Original title: AffectDelta: Beyond Emotion Labels for Image Editing
Xingzu Zhan, Lin Gu, Ruogu Fang · September 2026
Images often stir up a mix of feelings, not just one simple mood. Most image editing tools try to change an image using a single emotion word or written instructions. However, words cannot precisely say how much each feeling should rise or fall. To fix this, researchers built AffectDelta. Instead of one label, it tracks eight emotion scores at once. First, a frozen tool measures the feelings in the original picture. Then, AffectDelta calculates the exact difference needed to hit the target mood. An internal transition encoder and a diffusion backboneA type of image-generating neural network that starts from noise and gradually refines it into a clear picture, guided here by emotion signals. turn this difference into visual adjustments. The tool successfully shifts the emotional feel while keeping the original scene intact. It outperformed six other systems in tests, balancing accurate emotions with strong image preservation.
You're seasoning a soup that already has several flavors in it. The recipe says 'make it spicier,' but that one instruction doesn't tell you whether to also add more salt, leave the sweetness alone, or tone down the sourness. A single word like 'spicier' can't capture the full picture of what needs to change and by how much. In this paper, the editor works the same way—instead of translating a single emotion word into changes, it receives eight precise numbers, each one saying exactly how much to increase, decrease, or hold steady each emotional dimension, and applies those targeted adjustments while leaving the rest of the soup's character intact.
When you edit a photo to feel more joyful, a single label forces a blunt, one-dimensional change that can flatten the image. This system can raise warmth while keeping the quiet melancholy that gives the scene its depth.
The abstract does not name the specific eight emotion dimensions or explain what they represent. It does not address how the system performs in direct user preference studies or whether it generalizes to images outside its training distribution. The approach depends on the accuracy of the frozen Emotion Distribution Predictor, but the abstract does not discuss what happens when that predictor is wrong.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
AffectDelta: Beyond Emotion Labels for Image Editing — Xingzu Zhan, Lin Gu, Ruogu Fang
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.