Titan vs Nova Canvas image masks: which colour is edited

Mask conventions are where image APIs disagree, and a flipped mask edits everything except the area you wanted. Amazon's two image models on Bedrock give the same answer for inpainting, but the two pages describe different file rules.

A mask checker

This script flags the three mistakes that cause most failures: wrong size, gray pixels and an alpha channel. It does not tell you which colour a given service edits; that is the table above.

from PIL import Image
import numpy as np

def check_mask(src_path: str, mask_path: str) -> list[str]:
    src, mask = Image.open(src_path), Image.open(mask_path)
    problems = []
    if src.size != mask.size:
        problems.append(f"size {mask.size} != {src.size}")
    if "A" in mask.getbands():
        problems.append("mask has an alpha channel")
    px = np.asarray(mask.convert("L"))
    gray = int(((px != 0) & (px != 255)).sum())
    if gray:
        problems.append(f"{gray} gray pixels")
    return problems

print(check_mask("photo.png", "mask.png") or "mask ok")

On Sume

Only the two ChatGPT Image 2.5 variants take mask_url. The Sume docs do not state the colour convention for it, and the earlier post on the mask alpha channel covers what to send. Treat the first call as a test: make a mask with a clear left half, send it, and see which half changed, then diff with the numpy check.