Side-by-side images comparing our method (left) with Flux.dev (right) demonstrate the performance of both approaches under the same Denoise Steps. The image file names follow the format: iteration_prompt_index_DenoiseSteps.png.
The prompts used for generating these images are as follows:
prompt_2: Corporate decisions about shining diamonds gemstones covering the boardroom with light color and jellyfish happy smiling party Muppet
prompt_3: Land of mushrooms, surreal objects, mystical clouds, marbling art style

To facilitate real-time monitoring of the path compression effect during training, we conduct a random test every 1,000 iterations. The number of denoising steps for these tests is randomly selected from 1 to 10. From these test results, it can be observed that our results become increasingly clear and accurate compared to the original Flux.