AtmoSteer: Personalized Scene Control with Reusable Atmospheric Recipes

Generating scenes that convey atmospheres such as “cozy” or “contemplative” is difficult because people interpret them subjectively and often struggle to express them concretely. We present AtmoSteer, a human-in-the-loop approach that captures a user’s personal interpretation of an atmosphere as an atmospheric recipe: a weighted set of descriptive tags that can be reused across scenes for tasks such as image generation and image search. AtmoSteer supports users in discovering their own atmospheric recipes by exploring diverse interpretations with images and constructing a personalized vocabulary of visually grounded tags that describe style, layout, objects, and lighting. The system then supports preference refinement through two modes: slider-based weight adjustment of tags and visual example-based elicitation. In a user study on image generation (N=20), images generated from recipes produced by AtmoSteer were significantly more preferred and better aligned with participants’ subjective interpretations than those generated using generic descriptors or style transfer techniques.

@inproceedings {Sun26Atmosteer, 
 author = {Sun, Ruonan Nancy and Cho, Hyunsung and Lindlbauer, David}, 
 title = {AtmoSteer: Personalized Scene Control with Reusable Atmospheric Recipes}, 
 year = {2026}, 
 publisher = {Association for Computing Machinery}, 
 address = {New York, NY, USA}, 
 keywords = {Text-to-image generation, human-in-the-loop preference elicitation}, 
 location = {Detroit, MI, USA}, 
 doi = {10.1145/3830398.3830683}, 
 series = {UIST '26} 
 }