Regressor-Guided Generative Image Editing Balances User Emotions to Reduce Time Spent Online

TL;DR

This work tackles non-coercive reduction of online engagement by editing images to modulate emotional responses. It develops three regressor-guided editing strategies—parametric optimization, style optimization, and RGDR diffusion resampling—each guided by a valence/arousal regressor anchored to the Circumplex Model of Affect, with RGDR enabling high-level content edits. Technical evaluation shows all approaches can move images toward a neutral, low-arousal target while preserving image fidelity, but RGDR uniquely achieves emotionally neutral perception without quality loss. A controlled image-rating study and a social-media experiment demonstrate that RGDR-edited images reduce perceived arousal and browsing duration, suggesting a practical, user-autonomy-preserving path for digital well-being interventions. The findings highlight a diffusion-based emotion-editing design space that balances emotional impact with perceptual fidelity and ethical deployment considerations.

Abstract

Internet overuse is a widespread phenomenon in today's digital society. Existing interventions, such as time limits or grayscaling, often rely on restrictive controls that provoke psychological reactance and are frequently circumvented. Building on prior work showing that emotional responses mediate the relationship between content consumption and online engagement, we investigate whether regulating the emotional impact of images can reduce online use in a non-coercive manner. We introduce and systematically analyze three regressor-guided image-editing approaches: (i) global optimization of emotion-related image attributes, (ii) optimization in a style latent space, and (iii) a diffusion-based method using classifier and classifier-free guidance. While the first two approaches modify low-level visual features (e.g., contrast, color), the diffusion-based method enables higher-level changes (e.g., adjusting clothing, facial features). Results from a controlled image-rating study and a social media experiment show that diffusion-based edits balance emotional responses and are associated with lower usage duration while preserving visual quality.

Figures (21)

• Figure 1: Style optimization: To adapt an input image $I$, its latent style vector $s$ is optimized to generate a similar image that elicits a reduced emotional response when decoded with $I$'s latent content $c_0 = E_c(I)$, resulting in $\hat{I} = D(c_0, s)$. The optimization process minimizes the distance between the predicted valence and arousal of the adapted image, $[\hat{v}, \hat{a}] = R(\hat{I})$, and the target values specified by emotional reference, $[v', a']$. Additionally, the optimization incorporates a term to ensure content consistency by minimizing the L1 loss between the encoded latent content of the generated image, $\hat{c} = E_c(\hat{I})$, and $c_0$. The style vector $s$ is initialized as the encoded style vector of the input image, $s_0 = E_s(I)$.
• Figure 2: Parametric optimization: To adapt an input image $I$, the parameters $p$ of the differentiable image transformations $T$ are optimized to produce a similar image that elicits a reduced emotional response, represented by $\hat{I} = T(I, p)$. The optimization minimizes the distance between the predicted valence and arousal of the adapted image, $[\hat{v}, \hat{a}] = R(\hat{I})$, and the emotional reference, $[v', a']$. It further ensures a high cosine similarity between the CLIP-space embeddings of the input image $c = CLIP(I)$ and the transformed image $\hat{c} = CLIP(\hat{I})$.
• Figure 3: Regressor-guided diffusion resampling: To adapt an input image $I$, it is first encoded into a latent vector $z_0 = E(I)$ and then inverted to its corresponding noise vector $z_T = DDIM_{\text{inv}}(z_0, t, \mathcal{C}, \beta)$, while generating unconditional text embeddings for each timestep $\{\emptyset_t\}_{t=1}^T = NTO([z_T, \ldots, z_0], \mathcal{C})$, ensuring editability of $\hat{z}_t$ and alignment with $I$. At each step of the denoising process, $\hat{z}_t$ is updated by blending the predicted unconditional and conditional noise $\epsilon_{\theta}$, further refined by a score $\nabla_{z_t}$ that quantifies alignment with the emotional reference. With the resulting noise vector $\Tilde{\epsilon}_{\theta}$, $\hat{z}_{t-1}$ can be obtained. The final latent vector $\hat{z}_0$ is decoded into the image $\hat{I} = D(\hat{z}_0)$, which closely resembles $I$ while modulating the emotional response.
• Figure 4: Criteria alignment analysis: The plots show how the metrics evolve as the weighting of the objective terms that introduce emotional changes in the images increases. In each plot, the upper x-axis represents the values of $w_2$ for parametric optimization and style optimization, while the lower x-axis indicates the values of $s$ for RGDR and CG. The y-axes show the metric values.
• Figure 5: Example result. Predicted valence–arousal values are shown top-right [target: (valence = 0.5, arousal = 0.0)].