Repository navigation
InpaintProcessor.preprocess returns a bare tensor when no mask is passed, breaking its own 3-value contract #14470
Description
Describe the bug
InpaintProcessor.preprocess returns (image, mask, postprocessing_kwargs) on every path except the mask is None early return (image_processor.py#L895-L897), which returns the processed image on its own. Unpacking the three values the rest of the method promises then raises:
ValueError: not enough values to unpack (expected 3, got 1)
The return annotation is tuple[torch.Tensor, torch.Tensor], which no path actually returns, every masked path returns three values.
Both the early return and the three-value contract were introduced in the same commit (f50b18eec, #12220), so this looks like the early return simply not being updated when postprocessing_kwargs was added, rather than an intentional two-shape API.
Both in-repo call sites unpack three values (modular_pipelines/qwenimage/encoders.py#L1144 and #L1236), but their mask_image input is required=True, so they never reach the branch. It is reachable through direct use of InpaintProcessor, which is exported from diffusers and documented in docs/source/en/api/image_processor.md.
Reproduction
import numpy as np
import PIL.Image
from diffusers.image_processor import InpaintProcessor
image = PIL.Image.fromarray(np.zeros((64, 64, 3), dtype=np.uint8))
mask = PIL.Image.fromarray(np.zeros((64, 64), dtype=np.uint8))
processor = InpaintProcessor()
print(len(processor.preprocess(image, mask=mask, height=64, width=64))) # 3
print(type(processor.preprocess(image, height=64, width=64))) # <class 'torch.Tensor'>
img, msk, kwargs = processor.preprocess(image, height=64, width=64)Output:
3
<class 'torch.Tensor'>
Traceback (most recent call last):
...
ValueError: not enough values to unpack (expected 3, got 1)
Logs
System Info
- 🤗 Diffusers version: 0.40.0.dev0 (
main) - Platform: macOS-26.6-arm64-arm-64bit-Mach-O
- Python version: 3.13.7
- PyTorch version (GPU?): 2.13.0 (False)
- Huggingface_hub version: 1.27.0
- Safetensors version: 0.8.0
- Accelerator: Apple M3
- Using GPU in script?: No
- Using distributed or parallel set-up in script?: No
Activity
xquantize commentedon Aug 14, 2026
I dug into this and can confirm the inconsistency. The
mask is Noneearly return yields a single tensor while every masked path returns(image, mask, postprocessing_kwargs)and both in-repo call sites (qwenimage/encoders.py) unpack three, the-> tuple[torch.Tensor, torch.Tensor]annotation matches neither. It traces to #12220 (f50b18eec), where the early return wasn't updated whenpostprocessing_kwargswas added.I have a fix plus a regression test ready: align the
mask=Nonepath to the same 3-tuple,mask=None,postprocessing_kwargswithNonevalues (mirroring the existing no-crop branch) and correct the annotation totuple[torch.Tensor, torch.Tensor | None, dict]. Since this changes the public return shape on themask=Nonepath I wanted to confirm you're happy with that direction before opening a PR, happy to instead keep two shapes and only fix the annotation if you'd prefer.lindicaphxag-tech commentedon Sep 19, 2026
PR #14807 follows the return-contract fix proposed here: the no-mask path returns the same three-value shape (
image,None, postprocessing kwargs) as the masked path, and the annotation/regression test are updated. The repository checks for the PR are green. Could a maintainer assign/review it?fix(image_processor): maintain 3-tuple return contract in InpaintProc…