This repository provides the LoRA finetune codes of Qwen-Image.

• Superior Text Rendering: Qwen-Image excels at complex text rendering, including multiline layouts, paragraph-level semantics, and fine-grained details. It supports both alphabetic languages (e.g., English) and logographic languages (e.g., Chinese) with high fidelity.

• Consistent Image Editing: Through our enhanced multi-task training paradigm, Qwen-Image achieves exceptional performance in preserving both semantic meaning and visual realism during editing operations.

• Strong Cross-Benchmark Performance: Evaluated on multiple benchmarks, Qwen-Image consistently outperforms existing models across diverse generation and editing tasks, establishing a strong foundation model for image generation.

• Qwen-Image (Text-to-Image Model)
  • LoRA finetune

Clone the repo:

git clone https://github.com/mindspore-lab/mindone.git
cd mindone/examples/diffusers/qwenimage

Download Model Weights:

# Download from HuggingFace
hf download Qwen/Qwen-Image

export ASCEND_RT_VISIBLE_DEVICES=0,1
NPUS=2
MASTER_PORT=9000
LOG_DIR=outputs/lora
msrun --bind_core=True --worker_num=${NPUS} --local_worker_num=${NPUS} --master_port=${MASTER_PORT} --log_dir=${LOG_DIR}/parallel_logs \
python finetune_lora_with_mindspore_trainer.py \
    --output_dir ${LOG_DIR} \
    --num_train_epochs 1 \
    --learning_rate 1e-5 \
    --save_strategy no \
    --bf16