Source-linked AI summary
Measured Sliders: Learning Continuous Controls from Differentiable Image Measurements
Yijia Chen, Boyu Wei, Xuanhua Yin
TL;DR
Diffusion sliders often lack predictable, measurable scales and advance knowledge of learnability or interference. Measured Sliders defines controls through differentiable image measurements, using observability testing, measurement-guided objectives, and decoded calibration. Across SDXL and FLUX.1-dev, the controls are ordered, selective, and composable, with strong lighting-direction and multi-attribute results.
Problem
Existing diffusion sliders derive axes from text or learned representations, leaving coefficient scales disconnected from observable image properties.
Method
Measured Sliders uses a shared measurement space for observability testing, target-and-non-target measurement supervision, multi-branch LoRA learning, and decoded calibration.
Results
Across SDXL and FLUX.1-dev, controls are ordered, selective, and composable; lighting direction reaches ρ = 0.995 with 98.9% monotone sweeps, while five-attribute selectivity is 2.59 versus 1.50 for the strongest baseline average.
Takeaways & Limitations
Image-space measurement provides a common basis for learning, diagnosing, calibrating, and composing continuous generative controls.
Abstract
from arXiv · showhide
Continuous sliders are useful only when coefficient changes produce predictable image changes. Yet most diffusion sliders derive their axes from text or learned representations, leaving their scales disconnected from observable image properties. Consequently, we cannot tell in advance which attributes are learnable, compare control strengths directly, or anticipate interference when multiple controls are combined. We propose Measured Sliders, a framework that defines continuous controls through closed-form differentiable image measurements. A common measurement space unifies the pipeline. Before training, an observability test identifies usable supervision. During training, a measurement-guided objective learns target movement while suppressing non-target changes. After training, decoded calibration expresses controls in comparable units of realized image change. Multiple LoRA branches are stored in one checkpoint and composed without training on joint activations. Across SDXL and FLUX.1-dev, the resulting controls are ordered, selective, and composable. On 553 prompts, lighting direction reaches rho = 0.995 and 98.9% monotone sweeps. A five-attribute checkpoint achieves average selectivity 2.59, compared with 1.50 for the strongest baseline, and preserves every requested direction in 96.7% of pair and 86.1% of triple compositions. The observability test also separates every subsequently successful measurement from the failed candidate. Overall, image-space measurement provides a common basis for learning, diagnosing, calibrating, and composing continuous generative controls.
1 Introduction
Measured Sliders address the lack of predictable, measurable continuous control by defining LoRA branches through differentiable image measurements. The framework tests observability before training, suppresses non-target changes during learning, and calibrates decoded effects afterward.
- Motivation: Text-defined sliders can name edits such as “slightly brighter” without specifying the image-space magnitude of change.Scene-level attributes also require coordinated image changes while preserving the scene.
- Motivation: Raw coefficients lack predictable ordering, shared output units, and direct guarantees of measured non-target invariance.These gaps complicate learning, evaluation, and comparisons across controls.
- Framework: Measured Sliders convert differentiable image measurements into reusable continuous controls with pre-training observability, measurement-guided learning, and decoded calibration.Each LoRA branch targets one measurement while suppressing changes in the others, and branches share one checkpoint.
- Results: On 553 prompts, lighting direction reaches ρ = 0.995 with 98.9% monotone sweeps and 3.3× the selectivity of the strongest external slider.The evaluation covers decoded effect, drift, ordering, consistency, and zero-shot composition.
2 Related Work
Prior work learns continuous controls from parameters, representations, prompts, examples, or differentiable guidance, and increasingly supports multiple attributes. Existing evaluation measures edit response and preservation, while Measured Sliders targets the cross-axis relation between coefficients and measured effects.
- Control representations: Concept Sliders and related methods learn directions in parameter, text-representation, feature, or editing spaces from prompts, paired images, or instructions.These approaches provide different mechanisms for discovering and traversing continuous controls.
- Composability: Unified and multi-attribute methods store or compose several controls, while other systems regulate competing inputs or improve preservation.Examples include sparse text-embedding directions, conditional priors, unified adapters, and causal or context-aware representations.
- Differentiable guidance: Differentiable output objectives steer diffusion through auxiliary gradients, general guidance functions, pretrained energies, or joint reward and distributional constraints.These methods motivate controlling generation directly through differentiable image-space objectives.
- Evaluation: Continuous-control benchmarks measure response, monotonicity, preservation, saturation, and interference across edits and attributes.Measured Sliders addresses the additional question of whether coefficients from different controls correspond to comparable measured effects.
3 Method
Measured Sliders builds continuous controls in a shared image-measurement space, using observability screening, measurement-supervised LoRA training, and decoded calibration. The resulting interface maps controls to comparable image-space effects and supports branch composition without retraining.
- Observability: The pipeline first tests whether preview measurements resolve decoded natural variation, retaining candidates with a usable optimization signal.A candidate is retained when its pre-training observability ratio exceeds 1.
- Measurements: Closed-form measurements represent attributes such as brightness, saturation, warmth, contrast, tint, and lighting direction in a common vector.Preview and decoded implementations share the same measurement panel, while brightness uses a differentiable threshold approximation in preview space.
- Multi-Control Learning: All branch-specific LoRA updates are optimized together and stored in one checkpoint while retaining separate factors.The control vector activates individual branches or multiple branches simultaneously without requiring joint-activation retraining.
- Measurement-Supervised Learning: Each branch trains toward movement in its active measurement while assigning zero targets to all other measurements.Normalization prevents measurements with larger units from dominating the joint objective; optional anchor and color-balance terms preserve content without redefining the measurement target.
- Decoded Calibration: Decoded calibration converts branch-specific coefficient ladders into comparable image-space effect units using responses on a fixed calibration set.The method uses running maxima, earliest interpolated crossings, clipping, and no weight changes or extra denoising steps.
- Evaluation: Table 1 evaluates lighting-direction control on 553 GenEval prompts using decoded-image statistics and reports monotonicity as a percentage.Effect and drift use fixed backbone-specific natural ranges and are not compared across panels.
4 Experiments
Experiments show that measurement supervision produces ordered, selective controls, while observability predicts which measurements are learnable and decoded calibration aligns effect magnitudes. A single checkpoint supports selective, coherent, and zero-shot composable multi-control behavior, although strict composition criteria expose remaining failures.
- Main results: Measurement supervision raises FLUX lighting response by 295×, with ρ increasing from 0.111 to 0.998 and monotone sweeps from 9.6% to 99.8%.The matched comparison replaces text prediction-difference supervision with the measurement target.
- Main results: On SDXL, selectivity reaches 1.761, a 3.3× gain over the strongest external slider, with ρ = 0.995 and 98.9% monotone sweeps.The selectivity 95% confidence interval is [1.692, 1.832].
- Observability and calibration: Observability ratios of 2.82–9.13 yield mean decoded ρ ≥0.967, while sharpness at SNRk = 0.419 produces ρ = −0.232.The result was predicted before adapter training; five retained measurements exceed both observability and learned-control thresholds.
- Observability and calibration: Decoded calibration reduces endpoint dispersion from 63.2× with one preview-scaled coefficient to 2.2× on the unseen split without changing model weights.Calibration uses four separate contents to express controls in comparable realized-effect units.
- Multi-control composition: A five-branch checkpoint reaches average selectivity 2.59 versus 1.50 for the strongest baseline and preserves all requested directions in 96.7% of pairs and 86.1% of triples.These zero-shot compositions require no joint-activation training; Figure 5 links the measured responses to coherent visual traversals from a shared neutral generation.
- Multi-control composition: Under the strict composition criterion, 45.2% of pairs and 15.1% of triples pass because directional success does not ensure adequate magnitude, interference, or consistency.The criterion requires correct signs, minimum effects, bounded target error and inactive drift, and DINO similarity of at least 0.8.
5 Conclusion
Measured Sliders grounds continuous controls in differentiable image measurements shared by observability, learning, and calibration. Across SDXL and FLUX.1-dev, it provides ordered lighting control, calibrated multi-attribute control, and zero-shot composition, while content-adaptive calibration and broader measurement panels remain open directions.
- Conclusion: Measured Sliders uses differentiable image measurements as a shared basis for observability, learning, and calibration.The framework yields ordered lighting control and a five-branch calibrated appearance checkpoint across SDXL and FLUX.1-dev.
- Conclusion: The pre-training test predicts all six learnability outcomes, while calibration reduces endpoint dispersion by nearly 29×.Decoded units also separate control quality from LoRA scale and reveal higher-order branch interference.
- Conclusion: Content-adaptive calibration and broader measurement panels remain open directions.