fix: prevent clip_preprocess center crop from exceeding the resized image

[akhenakh]

Summary

clip_preprocess computed the aspect-preserving resize size by truncating scale * dim to int64_t. For some aspect ratios the single-precision result lands a hair below an integer (e.g. 736.0f / 730.0f * 730.0f == 735.999…), so the resized side is one pixel smaller than the crop target. The center crop then reads one element past the end of the resized tensor and throws an uncaught exception:

sd::Tensor error: Tensor index out of range: shape=[650, 735, 3, 1]
terminate called after throwing an instance of 'std::invalid_argument'
  what():  Tensor index out of range: shape=[650, 735, 3, 1]
SIGABRT: abort

Reproduction (Qwen3VL / Krea2 reference images): a second reference of 645×730 is snapped to w_bar=640, h_bar=736; scale = max(640/645, 736/730) = 1.008219, so resized_height = (int64_t)(scale * 730) = (int64_t)735.999… = 735, while the crop loops up to target_height = 736. Any caller of clip_preprocess can hit this (Qwen3VL/Krea2 vision encoder, CLIP vision, PhotoMaker) — it only depends on the image aspect ratio.

The fix rounds the resized dimensions up and clamps them to at least the crop target, so the crop window is always inside the resized image:

int64_t resized_width  = static_cast<int64_t>(std::ceil(scale * static_cast<float>(image.shape()[0])));
int64_t resized_height = static_cast<int64_t>(std::ceil(scale * static_cast<float>(image.shape()[1])));
resized_width  = std::max<int64_t>(resized_width, target_width);
resized_height = std::max<int64_t>(resized_height, target_height);

Related Issue / Discussion

Additional Information

Verification (exact failing setup):

• Model: Krea-2-Turbo-Q8_0.gguf + Qwen3VL text encoder, --backend vulkan1,te=cpu
• Args: -W 960 -H 720 --steps 10 --cfg-scale 1 --img-cfg-scale 1 --flow-shift 3
• Refs: -r ref0.png (960×720, i.e. 720×960 H×W) -r ref1.png (645×730) with --ref-image-args preset=krea2_ostris_edit
• Prompt: an edit instruction with two Picture N: references.

Before the fix:

[INFO] image.cpp:271  - EDIT mode
[INFO] conditioner resize ... ref image 1 from 730x645 to 736x640
sd::Tensor error: Tensor index out of range: shape=[650, 735, 3, 1]
[abort]

After the fix (same command, same refs):

[INFO] image.cpp:516  - get_learned_condition completed, taking 26.60s
[INFO] image.cpp:886  - sampling completed, taking 100.81s
[INFO] main.cpp:504  - save result image 0 to '...png' (success)
exit 0

• Platform: Arch Linux, RX 7900 XTX (RADV), Mesa 26.2.2, Vulkan 1.4.354.
• Formatting: clang-format clean (.clang-format, Chromium base).

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[leejet]

I think the crash fix is valid, but using ceil here changes the resize result for inputs that were previously handled correctly.

For example, in the reported 645x730 -> 640x736 case, only the resized height needs to be clamped from 735 to 736. Changing to ceil also changes the width from 650 to 651, which unnecessarily alters the CLIP preprocessing result.

Could we keep the existing truncation behavior and only enforce the crop invariant instead?

int64_t resized_width =
    static_cast<int64_t>(scale * static_cast<float>(image.shape()[0]));
int64_t resized_height =
    static_cast<int64_t>(scale * static_cast<float>(image.shape()[1]));

resized_width = std::max<int64_t>(resized_width, target_width);
resized_height = std::max<int64_t>(resized_height, target_height);

This should fix the out-of-bounds issue while minimizing behavior changes for existing inputs.

[akhenakh]

thanks for catching the width regression. Updated to the clamp-only version