Qwen-Image-Flash: Beyond Objective Design
Abstract
Few-step distillation for visual generative models benefits from systematic investigation of training recipes beyond just distillation objectives, leading to improved student performance through optimized data composition, teacher guidance, and task mixture.
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.
Community
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.
Guys, you're legends, but why are you still keeping your new Qwen Image models closed? Even Ideogram has already released weights, and Krea 2 is expected to go open soon.
You did such an incredible job with the first generations of Qwen Image. In many ways, it's still one of the best image models out there - especially when it comes to character training and consistency. But right now, it feels like you're losing ground for no good reason.
It's genuinely disappointing that your management doesn't seem to understand this ((.
Wishing you all the best and good luck with everything.
i find your comment a bit respect less, after all they released so many totally mind blowing models for free already!
you wont see this happening from companies like google or xAi etc!
Get this paper in your agent:
hf papers read 2606.03746 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper