Thomas HollierandClaude Opus 4.6
b3fac23393
9 ComfyUI nodes for 360 panoramic image generation using FLUX.1-dev with DiT360 LoRA. Supports text-to-panorama, inpainting, and outpainting via RF-Inversion with PersonalizeAnything attention. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
69 lines
2.2 KiB
Python
69 lines
2.2 KiB
Python
"""Equirectangular projection utilities for panoramic images."""
|
|
|
|
import torch
|
|
import math
|
|
from typing import Tuple, Literal
|
|
|
|
|
|
def validate_aspect_ratio(width: int, height: int, tolerance: float = 0.01) -> bool:
|
|
"""Check if dimensions have a 2:1 ratio (equirectangular requirement)."""
|
|
ratio = width / height
|
|
return abs(ratio - 2.0) < tolerance
|
|
|
|
|
|
def get_equirect_dimensions(width: int, alignment: int = 16) -> Tuple[int, int]:
|
|
"""Calculate valid equirectangular dimensions with proper alignment."""
|
|
width = (width // alignment) * alignment
|
|
height = width // 2
|
|
return width, height
|
|
|
|
|
|
def blend_edges(
|
|
image: torch.Tensor,
|
|
blend_width: int = 10,
|
|
mode: Literal["linear", "cosine", "smooth"] = "cosine",
|
|
) -> torch.Tensor:
|
|
"""Blend left and right edges for seamless wraparound.
|
|
|
|
Args:
|
|
image: (B, H, W, C) ComfyUI format tensor.
|
|
blend_width: Width of blend region in pixels.
|
|
mode: Blending curve type.
|
|
"""
|
|
B, H, W, C = image.shape
|
|
|
|
if blend_width <= 0 or blend_width >= W // 2:
|
|
return image
|
|
|
|
left_edge = image[:, :, :blend_width, :]
|
|
right_edge = image[:, :, -blend_width:, :]
|
|
|
|
if mode == "linear":
|
|
weights = torch.linspace(0, 1, blend_width, device=image.device)
|
|
elif mode == "cosine":
|
|
t = torch.linspace(0, math.pi, blend_width, device=image.device)
|
|
weights = (1 - torch.cos(t)) / 2
|
|
elif mode == "smooth":
|
|
weights = torch.linspace(0, 1, blend_width, device=image.device) ** 2
|
|
else:
|
|
raise ValueError(f"Unknown blend mode: {mode}")
|
|
|
|
weights = weights.view(1, 1, -1, 1)
|
|
|
|
blended_left = left_edge * (1 - weights) + right_edge * weights
|
|
blended_right = right_edge * (1 - weights) + left_edge * weights
|
|
|
|
result = image.clone()
|
|
result[:, :, :blend_width, :] = blended_left
|
|
result[:, :, -blend_width:, :] = blended_right
|
|
|
|
return result
|
|
|
|
|
|
def check_edge_continuity(image: torch.Tensor, threshold: float = 0.05) -> bool:
|
|
"""Check if left and right edges are continuous."""
|
|
left_edge = image[:, :, 0, :]
|
|
right_edge = image[:, :, -1, :]
|
|
diff = torch.abs(left_edge - right_edge).mean()
|
|
return diff.item() < threshold
|