Isulion
fc903d4acf
Breaking changes: - Same seeds now yield different prompts than 5.x (reproducibility fixed) - Civitai Trending: removed broken 'model' filter widget - Shutdown node is a no-op unless ISULION_ENABLE_SHUTDOWN=1 Highlights: - Seeding: all 109 theme handlers share the seeded ConfigManager PRNG (284 global random.* call sites migrated); same seed => same prompt - Theme registry rewritten with @register_theme decorators; removed HANDLER_CLASS_NAMES_MAP, star-imports and globals() lookups - Shared ThemeRegistry singleton: INPUT_TYPES no longer re-reads ~100 JSON files per call (2.6ms -> 0.003ms cached) - load_images_node: paths sandboxed to ComfyUI input dir, deterministic sha1 IS_CHANGED, PIL context managers, no bare excepts - Overlay node: full batch support, base channel count preserved (RGB in -> RGB out), pass-through on internal errors - QR node: vectorized module drawing (np.kron), logging instead of prints - New net_utils.safe_download: http(s)-only scheme + 20MB cap for URL fetching nodes (SSRF/decompression-bomb hardening) - Removed unused soft_empty_cache() calls, dead code (culinary_handler, IsulionCategorySelector, orphan Core_Nodes/config dir, dead js ext), invalid WEB_DIRECTORY; fixed display-name mappings - s300 handler: fixed call to non-existent config method (latent crash) - Added MIT LICENSE and requirements.txt (requests)
101 lines
2.8 KiB
Python
101 lines
2.8 KiB
Python
import torch
|
|
import math
|
|
import random
|
|
import numpy
|
|
|
|
|
|
class IsuCollageNode:
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"images": ("IMAGE",),
|
|
"seed": ("INT", {"optional": True}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("IMAGE",)
|
|
RETURN_NAMES = ("collage_image",)
|
|
FUNCTION = "create_collage"
|
|
CATEGORY = "Isulion/Image"
|
|
|
|
def create_collage(self, images, seed=None):
|
|
if seed is not None:
|
|
torch.manual_seed(seed)
|
|
random.seed(seed)
|
|
numpy.random.seed(seed)
|
|
|
|
imgs = list(images)
|
|
if seed is not None:
|
|
random.shuffle(imgs)
|
|
|
|
if not imgs:
|
|
raise ValueError("No images provided")
|
|
|
|
imgs = [img.squeeze(0) if len(img.shape) == 4 else img for img in imgs]
|
|
imgs = [img.float() for img in imgs]
|
|
|
|
n = len(imgs)
|
|
if n == 1:
|
|
return (imgs[0].unsqueeze(0),)
|
|
|
|
W = 1024
|
|
|
|
numRows = math.ceil(math.sqrt(n))
|
|
maxImagesPerRow = math.ceil(n / numRows)
|
|
|
|
rows = []
|
|
cur_i = []
|
|
cur_r = []
|
|
cur_w = 0.0
|
|
|
|
for i in range(n):
|
|
ar = imgs[i].shape[1] / imgs[i].shape[0]
|
|
cur_i.append(i)
|
|
cur_r.append(ar)
|
|
cur_w += 150.0 * ar
|
|
if cur_w >= W or i == n - 1 or len(cur_i) >= maxImagesPerRow:
|
|
rows.append((cur_i[:], cur_r[:], cur_w))
|
|
cur_i = []
|
|
cur_r = []
|
|
cur_w = 0.0
|
|
|
|
rects = []
|
|
y = 0
|
|
for idxs, rats, rw in rows:
|
|
scale = W / rw
|
|
rh = max(1, int(round(150.0 * scale)))
|
|
fws = [rh * r for r in rats]
|
|
rem = W
|
|
int_ws = []
|
|
for j, fw in enumerate(fws):
|
|
if j == len(fws) - 1:
|
|
int_ws.append(rem)
|
|
else:
|
|
iw = max(1, int(round(fw)))
|
|
int_ws.append(iw)
|
|
rem -= iw
|
|
x = 0
|
|
for j, idx in enumerate(idxs):
|
|
iw = int_ws[j]
|
|
rects.append((idx, x, y, iw, rh))
|
|
x += iw
|
|
y += rh
|
|
|
|
H = y
|
|
|
|
canvas = torch.full((H, W, 3), 245.0 / 255.0, dtype=torch.float32, device=imgs[0].device)
|
|
|
|
for idx, x, y, iw, ih in rects:
|
|
if iw < 1 or ih < 1:
|
|
continue
|
|
img = imgs[idx]
|
|
if img.shape[0] != ih or img.shape[1] != iw:
|
|
t = img.unsqueeze(0).permute(0, 3, 1, 2)
|
|
t = torch.nn.functional.interpolate(t, size=(ih, iw), mode='bilinear', align_corners=False)
|
|
r = t.permute(0, 2, 3, 1).squeeze(0)
|
|
else:
|
|
r = img
|
|
canvas[y:y+ih, x:x+iw] = r
|
|
|
|
return (canvas.unsqueeze(0),) |