+
|
| 688 |
+
decoded = decode_latent_batch(p.sd_model, samples,
|
| 689 |
+
target_device=devices.cpu, check_for_nans=True)
|
| 690 |
+
decoded = torch.stack(decoded).float()
|
| 691 |
+
decoded = torch.clamp((decoded + 1.0) / 2.0, 0.0, 1.0)
|
| 692 |
+
source_img = decoded * 2.0 - 1.0
|
| 693 |
+
|
| 694 |
+
p.image_conditioning = p.img2img_image_conditioning(source_img, samples)
|
| 695 |
+
|
| 696 |
+
# stage dims derived from the current latent (samples is already stage-sized)
|
| 697 |
+
stage_w = samples.shape[3] * opt_f
|
| 698 |
+
stage_h = samples.shape[2] * opt_f
|
| 699 |
+
|
| 700 |
+
return _stage_sample_img2img(
|
| 701 |
+
p, samples, noise, conditioning, unconditional_conditioning,
|
| 702 |
+
stage_w, stage_h, steps_for_refinement, sampler_profile,
|
| 703 |
+
)
|
| 704 |
+
|
| 705 |
+
|
| 706 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 707 |
+
# Version implementations
|
| 708 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 709 |
+
|
| 710 |
+
def sample_v1_exact(p, conditioning, unconditional_conditioning,
|
| 711 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 712 |
+
"""
|
| 713 |
+
v1 (exact) β original implementation, preserved 1:1 as a regression baseline.
|
| 714 |
+
Does NOT use the pre-built (deduped) plan for sampling; rebuilds the raw
|
| 715 |
+
stage list from min_scale/max_scale/steps exactly as the original did.
|
| 716 |
+
|
| 717 |
+
For infotext honesty, a v1-native StagePlan is constructed from the same
|
| 718 |
+
raw stage list so PG plan reflects what v1 actually ran, not the deduped plan
|
| 719 |
+
that the wrapper built for guard purposes.
|
| 720 |
+
"""
|
| 721 |
+
min_scale = float(p.progressive_growing_min_scale)
|
| 722 |
+
max_scale = float(p.progressive_growing_max_scale)
|
| 723 |
+
resolution_steps = np.linspace(min_scale, max_scale, int(p.progressive_growing_steps))
|
| 724 |
+
|
| 725 |
+
def _snap_v1(v):
|
| 726 |
v_int = int(v)
|
| 727 |
v_int = max(opt_f, v_int)
|
| 728 |
v_int = (v_int // opt_f) * opt_f
|
| 729 |
return max(opt_f, v_int)
|
| 730 |
|
| 731 |
+
# build the honest v1 stage list (duplicates preserved, no forced final snap)
|
| 732 |
+
v1_sizes = [
|
| 733 |
+
(_snap_v1(p.width * s), _snap_v1(p.height * s))
|
| 734 |
+
for s in resolution_steps
|
| 735 |
+
]
|
| 736 |
+
v1_plan = StagePlan(
|
| 737 |
+
sizes=v1_sizes,
|
| 738 |
+
final_w=v1_sizes[-1][0],
|
| 739 |
+
final_h=v1_sizes[-1][1],
|
| 740 |
+
)
|
| 741 |
+
|
| 742 |
+
_write_params(p, v1_plan, refine_policy='all stages',
|
| 743 |
+
refine_step_policy='uniform (v1 fixed)')
|
| 744 |
+
sampler_profile, _ = _get_sampler_profile(p)
|
| 745 |
+
# v1 always uses bicubic regardless of the UI dropdown
|
| 746 |
+
try:
|
| 747 |
+
p.extra_generation_params['PG interp'] = 'bicubic (v1 fixed)'
|
| 748 |
+
if sampler_profile != 'standard':
|
| 749 |
+
p.extra_generation_params['PG sampler profile'] = sampler_profile
|
| 750 |
+
# v1 builds its stage list from the manual Stages slider, not from
|
| 751 |
+
# auto stage count β override the fields _write_params may have set
|
| 752 |
+
# so infotext describes what v1 actually ran, not what the user expected
|
| 753 |
+
if getattr(p, 'progressive_growing_auto_stages', False):
|
| 754 |
+
p.extra_generation_params['PG stages mode'] = 'manual (v1 legacy β auto ignored)'\
|
| 755 |
+
|
| 756 |
+
p.extra_generation_params.pop('PG target jump', None)
|
| 757 |
+
p.extra_generation_params.pop('PG requested', None)
|
| 758 |
+
p.extra_generation_params.pop('PG actual', None)
|
| 759 |
+
except Exception:
|
| 760 |
+
pass
|
| 761 |
+
|
| 762 |
+
initial_width, initial_height = v1_sizes[0]
|
| 763 |
|
|
|
|
| 764 |
x = create_random_tensors(
|
| 765 |
(opt_C, initial_height // opt_f, initial_width // opt_f),
|
| 766 |
+
seeds, subseeds=subseeds, subseed_strength=subseed_strength,
|
| 767 |
+
seed_resize_from_h=p.seed_resize_from_h,
|
| 768 |
+
seed_resize_from_w=p.seed_resize_from_w, p=p,
|
|
|
|
|
|
|
|
|
|
| 769 |
)
|
| 770 |
|
| 771 |
+
image_cond = p.txt2img_image_conditioning(x, width=initial_width, height=initial_height)
|
| 772 |
+
|
| 773 |
+
samples = _stage_sample_txt2img(
|
| 774 |
+
p, x, conditioning, unconditional_conditioning,
|
| 775 |
+
image_cond, initial_width, initial_height, sampler_profile,
|
|
|
|
|
|
|
| 776 |
)
|
| 777 |
|
| 778 |
total_stages = len(resolution_steps)
|
| 779 |
|
|
|
|
| 780 |
for i in range(1, total_stages):
|
| 781 |
+
target_width, target_height = v1_sizes[i]
|
|
|
|
| 782 |
|
|
|
|
| 783 |
samples = torch.nn.functional.interpolate(
|
| 784 |
samples,
|
| 785 |
size=(target_height // opt_f, target_width // opt_f),
|
| 786 |
+
mode='bicubic', align_corners=False,
|
|
|
|
| 787 |
)
|
| 788 |
|
| 789 |
+
if p.progressive_growing_refinement:
|
| 790 |
+
samples = _run_img2img_refinement(
|
| 791 |
+
p, samples, seeds, subseeds, subseed_strength,
|
| 792 |
+
conditioning, unconditional_conditioning,
|
| 793 |
+
refine_idx=i - 1,
|
| 794 |
+
n_refine=total_stages - 1,
|
| 795 |
+
refine_step_mode='uniform', # v1 legacy always uniform
|
| 796 |
+
sampler_profile=sampler_profile,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 797 |
)
|
| 798 |
|
| 799 |
+
return samples
|
| 800 |
+
|
| 801 |
+
|
| 802 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 803 |
+
|
| 804 |
+
def _run_plan_with_policy(p, plan: StagePlan, conditioning, unconditional_conditioning,
|
| 805 |
+
seeds, subseeds, subseed_strength,
|
| 806 |
+
refine_predicate) -> torch.Tensor:
|
| 807 |
+
"""
|
| 808 |
+
Shared loop for v2 / v4 / v5 / v6.
|
| 809 |
+
|
| 810 |
+
refine_predicate(stage_idx, w, h, plan) -> bool
|
| 811 |
+
Called for every stage after the first; return True to run refinement.
|
| 812 |
+
|
| 813 |
+
Reads p.progressive_growing_interp_mode for latent upscale (default: bicubic).
|
| 814 |
+
Reads p.progressive_growing_refine_step_mode for per-pass step budget.
|
| 815 |
+
|
| 816 |
+
Pre-computes the number of refine passes so _get_refine_steps can allocate
|
| 817 |
+
a budget that accounts for the total workload (needed by late-heavy / final-heavy).
|
| 818 |
+
|
| 819 |
+
Sampler compatibility
|
| 820 |
+
--------------------
|
| 821 |
+
_get_sampler_profile() is called once to determine the active profile.
|
| 822 |
+
'smea' profile wraps every sampler.sample() / sample_img2img() call with
|
| 823 |
+
_stage_sampler_context(), which temporarily sets p.width/p.height to stage
|
| 824 |
+
dims and rescales p.init_latent/mask/nmask. This fixes spatial-sigma
|
| 825 |
+
schedulers and model-wrapper state mismatches without touching the rest of
|
| 826 |
+
the pipeline. The profile name is written to extra_generation_params as
|
| 827 |
+
'PG sampler profile' when non-standard.
|
| 828 |
+
"""
|
| 829 |
+
first_w, first_h = plan.sizes[0]
|
| 830 |
+
interp_mode = getattr(p, 'progressive_growing_interp_mode', LATENT_INTERP_DEFAULT)
|
| 831 |
+
step_mode = getattr(p, 'progressive_growing_refine_step_mode', REFINE_STEP_DEFAULT)
|
| 832 |
+
do_refine = getattr(p, 'progressive_growing_refinement', True)
|
| 833 |
+
sampler_profile, _ = _get_sampler_profile(p)
|
| 834 |
+
|
| 835 |
+
# record non-standard profile in infotext
|
| 836 |
+
if sampler_profile != 'standard':
|
| 837 |
+
try:
|
| 838 |
+
p.extra_generation_params['PG sampler profile'] = sampler_profile
|
| 839 |
+
except Exception:
|
| 840 |
+
pass
|
| 841 |
+
|
| 842 |
+
# pre-count how many stages will actually refine so budget can be distributed
|
| 843 |
+
n_refine = sum(
|
| 844 |
+
1 for i, (w, h) in enumerate(plan.sizes[1:], start=1)
|
| 845 |
+
if do_refine and refine_predicate(i, w, h, plan)
|
| 846 |
+
) if do_refine else 0
|
| 847 |
+
|
| 848 |
+
x = _initial_latent(p, first_w, first_h, seeds, subseeds, subseed_strength)
|
| 849 |
+
image_cond = p.txt2img_image_conditioning(x, width=first_w, height=first_h)
|
| 850 |
+
|
| 851 |
+
samples = _stage_sample_txt2img(
|
| 852 |
+
p, x, conditioning, unconditional_conditioning,
|
| 853 |
+
image_cond, first_w, first_h, sampler_profile,
|
| 854 |
+
)
|
| 855 |
+
|
| 856 |
+
refine_idx = 0
|
| 857 |
+
for i, (w, h) in enumerate(plan.sizes[1:], start=1):
|
| 858 |
+
samples = _upscale_latent(samples, w, h, interp_mode=interp_mode)
|
| 859 |
+
|
| 860 |
+
if do_refine and refine_predicate(i, w, h, plan):
|
| 861 |
+
samples = _run_img2img_refinement(
|
| 862 |
+
p, samples, seeds, subseeds, subseed_strength,
|
| 863 |
+
conditioning, unconditional_conditioning,
|
| 864 |
+
refine_idx=refine_idx,
|
| 865 |
+
n_refine=max(1, n_refine),
|
| 866 |
+
refine_step_mode=step_mode,
|
| 867 |
+
sampler_profile=sampler_profile,
|
| 868 |
)
|
| 869 |
+
refine_idx += 1
|
| 870 |
|
| 871 |
return samples
|
| 872 |
|
| 873 |
|
| 874 |
+
def sample_v2_safe(p, conditioning, unconditional_conditioning,
|
| 875 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 876 |
+
"""
|
| 877 |
+
v2 (safe) β validated plan, dedup, correct per-stage conditioning size, script hooks.
|
| 878 |
+
Behaves identically to v1 where plans agree; differs only in edge cases.
|
| 879 |
+
"""
|
| 880 |
+
_write_params(p, plan, refine_policy='all stages',
|
| 881 |
+
refine_step_policy=getattr(p, 'progressive_growing_refine_step_mode', REFINE_STEP_DEFAULT))
|
| 882 |
+
|
| 883 |
+
def always(i, w, h, plan):
|
| 884 |
+
return True
|
| 885 |
+
|
| 886 |
+
return _run_plan_with_policy(
|
| 887 |
+
p, plan, conditioning, unconditional_conditioning,
|
| 888 |
+
seeds, subseeds, subseed_strength,
|
| 889 |
+
refine_predicate=always,
|
| 890 |
+
)
|
| 891 |
+
|
| 892 |
+
|
| 893 |
+
def sample_v3_fast(p, conditioning, unconditional_conditioning,
|
| 894 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 895 |
+
"""
|
| 896 |
+
v3 (fast) β refinement only on the final stage.
|
| 897 |
+
Avoids expensive VAE decode on every intermediate stage.
|
| 898 |
+
Best for quick iteration or large stage counts.
|
| 899 |
+
"""
|
| 900 |
+
_write_params(p, plan, refine_policy='final stage only',
|
| 901 |
+
refine_step_policy=getattr(p, 'progressive_growing_refine_step_mode', REFINE_STEP_DEFAULT))
|
| 902 |
+
|
| 903 |
+
def only_last(i, w, h, plan):
|
| 904 |
+
return i == plan.n_stages - 1
|
| 905 |
+
|
| 906 |
+
return _run_plan_with_policy(
|
| 907 |
+
p, plan, conditioning, unconditional_conditioning,
|
| 908 |
+
seeds, subseeds, subseed_strength,
|
| 909 |
+
refine_predicate=only_last,
|
| 910 |
+
)
|
| 911 |
+
|
| 912 |
+
|
| 913 |
+
def sample_v4_balanced(p, conditioning, unconditional_conditioning,
|
| 914 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 915 |
+
"""
|
| 916 |
+
v4 (balanced) β refinement on stages whose linear scale >= BALANCED_REFINE_THRESHOLD.
|
| 917 |
+
|
| 918 |
+
The threshold is a *linear* dimension ratio (sqrt of area ratio), so
|
| 919 |
+
BALANCED_REFINE_THRESHOLD=0.70 means: refine when the stage side is >= 70 %
|
| 920 |
+
of the final side, which corresponds to β 49 % of the final pixel area.
|
| 921 |
+
This avoids refinement on cheap small upscales while still covering the
|
| 922 |
+
detail-sensitive near-final stages.
|
| 923 |
+
"""
|
| 924 |
+
_write_params(p, plan, refine_policy=f'linear scale >= {BALANCED_REFINE_THRESHOLD}',
|
| 925 |
+
refine_step_policy=getattr(p, 'progressive_growing_refine_step_mode', REFINE_STEP_DEFAULT))
|
| 926 |
+
|
| 927 |
+
final_area = plan.final_w * plan.final_h
|
| 928 |
+
|
| 929 |
+
def large_stages_only(i, w, h, plan):
|
| 930 |
+
# linear scale = sqrt(stage_area / final_area)
|
| 931 |
+
linear_scale = math.sqrt((w * h) / final_area) if final_area > 0 else 0.0
|
| 932 |
+
return linear_scale >= BALANCED_REFINE_THRESHOLD
|
| 933 |
+
|
| 934 |
+
return _run_plan_with_policy(
|
| 935 |
+
p, plan, conditioning, unconditional_conditioning,
|
| 936 |
+
seeds, subseeds, subseed_strength,
|
| 937 |
+
refine_predicate=large_stages_only,
|
| 938 |
+
)
|
| 939 |
+
|
| 940 |
+
|
| 941 |
+
def sample_v5_latent(p, conditioning, unconditional_conditioning,
|
| 942 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 943 |
+
"""
|
| 944 |
+
v5 (latent only) β pure latent upscale, no refinement at any stage.
|
| 945 |
+
Fastest mode; useful to study the raw effect of latent-space growing.
|
| 946 |
+
"""
|
| 947 |
+
_write_params(p, plan, refine_policy='none', refine_step_policy='none')
|
| 948 |
+
|
| 949 |
+
def never(i, w, h, plan):
|
| 950 |
+
return False
|
| 951 |
+
|
| 952 |
+
return _run_plan_with_policy(
|
| 953 |
+
p, plan, conditioning, unconditional_conditioning,
|
| 954 |
+
seeds, subseeds, subseed_strength,
|
| 955 |
+
refine_predicate=never,
|
| 956 |
+
)
|
| 957 |
+
|
| 958 |
+
|
| 959 |
+
def sample_v6_adaptive(p, conditioning, unconditional_conditioning,
|
| 960 |
+
seeds, subseeds, subseed_strength, prompts, plan: StagePlan):
|
| 961 |
+
"""
|
| 962 |
+
v6 (adaptive) β refinement policy scales with stage count.
|
| 963 |
+
|
| 964 |
+
2 stages β refine every post-initial stage (progressive would be pointless otherwise)
|
| 965 |
+
3β4 stages β refine the last 2 stages only
|
| 966 |
+
5+ stages β refine stages whose linear scale >= ADAPTIVE_REFINE_THRESHOLD,
|
| 967 |
+
plus always force-refine the final stage
|
| 968 |
+
|
| 969 |
+
Sits between v3 (final only) and v2 (all stages): cheaper than v2 on long
|
| 970 |
+
plans, more thorough than v3 on short ones.
|
| 971 |
+
"""
|
| 972 |
+
_write_params(p, plan, refine_policy=f'adaptive (threshold >= {ADAPTIVE_REFINE_THRESHOLD:.2f})',
|
| 973 |
+
refine_step_policy=getattr(p, 'progressive_growing_refine_step_mode', REFINE_STEP_DEFAULT))
|
| 974 |
+
|
| 975 |
+
final_area = plan.final_w * plan.final_h
|
| 976 |
+
|
| 977 |
+
def adaptive(i, w, h, plan):
|
| 978 |
+
n = plan.n_stages
|
| 979 |
+
if n <= 2:
|
| 980 |
+
# only 1 post-initial stage β always refine it
|
| 981 |
+
return True
|
| 982 |
+
if n <= 4:
|
| 983 |
+
# short plan: refine the last 2 stages
|
| 984 |
+
return i >= n - 2
|
| 985 |
+
# long plan: threshold + guaranteed final
|
| 986 |
+
linear_scale = math.sqrt((w * h) / final_area) if final_area > 0 else 0.0
|
| 987 |
+
return i == n - 1 or linear_scale >= ADAPTIVE_REFINE_THRESHOLD
|
| 988 |
+
|
| 989 |
+
return _run_plan_with_policy(
|
| 990 |
+
p, plan, conditioning, unconditional_conditioning,
|
| 991 |
+
seeds, subseeds, subseed_strength,
|
| 992 |
+
refine_predicate=adaptive,
|
| 993 |
+
)
|
| 994 |
+
|
| 995 |
+
|
| 996 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 997 |
+
# Version registry
|
| 998 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 999 |
+
|
| 1000 |
+
_VERSIONS: dict[str, callable] = {
|
| 1001 |
+
"v2 (safe)": sample_v2_safe,
|
| 1002 |
+
"v3 (fast)": sample_v3_fast,
|
| 1003 |
+
"v4 (balanced)": sample_v4_balanced,
|
| 1004 |
+
"v5 (latent)": sample_v5_latent,
|
| 1005 |
+
"v6 (adaptive)": sample_v6_adaptive,
|
| 1006 |
+
"v1 (exact)": sample_v1_exact, # legacy / regression reference
|
| 1007 |
}
|
| 1008 |
|
| 1009 |
+
_DEFAULT_VERSION = "v2 (safe)"
|
| 1010 |
+
|
| 1011 |
|
| 1012 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 1013 |
+
# Monkey-patch β thin wrapper, applied exactly once
|
| 1014 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 1015 |
|
| 1016 |
+
_PATCHED = False
|
| 1017 |
_ORIG_SAMPLE = None
|
| 1018 |
|
| 1019 |
|
| 1020 |
def _apply_patch_once() -> None:
|
| 1021 |
+
"""Patch StableDiffusionProcessingTxt2Img.sample once at first use."""
|
| 1022 |
|
| 1023 |
global _PATCHED, _ORIG_SAMPLE
|
| 1024 |
if _PATCHED:
|
|
|
|
| 1028 |
if cls is None:
|
| 1029 |
return
|
| 1030 |
|
| 1031 |
+
if getattr(cls, '_pg_ext_patched', False):
|
|
|
|
| 1032 |
_PATCHED = True
|
| 1033 |
return
|
| 1034 |
|
| 1035 |
_ORIG_SAMPLE = cls.sample
|
| 1036 |
|
| 1037 |
+
def _sample_wrapper(self,
|
| 1038 |
+
conditioning, unconditional_conditioning,
|
| 1039 |
+
seeds, subseeds, subseed_strength, prompts):
|
| 1040 |
+
|
| 1041 |
+
ok, reason, plan = _should_use_progressive(self)
|
| 1042 |
|
| 1043 |
+
if ok:
|
| 1044 |
+
# For 'standard' profile: create sampler once here (mirrors what the
|
| 1045 |
+
# original sample() does internally).
|
| 1046 |
+
# For 'smea' profile: sampler is created fresh per stage-pass inside
|
| 1047 |
+
# _stage_sample_txt2img / _stage_sample_img2img so that all
|
| 1048 |
+
# size-dependent state (sigma schedules, spatial caches) is
|
| 1049 |
+
# initialised against stage dims, not the final target size.
|
| 1050 |
+
profile, _ = _get_sampler_profile(self)
|
| 1051 |
+
if profile == 'standard':
|
| 1052 |
+
self.sampler = sd_samplers.create_sampler(self.sampler_name, self.sd_model)
|
| 1053 |
|
| 1054 |
+
ver = getattr(self, 'progressive_growing_version', _DEFAULT_VERSION)
|
| 1055 |
+
fn = _VERSIONS.get(ver, sample_v2_safe)
|
| 1056 |
+
# plan was already built by _should_use_progressive β pass it through
|
| 1057 |
+
# so version fns don't call build_plan() a second time
|
| 1058 |
+
return fn(self, conditioning, unconditional_conditioning,
|
| 1059 |
+
seeds, subseeds, subseed_strength, prompts, plan)
|
| 1060 |
|
| 1061 |
+
# record skip reason if there was one
|
| 1062 |
+
if reason:
|
| 1063 |
+
try:
|
| 1064 |
+
self.extra_generation_params['PG skip'] = reason
|
| 1065 |
+
except Exception:
|
| 1066 |
+
pass
|
| 1067 |
+
|
| 1068 |
+
return _ORIG_SAMPLE(self, conditioning, unconditional_conditioning,
|
| 1069 |
+
seeds, subseeds, subseed_strength, prompts)
|
| 1070 |
+
|
| 1071 |
+
cls.sample = _sample_wrapper
|
| 1072 |
+
cls._pg_ext_patched = True
|
| 1073 |
_PATCHED = True
|
| 1074 |
|
| 1075 |
|
| 1076 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 1077 |
# Always-visible UI script
|
| 1078 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
|
|
|
| 1079 |
|
| 1080 |
class ProgressiveGrowingAlwaysVisible(scripts.Script):
|
| 1081 |
+
|
| 1082 |
def title(self):
|
| 1083 |
return "Progressive Growing"
|
| 1084 |
|
| 1085 |
def show(self, is_img2img):
|
|
|
|
| 1086 |
return scripts.AlwaysVisible if not is_img2img else False
|
| 1087 |
|
| 1088 |
def ui(self, is_img2img):
|
| 1089 |
with gr.Accordion("Progressive Growing", open=False):
|
| 1090 |
enabled = gr.Checkbox(value=False, label="Enable")
|
| 1091 |
+
version = gr.Dropdown(
|
| 1092 |
+
choices=list(_VERSIONS.keys()),
|
| 1093 |
+
value=_DEFAULT_VERSION,
|
| 1094 |
+
label="Mode",
|
| 1095 |
+
)
|
| 1096 |
+
|
| 1097 |
+
with gr.Row():
|
| 1098 |
+
min_scale = gr.Slider(
|
| 1099 |
+
minimum=0.1, maximum=1.0, step=0.05,
|
| 1100 |
+
value=0.25, label="Min scale",
|
| 1101 |
+
)
|
| 1102 |
+
max_scale = gr.Slider(
|
| 1103 |
+
minimum=0.1, maximum=1.0, step=0.05,
|
| 1104 |
+
value=1.0, label="Max scale",
|
| 1105 |
+
)
|
| 1106 |
|
| 1107 |
+
stages = gr.Slider(
|
| 1108 |
+
minimum=2, maximum=16, step=1,
|
| 1109 |
+
value=4, label="Stages (ignored when Auto is on)",
|
| 1110 |
+
)
|
| 1111 |
+
with gr.Row():
|
| 1112 |
+
auto_stages = gr.Checkbox(value=False, label="Auto stage count")
|
| 1113 |
+
auto_jump = gr.Slider(
|
| 1114 |
+
minimum=1.1, maximum=2.0, step=0.05,
|
| 1115 |
+
value=AUTO_STAGE_JUMP_DEFAULT,
|
| 1116 |
+
label="Target jump per stage",
|
| 1117 |
+
)
|
| 1118 |
+
auto_max = gr.Slider(
|
| 1119 |
+
minimum=2, maximum=12, step=1,
|
| 1120 |
+
value=AUTO_STAGE_MAX_DEFAULT,
|
| 1121 |
+
label="Max auto stages",
|
| 1122 |
+
)
|
| 1123 |
+
with gr.Row():
|
| 1124 |
+
refinement = gr.Checkbox(value=True, label="Refinement between stages")
|
| 1125 |
+
refine_step_mode = gr.Dropdown(
|
| 1126 |
+
choices=REFINE_STEP_MODES,
|
| 1127 |
+
value=REFINE_STEP_DEFAULT,
|
| 1128 |
+
label="Refinement step budget",
|
| 1129 |
+
)
|
| 1130 |
+
interp_mode = gr.Dropdown(
|
| 1131 |
+
choices=LATENT_INTERP_MODES,
|
| 1132 |
+
value=LATENT_INTERP_DEFAULT,
|
| 1133 |
+
label="Latent upscale interpolation",
|
| 1134 |
+
)
|
| 1135 |
|
| 1136 |
gr.Markdown(
|
| 1137 |
+
"**Modes**\n"
|
| 1138 |
+
"- **v2 (safe)** β validated, deduped stages, refinement at every stage *(default)*\n"
|
| 1139 |
+
"- **v3 (fast)** β refinement only on the final stage; fastest, no VAE mid-pass\n"
|
| 1140 |
+
"- **v4 (balanced)** β refinement when stage side β₯ 70 % of final side (β 49 % of area)\n"
|
| 1141 |
+
"- **v5 (latent)** β pure latent upscale, zero refinement\n"
|
| 1142 |
+
"- **v6 (adaptive)** β 2 stages: all; 3β4 stages: last 2; 5+ stages: threshold + final\n"
|
| 1143 |
+
"- **v1 (exact)** β original implementation, kept for regression comparison; "
|
| 1144 |
+
"always uses manual Stages, ignores Auto stage count\n\n"
|
| 1145 |
+
"**Auto stage count** β ignores Stages slider; computes count so each upscale grows "
|
| 1146 |
+
"the latent side by ~Target jump (1.35 β 35 %). 0.25β1.0 β 6 stages, 0.5β1.0 β 3 stages.\n\n"
|
| 1147 |
+
"**Interpolation** β bicubic: smooth/photo; bilinear: softer; nearest: fastest/pixel-art; area: anti-aliased\n\n"
|
| 1148 |
+
"**Refinement budget** β uniform: equal steps per pass (legacy); "
|
| 1149 |
+
"late-heavy: growing budget towards final pass; "
|
| 1150 |
+
"final-heavy: minimal steps on all but the last refinement pass\n\n"
|
| 1151 |
+
"β Incompatible with **Hires. fix** β enabling both disables Progressive Growing."
|
| 1152 |
)
|
| 1153 |
|
| 1154 |
+
return [enabled, version, min_scale, max_scale,
|
| 1155 |
+
stages, auto_stages, auto_jump, auto_max,
|
| 1156 |
+
refinement, refine_step_mode, interp_mode]
|
| 1157 |
|
| 1158 |
+
def process(self, p,
|
| 1159 |
+
enabled, version, min_scale, max_scale,
|
| 1160 |
+
stages, auto_stages, auto_jump, auto_max,
|
| 1161 |
+
refinement, refine_step_mode, interp_mode):
|
| 1162 |
|
| 1163 |
+
_apply_patch_once()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1164 |
|
| 1165 |
+
p.enable_progressive_growing = bool(enabled)
|
| 1166 |
+
p.progressive_growing_version = str(version)
|
| 1167 |
+
p.progressive_growing_min_scale = float(min_scale)
|
| 1168 |
+
p.progressive_growing_max_scale = float(max_scale)
|
| 1169 |
+
p.progressive_growing_steps = int(stages)
|
| 1170 |
+
p.progressive_growing_auto_stages = bool(auto_stages)
|
| 1171 |
+
p.progressive_growing_auto_jump = float(auto_jump)
|
| 1172 |
+
p.progressive_growing_auto_max = int(auto_max)
|
| 1173 |
+
p.progressive_growing_refinement = bool(refinement)
|
| 1174 |
+
p.progressive_growing_refine_step_mode = str(refine_step_mode) if refine_step_mode in REFINE_STEP_MODES else REFINE_STEP_DEFAULT
|
| 1175 |
+
p.progressive_growing_interp_mode = str(interp_mode) if interp_mode in LATENT_INTERP_MODES else LATENT_INTERP_DEFAULT
|
|
|