Model Features This ControlNet is added to 15 layer blocks and 2 refinement layer blocks. The model was trained from scratch for 70,000 steps using a dataset of 1 million high-quality images covering general and human-centric content. The training resolution was 1328, with BFloat16 precision, a batch size of 64, a learning rate of 2e-5, and a text dropout rate of 0.10. It supports various control conditions—including Canny, HED, depth, pose, and MLSD—and can be used just like a standard ControlNet. We found that using different step counts at varying strength levels impacts the realism and clarity of the results. For strength and step tests, please refer to "Scale Test Results". You can adjust control_context_scale for stronger control and better detail preservation. For better stability, we strongly recommend using detailed prompts. The optimal range for control_context_scale is between 0.65 and 0.90. Note on steps: As the control strength (control_context_scale value) increases, it is recommended to increase the number of inference steps accordingly to achieve better results and maintain generation quality. This may be because the control model has not yet been fully refined. It also supports inpainting mode.