Measured Sliders: Learning Continuous Controls from Differentiable Image Measurements

The researchers proposed the Measured Sliders framework, aimed at defining continuous control through closed-form differentiable image measurements. This framework achieved ordered, selective, and composable control effects on the SDXL and FLUX.1-dev models. Experimental data showed that in 553 prompt word tests, the monotonicity of the lighting direction reached rho = 0.995, and the average selection rate for five attribute checkpoints was 2.59, which is better than the baseline of 1.50. The framework utilized strategies such as pre-training observable filtering, measurement-based guidance to suppress non-target changes during training, and post-training decoding calibration to express control as comparable image change units. Additionally, multiple LoRA branches could be stored at the same checkpoint and used in combination without joint activation, maintaining 96.7% and 86.1% required direction accuracy in paired and triple combinations, respectively.