Bug: Loading diffusers-trained Qwen-Image-Edit weights causes unexpected behavior #20
Description
Hi, thank you for your great work on this repository.
We noticed that your code uses a fixed max_sequence_length to support batch processing. However, in the diffusers implementation of Qwen Image Edit, max_sequence_length is determined dynamically based on the maximum token length.
This discrepancy leads to unexpected results when attempting to load weights trained with diffusers using your codebase.