Fix float truncation in Flux2ImageProcessor._resize_to_target_area

`_resize_to_target_area` scaled the image with `int(side * scale)`. For any
input whose scaled side should be an exact integer (every square image, for
one) `n * sqrt(target / n**2)` evaluates to 1023.999... in floating point for
roughly 16% of sizes, so `int` truncated it to 1023. The Flux2, Flux2 Klein,
Flux2 modular and Z-Image Omni pipelines then floor to a multiple of 16, so the
conditioning image and default output silently became 1008x1008 instead of
1024x1024 (e.g. for a 1039x1039 input).

Use `round` instead. It may exceed `target_area` by at most one pixel per side,
which every caller floors away.

Author: rootonchair

src/diffusers/pipelines/flux2/image_processor.py

@@ -109,8 +109,11 @@ def _resize_to_target_area(image: PIL.Image.Image, target_area: int = 1024 * 102
         image_width, image_height = image.size
 
         scale = math.sqrt(target_area / (image_width * image_height))
-        width = int(image_width * scale)
-        height = int(image_height * scale)
+        # Use `round` rather than `int`: truncation turns floating-point noise (e.g. 1023.999...) into a lost pixel,
+        # which callers then floor to a smaller multiple of the VAE scale factor. Rounding may exceed `target_area` by
+        # at most one pixel per side, which every caller floors away.
+        width = round(image_width * scale)
+        height = round(image_height * scale)
 
         return image.resize((width, height), PIL.Image.Resampling.LANCZOS)