gmqewn_qwen_lightning_4step_comp: comparison between GM-Qwen and Qwen lightning, 4 steps. 
For each prompt there are two generations, our model has much better diversity than DMD-based QWen-lightning. 

gmflux_hyperflux_fluxturbo_8step_comp: comparison between GMFlux, HpyerFlux and FluxTurbo, 8 steps. 
Our model has better fine-grained details than HyperFlux and better text rendering than FluxTurbo.

gmflux_senseflow_4step_comp: comparison between GMFlux and SenseFlow 4 steps. Our model has better 
structures, more faithful styles and better diversity. 

For each comparison, we also include the results of the teacher model (last column) as reference.
 