cs.HCApr 2, 2026
SaveAromaGen: Interactive Generation of Rich Olfactory Experiences with Multimodal Language Models
Organizations: MIT Media Lab, Cambridge, MA, USA · University of Washington, Seattle, WA, USA · Harvard Graduate School of Design, Cambridge, MA, USA · Harvard GSD/SEAS, Cambridge, MA, USA ·
Abstract
Smell's connections with food, memory, and social experience have motivated researchers to bring olfaction into interactive systems. Multimodal AI opens new possibilities for generating smells from natural language, yet it remains unclear whether pretrained models encode sufficient olfactory knowledge to translate language into perceptible compositions. We present AromaGen, an AI-powered 12-channel wearable olfactory system that maps free-form natural-language descriptions to 12 base odorants selected to cover a semantically derived olfactory space, while supporting iterative refinement through natural-language feedback. In a between-subjects study (N=60) across 50 real-world benchmark smells, participants distinguished AromaGen-generated target smells above chance in a three-alternative forced-choice task, with retrieval-augmented AI composition performing comparably to human-expert composition and zero-shot AI composition. Natural-language feedback significantly improved the perceived similarity of both AI- and human-expert compositions. Our findings demonstrate the feasibility of using pretrained multimodal AI, with human aroma composition data, to generate perceptible olfactory experiences from natural language.
Figures & tables
Figure 1. AromaGen maps a free-form natural-language description of a desired olfactory experience to a 12-dimensional odorant ratio vector over 12 base odorants, physically rendered through a neck-worn olfactory display, and supports iterative refinement through free-form natural-language feedback.
Figure 2. Two-dimensional describable olfactory space of English smell descriptors, grouped into 12 perceptually informed clusters (same color indicates the same cluster). X-axis represents the pleasantness / unpleasantness axis, Y-axis represents the concrete odor source / abstract evaluative axis.
| # | Odorant | Cluster | Notes |
|---|---|---|---|
| 1 | Benz Sal | Perfumed / Clean | Soft floral, powdery, balsamic |
| 2 | Sandalwood | Woody / Resinous | Creamy, warm wood |
| 3 | Clove Bud (50%) + Cumin (50%) | Spice | Warm pungent clove, earthy cumin |
| 4 | Lavender (50%) + Rose (50%) | Floral | Herbaceous lavender, dewy rose |
| 5 | Orange (50%) + Lemon (50%) | Citrus | Sweet-tart citrus |
| 6 | Vanilla Sugar (50%) + Almond Extract (50%) | Sweet / Gourmand | Creamy vanilla, marzipan almond |
Table 1. The 12 base odorants in the live AromaGen palette. The palette was selected to span diverse regions of the describable olfactory space under the 12-channel hardware constraint. Composite odorants were prepared by combining their constituent materials in the proportions shown. A 12-row palette table. Each row has a channel number, the odorant or composite blend, the matching olfactory cluster, and a short sensory note. Six rows are single materials and six are fixed blends of two or three oils.
Figure 3. The AromaGen composition pipeline. Users describe a desired olfactory experience through text or speech. The model combines pretrained knowledge with retrieved human aroma composition data to produce a 12-dimensional odorant-ratio vector over the 12 base odorants. The wearable renders the mixture, after which users can iteratively refine it through natural-language feedback.
| Olfactory Cluster | Real-World Smell Targets |
|---|---|
| Floral | Lavender; rose; jasmine tea; peony and rose oil shampoo |
| Citrus | Orange; mango; lemonade; lime soda (Sprite) |
| Woody & Resinous | Birch; patchouli; whiskey and oak candle; incense |
| Herbal & Cooling | Basil; cucumber; peppermint tea; mint chewing gum |
| Spice | Ginger; black pepper; chai tea; cinnamon roll |
| Sweet & Gourmand | Coke; dark chocolate; apple pie; sweet popcorn; chocolate and marshmallow Pop-Tarts |
Table 2. The 50 real-world smell targets used in the comparative user study, organized by AromaGen ’s 12 olfactory clusters.
Figure 4. An example of AromaGen ’s iterative refinement: the model decomposes a natural-language request, maps it onto perceptual qualities, and allocates a 12-dimensional odorant-ratio vector over the 12 base odorants. Spoken feedback (e.g., “less sour”) is applied as an in-context revision of the current mixture.
Figure 5. Experimenter-facing web interface for AromaGen : natural-language input (text or speech), a visualization of the 12-dimensional odorant-ratio vector, playback on the wearable, and natural-language feedback. Image input is supported by the architecture but was not evaluated.
Figure 6. AromaGen hardware, with 12 cartridges containing 12 base odorants
Figure 7. The 50 physical odor sources used as references for our real-world benchmark smells across 12 olfactory clusters, presented in odorless paper portion cups. Each row corresponds to one olfactory cluster, with odor sources ordered from left to right.
| Expert | Background | Overall | Highest-rated clusters | Lowest-rated cluster |
|---|---|---|---|---|
| 1 | Barista and taste researcher | 4.22 (1.36) | Chemical & Solvent 5.75; Citrus 5.50 | Fermented & Sour 2.50 |
| 2 | Pastry chef | 5.40 (0.90) | Chemical & Solvent 6.50; Citrus 5.75; Woody & Resinous 5.75 | Roasted & Smoky 4.50 |
| 3 | Perfumer | 3.86 (0.93) | Citrus 4.50; Woody & Resinous 4.50 | Putrid & Decay 2.75 |
Table 3. Mean 7-point similarity of each expert’s final compositions. Overall is the mean (SD) across all 50 targets. Cluster values are that expert’s mean rating for targets in the named cluster. A three-row table of expert elicitation ratings. Expert 1, a barista and taste researcher, averaged 4.22. Expert 2, a pastry chef, averaged 5.40. Expert 3, a perfumer, averaged 3.86. The highest-rated clusters were Chemical and Solvent, Citrus, and Woody and Resinous. The lowest-rated clusters were Fermented and Sour, Roasted and Smoky, and Putrid and Decay.
Figure 8. Overview of the comparative user study procedure across AromaGen composition, human-expert composition, and AromaGen zero-shot composition.
Figure 9. Comparative user study procedure. 1. Study environment. 2. Physical odor presentation. 3. Logging similarity ratings in AromaGen interface. 4. Capturing natural language based feedback in AromaGen interface.
Figure 10. 3-AFC test accuracy by composition condition ( ; 240 trials per condition). Accuracy was above chance across all three conditions ( AromaGen : 70.8%; Expert: 69.6%; AromaGen zero-shot: 63.7%), with no statistically significant differences between conditions.
Figure 11. Phrase cloud of participants’ natural-language feedback. Feedback commonly expressed directional changes to perceptual qualities ("more", "less" etc), relational comparisons to desired smells ("similar", "more like" etc), and uncertainty or approximation ("don’t know", "maybe" etc), illustrating how participants could steer generated smells without explicitly specifying odorants or compositional changes.
Figure 12. Three-alternative forced-choice identification accuracy by olfactory cluster, pooled across conditions (N = 60, 720 trials). (Left) Identification accuracy by smell cluster; all 12 clusters exceed chance (33%), ranging from 83% (Chemical & Solvent) to 45% (Sweet & Gourmand). (Right) Similarity ratings before and after spoken feedback for AromaGen and Human Expert; both conditions show a significant increase in perceived similarity following feedback ( AromaGen : 4.09→4.65; Human Expert: 3.79→4.05).
Figure 13. Percentage of compositions containing each of the 12 base odorants, by generation condition ( AromaGen , Human Expert Elicitation, AromaGen Zero-shot). Usage patterns differ notably by condition; Human-expert compositions rely most heavily on a small number of signature odorants (e.g., Orange + Lemon, 53%), while AromaGen Zero-shot spreads usage most evenly across the palette
Figure 14. Word cloud of participants’ freely generated target aroma, spanning nature ("forest", "mountain" etc), edibles ("kimchi", "sauce" etc), chemicals ("laundry", "detergent" etc.), places ("Colombia", "Seoul" etc), people ("grandma", "sister" etc) and more.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| AI+RAG | Expert | Zero-shot AI | |
| Demographic forms ( ) | 20 | 20 | 19 |
| Women | 17 | 9 | 12 |
| Men | 3 | 11 | 7 |
| Age 25–34 | 10 | 9 | 12 |
| English not primary | 5 | 10 | 7 |
| Ethnicity | |||
Table 4. Participant demographics across conditions in the comparative study. Demographic data were available for 20 participants in the AI+RAG and Expert conditions and 19 participants in the zero-shot AI condition.
| Odorant | Type | Brand | Ingredients / Formula |
|---|---|---|---|
| Cumin | Essential Oil | Silky Scents | 100% pure Cumin seed oil ( Cuminum cyminum ) |
| Eucalyptus | Essential Oil | Creative Flavours & Fragrances | Eucalyptus essential oil ( Eucalyptus spp.) |
| Onion | Powder | Badia | Onion powder |
| Ylang Ylang | Essential Oil | White Naturals | Pure Ylang Ylang essential oil; 1 fl oz |
| Sichuan Oil (Linalool) | Essential Oil | Creative Flavours & Fragrances | High concentration of Linalool derived from Sichuan peppercorns, providing a woody, floral, and citrus-spicy top note. |
| Red Clover | Tincture | Cambridge Naturals | Organic Red Clover flowering tops ( Trifolium pratense ); grain alcohol (45–55% by volume); deionized water; herb strength ratio 1:3 |
Table 5. Sources and compositions of the 12 base odorants used in AromaGen ’s palette.
Figure 15. Hardware Specification Sheet
More in this topic
Multimodal Generation
Multimodal generation creates coordinated text, images, audio, video, 3D, and other outputs from mixed inputs while preserving meaning and consistency.
See all 238 papers