GitHub Repo
Transformers Community Reports DeepSeek-VL Padding Color Mismatch; Image Processors May Ignore Saved Settings
A Transformers community report on October 15 says Janus and DeepSeek-VL image processors may ignore specified padding colors, causing inputs for some non-square images to differ from the original implementation. The report includes a patch and test results, but there is no official fix release yet, and the impact on model accuracy has not been measured.
On October 15, the Transformers community reported an image preprocessing consistency issue: some processors for Janus, DeepSeek-VL, and DeepSeek-VL-Hybrid recalculate the padding color, ignoring settings passed by the caller or saved in the checkpoint. The report reproduced the issue on that day’s main branch in a CPU environment, and noted that the default behavior may also differ between the PIL and torchvision backends. Issue report
A specific example involves deepseek-vl-7b-chat. DeepSeek’s public configuration specifies a background_color of RGB (122, 116, 104) and an image size of 1024. After loading a community-converted version, the reporter found that the low-resolution branch instead used (127, 127, 127), while the high-resolution branch retained the specified color. As a result, the same image requiring padding can produce different pixel inputs along different processing paths. Original model configuration
The original DeepSeek code derives the padding color from the image mean and pads rectangular images into squares; images that are already square are returned unchanged. This difference therefore primarily affects padded regions, and does not imply that all images are affected. The Transformers documentation also confirms that the Hybrid architecture uses SigLIP and SAM for low- and high-resolution information, respectively, making preprocessing consistency something to check when comparing implementations. Original code, architecture documentation
The accompanying patch prioritizes an explicitly specified or saved color, deriving one from processor attributes only when no color is provided. The author says 69 related tests passed, but did not run the slow 7B integration tests that require a GPU. As of this report, the issue remains open, with no linked pull request listed. Patch and validation details
For engineers porting models or reproducing evaluations, this offers a concrete checkpoint: in addition to pinning weights, compare preprocessing tensors, backend selection, and results after saving and reloading the configuration. Follow-up should track an official fix and end-to-end tests on non-square images; current evidence is insufficient to claim a substantial drop in recognition quality.