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.
Hi, thank you for your great work on this repository.
We noticed that your code uses a fixed
max_sequence_lengthto support batch processing. However, in the diffusers implementation of Qwen Image Edit,max_sequence_lengthis 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.