IxMxAMAR
3f9e39bb44
Critical security fixes ----------------------- - Sandbox Jinja2 in template_render + prompt_template_render (was RCE via shared workflows) - Drop user home dir from filesystem allow-list (was arbitrary delete/overwrite) - Confine save_image_with_manifest + sanitize filename basename - SSRF block on http_get/http_post/notify_webhook (catches 169.254.169.254 metadata) - Case-insensitive path confinement on Windows Critical correctness/perf ------------------------- - halftone_dots: O(N^2) -> O(cells*36); was freezing worker on 1024+ images - Preserve latent dict keys (noise_mask, batch_index) in latent_math/noise/upscale - Preserve alpha channel in invert + threshold_binary on RGBA inputs - Force odd kernel in mask_erode/mask_dilate to prevent H+1 shape mismatch - Snap aspect_ratio_pick output to multiples of 8 (was crashing SD UNet) Theme engine (JS) ----------------- - Hoist window.matchMedia() out of per-frame hot path -> recovers 200 FPS from 25 - Wrap-once draw method to stop stomping other extensions wrappers - nodeRemoved hook + WeakMap groups + in-place canvas reuse -> three memory leaks closed - Wire thickness now uses connections_width (the wrap was a no-op vs LiteGraph) - onGroupAdded auto-themes new groups; budget change triggers redraw - Gradient caching in synthwave/glassmorphic New ops (10) ------------ - json_path (programming): dotted/bracketed path extract - latent_nan_guard, latent_pad_crop (mask_latent) - mask_from_depth, mask_invert (mask_latent) - color_match_histogram, image_composite_over (image_pro/value) - frequency_separation (image_pro/style) - pad_to_multiple (image_pro/spatial) - save_images_zip (io_workflow) - llm_enhance_prompt via Ollama / LM Studio / llama.cpp (prompt) Theme UX -------- - Right-click "Theme..." submenu on every node (including vanilla ComfyUI nodes) - Auto-contrast title text (WCAG AA-based flip) Tests: 373 Python + 33 JS = 406 passing (added 33 new tests for security + features). Full review notes in CHANGELOG.md.
116 lines
3.6 KiB
Python
116 lines
3.6 KiB
Python
"""Resolution and aspect-ratio helpers."""
|
||
|
||
import math
|
||
|
||
from .. import op
|
||
|
||
|
||
_RATIO_PRESETS = {
|
||
"1:1 (square)": (1, 1),
|
||
"4:3": (4, 3),
|
||
"3:2": (3, 2),
|
||
"16:9": (16, 9),
|
||
"21:9": (21, 9),
|
||
"9:16 (portrait)": (9, 16),
|
||
"3:4 (portrait)": (3, 4),
|
||
"2:3 (portrait)": (2, 3),
|
||
}
|
||
|
||
|
||
_SDXL_BUCKETS = [
|
||
(1024, 1024),
|
||
(1152, 896),
|
||
(896, 1152),
|
||
(1216, 832),
|
||
(832, 1216),
|
||
(1344, 768),
|
||
(768, 1344),
|
||
(1536, 640),
|
||
(640, 1536),
|
||
]
|
||
|
||
|
||
@op(
|
||
op_id="aspect_ratio_pick",
|
||
display_name="Aspect Ratio Pick",
|
||
category="Resolution",
|
||
input_schema={"required": {
|
||
"preset": (list(_RATIO_PRESETS.keys()), {"default": "16:9"}),
|
||
"long_side": ("INT", {"default": 1024, "min": 64, "max": 8192}),
|
||
}},
|
||
output_indices=(5, 6, 8),
|
||
description="Returns width (out:5) and height (out:6) for the chosen ratio scaled to long_side.",
|
||
)
|
||
def aspect_ratio_pick(self, preset, long_side=1024):
|
||
rw, rh = _RATIO_PRESETS.get(preset, (1, 1))
|
||
if rw >= rh:
|
||
w = int(long_side)
|
||
h = int(round(long_side * rh / rw))
|
||
else:
|
||
h = int(long_side)
|
||
w = int(round(long_side * rw / rh))
|
||
# Snap to multiples of 8 — SD UNet/VAE require this (typically /8 or /64).
|
||
# Without snapping, long_side=1000 + 16:9 produced 1000×562 which crashed
|
||
# the model.
|
||
w = max(8, (w // 8) * 8)
|
||
h = max(8, (h // 8) * 8)
|
||
info = {"preset": preset, "ratio": [rw, rh], "width": w, "height": h, "snapped_to": 8}
|
||
return (w, h, info)
|
||
|
||
|
||
@op(
|
||
op_id="sdxl_bucket_pick",
|
||
display_name="SDXL Bucket Pick (9 official sizes)",
|
||
category="Resolution",
|
||
input_schema={"required": {"index": ("INT", {"default": 0, "min": 0, "max": 8})}},
|
||
output_indices=(5, 6, 8),
|
||
description="Pick one of the 9 official SDXL training resolutions by index.",
|
||
)
|
||
def sdxl_bucket_pick(self, index=0):
|
||
w, h = _SDXL_BUCKETS[int(index) % len(_SDXL_BUCKETS)]
|
||
return (w, h, {"index": int(index), "width": w, "height": h})
|
||
|
||
|
||
@op(
|
||
op_id="snap_to_multiple",
|
||
display_name="Snap to Multiple",
|
||
category="Resolution",
|
||
input_schema={"required": {
|
||
"value": ("INT", {"default": 1024}),
|
||
"multiple": ("INT", {"default": 8, "min": 1, "max": 1024}),
|
||
}},
|
||
output_indices=(5,),
|
||
description="Round `value` to the nearest multiple of `multiple`.",
|
||
)
|
||
def snap_to_multiple(self, value, multiple=8):
|
||
m = max(1, int(multiple))
|
||
return (int(round(int(value) / m) * m),)
|
||
|
||
|
||
@op(
|
||
op_id="megapixel_calculator",
|
||
display_name="Megapixel Calculator (target MP + ratio)",
|
||
category="Resolution",
|
||
input_schema={"required": {
|
||
"target_megapixels": ("FLOAT", {"default": 1.0, "min": 0.01, "max": 64.0}),
|
||
"ratio_w": ("INT", {"default": 16, "min": 1, "max": 999}),
|
||
"ratio_h": ("INT", {"default": 9, "min": 1, "max": 999}),
|
||
"snap_multiple": ("INT", {"default": 8, "min": 1, "max": 1024}),
|
||
}},
|
||
output_indices=(5, 6, 8),
|
||
description="Find W,H such that W*H ≈ target_megapixels * 1e6 with given ratio, snapped to multiple.",
|
||
)
|
||
def megapixel_calculator(self, target_megapixels, ratio_w, ratio_h, snap_multiple=8):
|
||
target_pixels = float(target_megapixels) * 1_000_000
|
||
ratio_w, ratio_h = int(ratio_w), int(ratio_h)
|
||
# W = sqrt(target * ratio_w / ratio_h); H = W * ratio_h / ratio_w
|
||
w = math.sqrt(target_pixels * ratio_w / ratio_h)
|
||
h = w * ratio_h / ratio_w
|
||
m = max(1, int(snap_multiple))
|
||
w_snapped = int(round(w / m) * m)
|
||
h_snapped = int(round(h / m) * m)
|
||
return (w_snapped, h_snapped, {
|
||
"width": w_snapped, "height": h_snapped,
|
||
"actual_megapixels": round((w_snapped * h_snapped) / 1_000_000, 4),
|
||
})
|