LTX2 HDR reference preprocessing applies an unintended stride-aligned resize #15037

preprocess_reference_video_hdr is documented to normalize frames without resizing before HDR's resize/pad step. However, height=None, width=None still selects stride-aligned default dimensions while do_resize=True.

With stride 32, a 1080×1920 (H×W) reference targeting 1088×1920 is first resized to 1056×1920, compressing its height before padding. Inputs smaller than the stride can hit a zero-sized resize.

This reproduces with no model downloads:

import torch
from diffusers.pipelines.ltx2 import LTX2VideoHDRProcessor

video = torch.rand(1, 3, 45, 61, generator=torch.Generator().manual_seed(0))
processor = LTX2VideoHDRProcessor()
actual = processor.preprocess_reference_video_hdr(video, height=64, width=64)
expected = torch.nn.functional.pad(video * 2 - 1, (0, 3, 0, 19), mode="reflect")
expected = expected.permute(1, 0, 2, 3).unsqueeze(0)
torch.testing.assert_close(actual, expected, rtol=0, atol=1e-6)

The assertion fails on the tested revision.

The HDR helper already skips resizing when the target contains the source, so this example should only normalize and pad. A pipeline test with tiny components also confirms that the discrepancy changes output latents.

I have a local fix that disables this preliminary resize on a copy, leaving the original processor unchanged. Would this scope be suitable for a PR?

Environment: Diffusers 0.41.0.dev0, PyTorch 2.14.1, Transformers 5.19.0, Python 3.11.15; macOS arm64.