169 lines
6.7 KiB
Python
169 lines
6.7 KiB
Python
# Created by MSXYZ (AI-assisted)
|
||
# Temporal Anti-Aliasing (TAA) + Lightweight DLAA-style Sharpening
|
||
# v0.1.1 - DLAANet with residual learning and orthogonal weights
|
||
|
||
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
import logging
|
||
import comfy.model_management as mm
|
||
|
||
logger = logging.getLogger("VideoTAADLAA")
|
||
|
||
class _DLAANet(nn.Module):
|
||
def __init__(self):
|
||
super().__init__()
|
||
|
||
# Sobel ve Jitter tanımları
|
||
self.register_buffer("sobel_x", torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=torch.float32).view(1, 1, 3, 3))
|
||
self.register_buffer("sobel_y", torch.tensor([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=torch.float32).view(1, 1, 3, 3))
|
||
self.register_buffer("jitter_offsets", torch.tensor([[0.25, 0.25], [-0.25, -0.25], [-0.25, 0.25], [0.25, -0.25]], dtype=torch.float32))
|
||
|
||
# Detay ayrıştırıcı
|
||
self.extract_feature = nn.Conv2d(3, 16, 3, padding=1, bias=False)
|
||
# Doku koruyucu
|
||
self.refiner = nn.Conv2d(16, 16, 3, padding=1, bias=False)
|
||
# Artifactleri temizle
|
||
self.reconstructor = nn.Conv2d(16, 3, 3, padding=1, bias=False)
|
||
|
||
# DLAANet with residual learning and orthogonal weights
|
||
self._init_weights()
|
||
|
||
def _init_weights(self):
|
||
nn.init.orthogonal_(self.extract_feature.weight)
|
||
nn.init.orthogonal_(self.refiner.weight)
|
||
nn.init.dirac_(self.reconstructor.weight)
|
||
|
||
def forward(self, x):
|
||
features = F.leaky_relu(self.extract_feature(x), 0.2)
|
||
refined = F.leaky_relu(self.refiner(features), 0.2)
|
||
residual = self.reconstructor(refined)
|
||
return x + residual
|
||
|
||
|
||
class _TAAState:
|
||
def __init__(self):
|
||
self.history = None
|
||
self.frame_id = 0
|
||
|
||
def reset(self):
|
||
self.history = None
|
||
self.frame_id = 0
|
||
|
||
def update(self, frame, alpha, sensitivity):
|
||
if self.history is None or self.history.shape != frame.shape:
|
||
self.history = frame.clone()
|
||
self.frame_id += 1
|
||
return frame
|
||
|
||
# Neighborhood clamping for stability
|
||
local_min = -F.max_pool2d(-frame, kernel_size=3, stride=1, padding=1)
|
||
local_max = F.max_pool2d(frame, kernel_size=3, stride=1, padding=1)
|
||
history_clipped = torch.clamp(self.history, local_min, local_max)
|
||
|
||
# Motion masking logic
|
||
diff = torch.abs(frame - history_clipped).mean(dim=1, keepdim=True)
|
||
motion_mask = torch.sigmoid((diff - sensitivity) * 45.0)
|
||
dynamic_alpha = alpha * (1.0 - motion_mask)
|
||
|
||
out = torch.lerp(frame, history_clipped, dynamic_alpha)
|
||
self.history = out.detach()
|
||
self.frame_id += 1
|
||
return out
|
||
|
||
class VideoTAADLAA:
|
||
def __init__(self):
|
||
self.net_cache = {}
|
||
self.taa = _TAAState()
|
||
|
||
@classmethod
|
||
def INPUT_TYPES(cls):
|
||
return {
|
||
"required": {
|
||
"images": ("IMAGE",),
|
||
"taa_strength": ("FLOAT", {"default": 0.65, "min": 0, "max": 1, "step": 0.05}),
|
||
"taa_alpha": ("FLOAT", {"default": 0.40, "min": 0, "max": 0.9, "step": 0.01}),
|
||
"motion_sensitivity": ("FLOAT", {"default": 0.05, "min": 0.01, "max": 0.5, "step": 0.01}),
|
||
"jitter_scale": ("FLOAT", {"default": 0.02, "min": 0, "max": 0.08, "step": 0.01}),
|
||
"dlaa_strength": ("FLOAT", {"default": 0.35, "min": 0, "max": 1, "step": 0.05}),
|
||
"edge_threshold": ("FLOAT", {"default": 0.25, "min": 0, "max": 0.5, "step": 0.01}),
|
||
"blur_radius": ("INT", {"default": 0, "min": 0, "max": 5, "step": 1}),
|
||
"reset_history": ("BOOLEAN", {"default": False}),
|
||
}
|
||
}
|
||
|
||
RETURN_TYPES = ("IMAGE",)
|
||
FUNCTION = "execute"
|
||
CATEGORY = "CustomPostProcess"
|
||
|
||
def _net(self, device):
|
||
if device not in self.net_cache:
|
||
self.net_cache[device] = _DLAANet().to(device).eval()
|
||
return self.net_cache[device]
|
||
|
||
def jitter(self, x, idx, scale, net):
|
||
if scale < 1e-5: return x
|
||
off = net.jitter_offsets[idx % 4]
|
||
B, C, H, W = x.shape
|
||
|
||
theta = torch.eye(2, 3, device=x.device).unsqueeze(0).repeat(B, 1, 1)
|
||
theta[:, 0, 2], theta[:, 1, 2] = off[0] * scale / W, off[1] * scale / H
|
||
|
||
grid = F.affine_grid(theta, x.shape, align_corners=False)
|
||
return F.grid_sample(x, grid, mode="bilinear", padding_mode="reflection", align_corners=False)
|
||
|
||
def edge_aa(self, x, thr, blur, net):
|
||
if blur <= 0: return x
|
||
|
||
# Grayscale for edge detection
|
||
gray = 0.299*x[:, 0:1] + 0.587*x[:, 1:2] + 0.114*x[:, 2:3]
|
||
sx, sy = F.conv2d(gray, net.sobel_x, padding=1), F.conv2d(gray, net.sobel_y, padding=1)
|
||
edge = torch.sqrt(sx*sx + sy*sy + 1e-6)
|
||
|
||
mask = torch.sigmoid((edge - thr) * 15.0)
|
||
blurred = F.avg_pool2d(F.pad(x, [blur]*4, mode="reflect"), blur*2+1, stride=1)
|
||
|
||
return x*(1-mask) + blurred*mask
|
||
|
||
def execute(self, images, taa_strength, taa_alpha, motion_sensitivity, jitter_scale, dlaa_strength, edge_threshold, blur_radius, reset_history=False):
|
||
device = mm.get_torch_device()
|
||
if reset_history: self.taa.reset()
|
||
net = self._net(device)
|
||
B, H, W, C = images.shape
|
||
out_list = []
|
||
|
||
with torch.no_grad():
|
||
for i in range(B):
|
||
# NCHW conversion
|
||
img = images[i:i+1].to(device).permute(0, 3, 1, 2).float()
|
||
rgb = img[:, :3]
|
||
|
||
# 1. Spatial Preprocessing
|
||
rgb = self.jitter(rgb, self.taa.frame_id, jitter_scale, net)
|
||
rgb = self.edge_aa(rgb, edge_threshold, blur_radius, net)
|
||
|
||
# 2. Temporal Pass
|
||
taa_out = self.taa.update(rgb, taa_alpha, motion_sensitivity)
|
||
rgb = torch.lerp(rgb, taa_out, taa_strength)
|
||
|
||
# 3. Post-Sharpen
|
||
if dlaa_strength > 0:
|
||
# Modelin etkisini dlaa_strength ile sınırla
|
||
refined_output = net(rgb)
|
||
rgb = rgb + (refined_output - rgb) * dlaa_strength
|
||
|
||
# Kontrast/clarity artırımı
|
||
luma = 0.299 * rgb[:, 0:1, :, :] + 0.587 * rgb[:, 1:2, :, :] + 0.114 * rgb[:, 2:3, :, :]
|
||
mean_luma = torch.mean(luma, dim=(1, 2, 3), keepdim=True)
|
||
|
||
rgb = (rgb - mean_luma) * 1.02 + mean_luma
|
||
rgb = torch.clamp(rgb, 0.0, 1.0)
|
||
|
||
out_list.append(rgb.permute(0, 2, 3, 1).cpu())
|
||
|
||
if i % 10 == 0: mm.soft_empty_cache()
|
||
|
||
return (torch.cat(out_list, dim=0),)
|
||
|
||
NODE_CLASS_MAPPINGS = {"VideoTAADLAA": VideoTAADLAA}
|
||
NODE_DISPLAY_NAME_MAPPINGS = {"VideoTAADLAA": "🎮 Video TAA + DLAA Anti-Aliasing"} |