65 lines
2.3 KiB
Python
65 lines
2.3 KiB
Python
from typing import Any
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
|
|
class MeldPixelate:
|
|
@classmethod
|
|
def INPUT_TYPES(cls) -> dict[str, Any]:
|
|
return {
|
|
"required": {
|
|
"image": ("IMAGE",),
|
|
"pixel_size": (
|
|
"INT",
|
|
{
|
|
"default": 16,
|
|
"min": 1,
|
|
"max": 256,
|
|
"step": 1,
|
|
"display": "number",
|
|
"tooltip": "The size of the pixel blocks (in pixels). Higher values create larger blocks.",
|
|
},
|
|
),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("IMAGE",)
|
|
RETURN_NAMES = ("image",)
|
|
FUNCTION = "pixelate"
|
|
CATEGORY = "Meld/Sandbox"
|
|
|
|
# --- Node Description ---
|
|
DESCRIPTION = "Applies a pixelation effect (mosaic) to an image by downscaling and then upscaling.\nUseful for creating retro pixel art effects or censoring content.\nFeatures:\n- High-quality pixelation using area downscaling\n- Sharp edges using nearest neighbor upscaling"
|
|
# ---------------------------
|
|
|
|
def pixelate(self, image: torch.Tensor, pixel_size: int) -> tuple[torch.Tensor]:
|
|
# image shape is [Batch, Height, Width, Channels]
|
|
B, H, W, C = image.shape
|
|
|
|
# 1. Calculate the target downscaled dimensions
|
|
# Corresponds to IntMath (//) in JSON
|
|
target_h = H // pixel_size
|
|
target_w = W // pixel_size
|
|
|
|
# Prevent dimension becoming 0
|
|
target_h = max(1, target_h)
|
|
target_w = max(1, target_w)
|
|
|
|
# ComfyUI images are BHWC, but PyTorch interpolate expects BCHW
|
|
img_permuted = image.permute(0, 3, 1, 2)
|
|
|
|
# 2. Downscale using 'area' (Average pooling effect)
|
|
# Corresponds to the first ImageScale (method: area) in JSON
|
|
downscaled = F.interpolate(img_permuted, size=(target_h, target_w), mode="area")
|
|
|
|
# 3. Upscale back to original size using 'nearest'
|
|
# Corresponds to the second ImageScale (method: nearest-exact) in JSON
|
|
# PyTorch's 'nearest' behaves identically to ComfyUI's 'nearest-exact'
|
|
upscaled = F.interpolate(downscaled, size=(H, W), mode="nearest")
|
|
|
|
# Permute back to BHWC
|
|
result = upscaled.permute(0, 2, 3, 1)
|
|
|
|
return (result,)
|