Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers

Harvey Mannering; Zhiwu Huang; Adam Prügel-Bennett

cs.CVOct 12, 2026

Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers

Authors: Harvey Mannering, Zhiwu Huang, Adam Prügel-Bennett

Organizations: University of Southampton

Abstract

Emotions are expressed on a continuum and are naturally described in the valence-arousal (VA) space. Existing VA-based image editors are typically training-based and domain-specific, limiting them to VA regions seen during training and making them brittle across datasets and domains. Meanwhile, despite producing high-quality edits, state-of-the-art foundation image editors are not continuous affective controllers: non-linear emotion representations in text encoders makes prompt interpolation poorly calibrated, so accurate and continuous VA control remains unreliable. To fill this gap, we introduce Barycentric Guidance, a training-free framework that turns pretrained foundation image editors into continuous affective controllers over VA, requiring no finetuning, extra data, or model changes. Our method maps any target VA point to simplex-based guidance weights, enabling VA controllability in three aspects: (i) accurate targeting of VA states, (ii) broad coverage across the VA plane, and (iii) continuous control. Across three face datasets, we improve VA target accuracy by 25-29%, expand VA-plane coverage by 28-50%, and increase expression diversity by 29-49%, over the strongest baseline. Unlike prior domain-specific methods, our training-free approach exploits the foundation model's inherent versatility, enabling generalization beyond human faces to animals, artworks, and complex scenes within a single framework. Code is available at https://github.com/harveymannering/BarycentricGuidance.

Figures & tables

Panel of Figure 1 — Figure 1: Valence-Arousal-based image editing results from our approach. (a) The VA model of emotion from Russell [1] (b) - (e) VA-based editing across a range of datasets including Manga Facial Expression [ 2 ] , AFHQ [ 3 ] , CelebA-HQ [ 4 ] , and EmoArt [ 5 ] . VA ranges [-1,1] with uniform steps.

Figure 2: PCA Performed on CLIP Text Embedddings.

Figure 3: Triangle Mesh created from VA Points.

Panel of Figure 4 — Figure 4: Colour editing comparison between prompt interpolation and barycentric guidance. Prompts: “Make this flower <colour>, keep everything else the same colour.”; model: FLUX.2 Klein 9B. Source image from EmoSet [ 51 ] .

Dataset	Model	VA Controllability				Image Quality/ Preservation
		RMSE $\downarrow$	VA Entropy $\uparrow$	VA Std $\uparrow$	$S^{*}_{2D}$ $\downarrow$	FID $\downarrow$	IS $\uparrow$	ID Sim. $\uparrow$	LPIPS $\downarrow$
AffectNet	GANmut †	-	0.616	0.47 / 0.38	-	36.89	3.90	0.76	0.128
	EmoStyle	0.610	0.423	0.46 / 0.31	0.147	36.72	2.81	0.31	0.276
	Ours	0.433	0.916	0.64 / 0.57	0.303	11.02	4.42	0.55	0.096
	Ours (Qwen)	0.431	0.813	0.55 / 0.49	0.278	10.67	4.71	0.72	0.062
FFHQ	GANmut †	-	0.687	0.50 / 0.36	-	73.20	4.49	0.71	0.282

Table 2: Facial Expression Editing comparison with other VA editing models. We treat Entropy, VA Std, RMSE, and

S^{*}_{2D}

as the primary controllability metrics. † : The method does not control VA directly; instead, it approximates VA editing using polar coordinates to steer emotion.

Panel of Figure 6 — Figure 6: A sweep of VA values around the circumference of the unit circle, where $\theta$ is the angle. GANmut, EmoStyle, and Our edits are shown across three datasets AffectNet (top), FFHQ (middle), and CelebA (bottom). Target VA values (shown below columns) apply to EmoStyle and Ours only.

Setting	VA Controllability				Image Quality / Preservation
	RMSE $\downarrow$	VA Entropy $\uparrow$	VA Std $\uparrow$	$S^{*}_{2D}$ $\downarrow$	FID $\downarrow$	IS $\uparrow$	ID Sim. $\uparrow$
Full Approach	0.433	0.916	0.64 / 0.57	0.303	11.02	4.42	0.55
$-$ Scaled CFG	0.439	0.906	0.59 / 0.51	0.305	11.10	4.56	0.61
$\leftrightarrows$ Bary. Guidance $\rightarrow$ Prompt Interp.	0.461	0.869	0.57 / 0.49	0.342	10.75	4.50	0.55
$-$ Face Morphing	0.451	0.756	0.56 / 0.50	0.377	8.94	4.62	0.66

Table 3: Ablation Study on AffectNet.

\leftrightarrows

: replaces Bary. Guidance with Prompt Interp.

Dataset	Method	VA Controllability				Image Quality / Preservation
		RMSE $\downarrow$	VA Entropy $\uparrow$	VA Std $\uparrow$	$S^{*}_{2D}$ $\downarrow$	FID $\downarrow$	IS $\uparrow$
EmoArt	Prompt Interp.	0.851	0.129	0.35 / 0.39	0.502	74.85	9.64
EmoArt	Ours	0.735	0.327	0.36 / 0.42	0.483	67.96	9.08
FindingEmo	EmotiCrafter	0.796	0.298	0.38 / 0.25	0.540	233.30	8.85
FindingEmo	Ours (klein)	0.783	0.163	0.40 / 0.21	0.497	221.51	10.68

Table 6: Non-face results. This includes VA-based editing of artworks on EmoArt [ 5 ] and VA-based generation of scenes using FindingEmo [ 64 ] .

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: Our approach can perform valence arousal based emotional image editing across a wide variety of domains. VA range [-1,1] with uniform step size for rows and columns.

Panel of Figure 10 — Figure 10: Qualitative results on AffectNet using Qwen Image Edit 2511 as our base model. VA range [-1,1] with uniform step size for rows and columns

Figure 11: Smoothness Metric Toy Example

Panel of Figure 12 — Figure 12: Comparison of facial expression intensity levels in the MEAD dataset, demonstrating the incremental progression from neutral to strong disgust.

Figure 13: On synthetic images, diversity of facial expression trades-off against ID similarity as measured by FaceNet and ArcFace.

Figure 14: Comparison of different morph step on AffectNet images. The morph step refers to the step at which we restart the diffusion process after face morphing is performed.

Panel of Figure 15 — Figure 15: Examples of compound emotions generated using two VA points with FLUX.2-dev used as a backbone.

Method	IS	ID Similarity
EmoStyle	2.85	0.666
Ours	3.40	0.686

Appendix

Table 7: Restricted edit strength magnitude experiment on EmoStyle and our approach. We compare 465 images with the same small edit strengths (less than 0.2 change in VA distance) the recalculate relevant metrics on this smaller subset. CelebA used for all source images.

Model	VA Controllability				Image Quality/ Preservation
	RMSE $\downarrow$	VA Entropy $\uparrow$	VA Std $\uparrow$	$S^{*}_{2D}$ $\downarrow$	FID $\downarrow$	IS $\uparrow$	ID Sim. $\uparrow$	LPIPS $\downarrow$
GPT-Image-2	0.530	0.327	0.52 / 0.35	0.402	9.38	4.71	0.84	0.066
Ours	0.433	0.916	0.64 / 0.57	0.303	11.02	4.42	0.55	0.096

Appendix

Table 8: Facial Expression Editing comparison between GPT-Image-2 and our approach.

Dataset	Entropy $\uparrow$	VA Std (V) $\uparrow$	VA Std (A) $\uparrow$	RMSE $\downarrow$
AffectNet	+48.7%	+36.2%	+50.0%	+29.0%
FFHQ	+29.5%	+28.0%	+41.7%	+26.4%
CelebA-HQ	+34.7%	+35.4%	+45.9%	+25.1%

Appendix

Table 9: Relative improvements over the strongest baseline. Numbers are derived from the results in Table 2 .

Panel of Figure 16 — Figure 16: Triangle meshes used for different domains. Nodes and their locations are derived from datasets for the specific domain, in this case, AffectNet, EmoArt, and FindingEmo.

	RMSE
Happy	0.054
Sad	0.149
Angry	0.306
Bitter	0.451
Droopy	0.825
At Ease	0.908

Appendix

Table 10: The VA accuracy of different emotion words when performing facial expression editing on AffectNet with FLUX.2-dev. Overall, 100 different emotion words were considered. Neither Barycentric Guidance face morphing were used here, just our standard prompt template, where only the emotion word is altered.

Panel of Figure 17 — Figure 17: 10 by 10 grids of our approach (backbone FLUX.2-dev) being used to edit an image from AffectNet. Our standard mesh used for all other facial expression editing results is shown in (a), whereas (b) shows editing results on the same mesh except with an additional anchor, “relaxed”, in the bottom right corner.

Panel of Figure 18 — Figure 18: PCA performed on different prompt templates on the CLIP ViT-L/14 model.

Panel of Figure 19 — Figure 19: PCA performed text embeddings from different text encoders. Prompt template used was “a portrait photo of a <emotion> face”.

Panel of Figure 20 — Figure 20: User study screenshots.

Panel of Figure 21 — Figure 21: Examples of image comparisons from our user study. Our edits are Edit A, Edit A, Edit A, and Edit B in (a), (b), (c), and (d), respectively.

Figure 22: Magnitudes of the valence-arousal (VA) edits shown in our user study, as measured by EmotiEffLib, for EmoStyle (left) and ours (middle). The plot of the right shows the magnitude of the VA edit, as measured by the distance between the source images VA values (measured by EmotiEffLib) and the target VA value that both editing models are aiming at.

Figure 23: Our approach on FFHQ. VA range [-1,1] with uniform step size for rows and columns.

Figure 24: EmoStyle on FFHQ. VA range [-1,1] with uniform step size for rows and columns.

Figure 25: Our approach on CelebA. VA range [-1,1] with uniform step size for rows and columns.

Figure 26: EmoStyle on CelebA. VA range [-1,1] with uniform step size for rows and columns.

Figure 27: Our approach on AffectNet. VA range [-1,1] with uniform step size for rows and columns.

Figure 28: EmoStyle on AffectNet. VA range [-1,1] with uniform step size for rows and columns.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 313 papers

Beyond Emotion Prompts: Fine-Grained Text-to-Image Generation Driven by Valence-Arousal-Dominance
Sep 21, 2026Minglang Li, Yueyue Fang, Xieping GaoConditional Image Generation Controllable Image Generation
College of Information Science and Engineering Hunan Normal University, Changsha, China
MooD: Perception-Enhanced Efficient Affective Image Editing via Continuous Valence-Arousal Modeling
May 4, 2026Xinyi Yin, Yiduo Wang, Tingqi Hu +6Valence-Arousal Modeling Image Editing
School of Cyber Science and Engineering, Zhengzhou University, Zhengzhou 450002, China · School of Computer Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China · School of Journalism and New Media, Xi’an Jiaotong University, Xi’an 710049, China +1
Scalable Minimal-Change Learning for Controllable Image Editing
Oct 5, 2026Shuo Chen, Fengming Huang, Yu Yao +2Text-Guided Image Editing Image Editing
Sydney AI Centre, The University of Sydney

Sep 21, 2026J/K move · Enter open · S save

Beyond Emotion Prompts: Fine-Grained Text-to-Image Generation Driven by Valence-Arousal-Dominance

Minglang Li, Yueyue Fang, Xieping Gao

College of Information Science and Engineering Hunan Normal University, Changsha, China

Conditional Image Generation Controllable Image Generation T2I Diffusion Models

Open paper AskPDF arXiv

Although text-to-image models can accurately depict subjects and scenes, creators still struggle to specify the fine-grained emotions an image should convey without rewriting its content description. Natural language can suggest emotions, but it offers no control scale with stable meanings and ordered intensities. We propose EMOTRANS, which transforms psychologically grounded valence-arousal-dominance (VAD) coordinates into generation conditions that are independent of the content text and modulated across denoising stages, making emotional style a finely adjustable creative variable. To support this goal, we construct EMOVAD, an art-painting dataset that pairs objective content descriptions with separately collected emotional ratings from multiple annotators. We also coordinate emotional expression and content preservation through dual-branch training with a shared model. Objective and human evaluations show that the framework improves the accuracy of three-dimensional emotion control and produces perceptible, orderable continuous changes while maintaining competitive text alignment and image quality. This work provides a practical emotion-driven approach to image generation that extends objective content depiction to fine-grained emotional adjustment.

Abstract

Image Editing

Explore similar work

Beyond Emotion Prompts: Fine-Grained Text-to-Image Generation Driven by Valence-Arousal-Dominance

MooD: Perception-Enhanced Efficient Affective Image Editing via Continuous Valence-Arousal Modeling

Scalable Minimal-Change Learning for Controllable Image Editing