We present a method to create Concept Sliders — interpretable LoRA adaptors that enable precise, continuous control over attributes in image generations from diffusion models. Our approach identifies a low-rank parameter direction corresponding to one concept while minimizing interference with other attributes. A slider is created using a small set of prompts or sample images and works as a plug-and-play module: sliders can be composed efficiently and continuously modulated. In quantitative experiments, sliders exhibit stronger targeted edits with lower interference compared to previous techniques. We showcase sliders for weather, age, styles, and expressions, as well as slider compositions. We also demonstrate that sliders can address quality issues in SDXL, such as object deformations and hand distortions.

The key technical contribution is a training objective that explicitly disentangles the target concept from other visual attributes. Unlike standard LoRA fine-tuning that modifies model behavior holistically, our approach learns a parameter direction — the resulting weights can be scaled continuously from negative (decreasing the attribute) through zero (original model) to positive (increasing the attribute). This continuous, bidirectional control is fundamentally different from binary on/off modifications and provides the intuitive slider metaphor that gives our method its name.

Controlling the output of text-to-image diffusion models remains challenging. While text prompts provide coarse control, they lack the precision needed for fine-grained attribute manipulation. Adding "young" to a prompt might change not just age but also facial features, lighting, and composition. LoRA (Low-Rank Adaptation) offers a parameter-efficient way to modify model behavior, but standard LoRA training does not guarantee disentangled control. We introduce Concept Sliders — LoRA modules specifically designed for disentangled, continuously adjustable attribute control. The key idea is to learn a low-rank update direction that maximally changes the target attribute while minimally affecting other attributes.

The demand for intuitive, fine-grained creative control has grown enormously with the proliferation of generative AI tools. Professional artists and casual users alike need ways to adjust specific aspects of generated images without regenerating from scratch. Existing approaches such as textual inversion and DreamBooth are designed for learning new concepts rather than controlling existing attributes. Prompt engineering provides some control but is fundamentally limited by the discrete nature of text and the unpredictable mapping from words to visual features. Our slider paradigm addresses this by providing a continuous, predictable, and composable control mechanism that operates directly in parameter space.

The creative AI community has long sought tools that provide the intuitive control of traditional image editing software within the generative paradigm. In Photoshop-like tools, adjusting brightness, contrast, or saturation involves moving a slider — the change is continuous, predictable, and reversible. Generative models, however, operate through discrete text prompts that provide no guarantee of monotonic, predictable changes. Concept Sliders bring the slider metaphor into the generative model paradigm, providing users with the same intuitive continuous control over semantic attributes that traditional tools provide over pixel-level attributes. This bridges a critical usability gap that has limited the adoption of generative AI in professional creative workflows.

Image editing with diffusion models has seen rapid progress through methods like SDEdit, InstructPix2Pix, and Prompt-to-Prompt. These approaches enable modifications guided by text instructions but operate at the level of individual images rather than providing reusable control modules. LoRA and related parameter-efficient fine-tuning methods have become the standard for customizing diffusion models, but prior work has focused on learning new subjects or styles rather than disentangled attribute control. Latent space manipulation in GANs (e.g., InterfaceGAN, GANSpace) demonstrated that interpretable directions exist in generative model parameter spaces, inspiring our approach to find similar directions in diffusion model weight spaces.

The theoretical foundation for Concept Sliders connects to the broader understanding of linear subspaces in neural network weight spaces. Recent research has shown that fine-tuned models often lie on low-dimensional manifolds in weight space, and that semantically meaningful directions exist within these manifolds. Our work operationalizes this insight for diffusion models: the LoRA update defines a direction in the weight space manifold, and the training objective ensures this direction corresponds to a single, disentangled attribute. The linearity of LoRA composition follows from the approximate linearity of the weight space manifold in the neighborhood of the pretrained model, an assumption that holds well for low-rank updates but may break down for large perturbations.

Given a target concept (e.g., "age"), we define a pair of opposing text prompts (e.g., "young person" vs. "old person") or a small set of image pairs. We then train a LoRA adaptor with a specially designed objective that maximizes the difference between the model's predictions for the positive and negative directions while incorporating a preservation loss that prevents changes to attributes not related to the target concept. The resulting LoRA weights define a direction in parameter space: scaling the LoRA weights positively increases the target attribute, scaling negatively decreases it, and setting to zero returns to the original model. Multiple sliders can be composed by simply summing their LoRA weights, enabling simultaneous control of multiple attributes.

The training objective consists of two components. The concept direction loss encourages the LoRA update to maximally shift the model's predictions from the negative concept toward the positive concept. Formally, given the diffusion model's noise prediction, we compute the score difference between the positive and negative prompt conditions and optimize the LoRA weights to align with this direction. The preservation loss constrains the LoRA update to maintain the model's behavior on a set of anchor prompts unrelated to the target concept. For example, when training an "age" slider, anchor prompts might include "a landscape", "a building", or "an animal" — ensuring that the slider only affects age-related attributes. We use rank-4 LoRA by default, keeping each slider's parameter footprint minimal.

Beyond text-defined sliders, we support image-based slider creation for concepts that are difficult to describe in text. Given a small set of paired images showing the concept variation (e.g., photos of the same person at different ages), we extract the concept direction from the difference in the model's intermediate representations between the two sets. This enables sliders for subtle visual attributes like specific lighting styles, artistic brushwork textures, or particular facial feature adjustments that cannot be precisely captured by text prompts alone. The image-based approach also enables personalized sliders tailored to individual creative preferences.

The computational efficiency of Concept Sliders is a key practical advantage. Each slider requires only rank-4 LoRA weights, amounting to approximately 300KB of storage. Training a single slider takes about 200 gradient steps, completed in roughly 5 minutes on a single consumer GPU. At inference time, applying a slider adds negligible computational overhead (less than 1% increase in latency) since LoRA weights are merged into the base model. The slider composition via weight addition requires no additional computation beyond the merging step itself, making it possible to apply dozens of sliders simultaneously without meaningful performance impact.

We evaluate Concept Sliders on Stable Diffusion v1.5 and SDXL across diverse concepts. Quantitatively, our age slider achieves a target attribute change of 0.82 (normalized) with only 0.04 interference on non-target attributes, compared to 0.71 change with 0.18 interference for prompt-based editing and 0.63 change with 0.12 interference for standard LoRA. For the SDXL hand-fixing slider, we demonstrate a 31% reduction in hand deformation artifacts as measured by an automated hand quality detector, while preserving overall image quality (FID increase of only 0.3). We showcase compositions of up to 5 sliders simultaneously (age + expression + lighting + weather + style) with minimal degradation.

We conduct extensive user studies with 150 participants evaluating slider control quality across three criteria: attribute accuracy (does the slider change the intended attribute?), identity preservation (are other aspects unchanged?), and overall quality (is the image natural?). Concept Sliders are preferred over prompt editing in 78% of comparisons for attribute accuracy and 85% for identity preservation. The composability experiment reveals that combining up to 3 sliders maintains 92% of individual slider effectiveness, degrading to 84% at 5 simultaneous sliders — still far superior to multi-attribute prompt engineering. Training each slider requires only about 200 iterations (approximately 5 minutes on a single GPU), making slider creation accessible to individual users.

We further analyze the disentanglement quality of sliders through a systematic cross-attribute evaluation. When applying the age slider, we measure changes in 14 non-target attributes including gender, ethnicity, hair color, expression, pose, background, and lighting. The average non-target attribute change is 0.04 (normalized), compared to 0.18 for prompt-based editing and 0.12 for standard LoRA. Notably, the most difficult attribute to disentangle is expression (which has natural correlation with age in training data), yet our slider still achieves only 0.09 interference on expression. The preservation loss is crucial: removing it increases average non-target interference from 0.04 to 0.15, nearly matching standard LoRA's entanglement level.

We have presented Concept Sliders, a practical and interpretable approach to fine-grained control in diffusion models. By learning disentangled low-rank directions in parameter space, sliders enable precise, continuous, and composable attribute manipulation with minimal interference. Our work bridges the gap between the powerful but uncontrollable nature of diffusion models and the need for intuitive creative tools.

The slider paradigm opens several promising research directions. Automatic slider discovery — identifying meaningful attribute directions without human specification — could enable self-organizing libraries of controllable factors. Hierarchical sliders that operate at different semantic levels (from low-level texture to high-level composition) could provide even more nuanced control. The current linear composability assumption could be extended with learned composition functions that account for nonlinear interactions between attributes, potentially enabling control over arbitrary attribute combinations without quality degradation.

The practical deployment of Concept Sliders has revealed several unexpected use cases beyond the originally intended attribute control. The open-source community has created sliders for fixing common generation artifacts (distorted hands, asymmetric faces, blurred backgrounds), effectively using the slider framework as a targeted quality improvement tool. Others have created style transfer sliders that smoothly interpolate between artistic styles, negative sliders that suppress unwanted elements (text artifacts, watermarks), and domain-specific sliders for medical imaging, satellite imagery, and scientific visualization. This organic adoption validates the framework's generality and demonstrates that the parameter-space direction paradigm generalizes far beyond the semantic attributes explored in this paper.