Qwen-Image-Flash: Beyond Objective Design
For researchers working on accelerating text-to-image and image editing models, this work provides empirical insights into training recipe design for few-step distillation, though it is incremental in nature.
The paper revisits few-step distillation for visual generative models, focusing on training recipe factors (data composition, teacher guidance, task mixture) rather than just distillation objectives. Using Qwen-Image-2.0, they develop Qwen-Image-Flash, showing that principled training pipeline organization is crucial for effective few-step distillation.
Few-step distillation has become an effective strategy for accelerating advanced visual generative models, yet prior work has largely focused on distillation objectives. In this work, we revisit few-step distillation from a complementary perspective, focusing on the training recipe that critically shapes student performance. Using Qwen-Image-2.0 as a representative case, we systematically investigate three factors in unified text-to-image generation and instruction-guided image editing distillation: data composition, teacher guidance, and task mixture. Our empirical analysis reveals several non-obvious behaviors, which motivate the development of Qwen-Image-Flash. Overall, our results suggest that effective few-step distillation requires not only carefully designed objectives, but also principled organization of the broader training pipeline.