TherA: Thermal-Aware Visual-Language Prompting for
Controllable RGB-to-Thermal Infrared Translation

Dong-Guw Lee^1*  Tai Hyoung Rhee^1*  Hyunsoo Jang¹
Young-Sik Shin²  Ukcheol Shin³  Ayoung Kim^1†

¹ Seoul National University   ² Kyungpook National University   ³ KENTECH
^* Equal Contribution   ^† Corresponding Author

CVPR 2026

News

• ⚡(2026-04-03): TherA & R2T2 Dataset repo opening

---

🔥 TherA: Thermal-Aware VLM-based Controllable RGB→TIR Translation

• 🧠 Thermal-Aware VLM Conditioning
• 🎯 Dual-Level Controllability
• 🌡 Physically Plausible Synthesis
• 🏆 SoTA Performance & Strong Zero-Shot Generalization

📦 R2T2: 100k+ Aligned RGB–TIR–Text Dataset

• 🖼 112,970 Aligned Triplets: RGB image + TIR image + Canonical thermal schema
• 🏙 Scene Diversity: Driving, CCTV, Aerial, Ego-view
• 🌗 Temporal Diversity: Day/Night, Diurnal transitions
• 🌦 Environmental Diversity: Weather, season, illumination variability
• 🧱 Material & Object-Level Annotation with Structured Canonicalization
• 📚 Compiled from 9 aligned RGB-TIR datasets with additional pseudo-aligned pairs

---

TherA Overview

Dataset Overview Table

Click to Expand

Category	Dataset	Train	Test	Extra	Scene	Weather	Season	Location	Avg. Resolution
Outdoor	MS²	93,746	18,896	84,358	Campus, Urban, Residential	Clear, Cloudy, Rainy	Summer	Korea	544 × 191
	STHeReO	46,437	9,745		Campus, Suburban	Clear	Summer	Korea	601 × 245
	ViViD	35,796	14,597		Campus	Clear, Cloudy	Spring	Korea	629 × 497
	NSAVP	65,333	78,823		Urban, Suburban	Clear, Cloudy	Summer	Korea	640 × 512
	CAMEL (Outdoor)	8,581	4,482		Campus, Road, Urban	Clear, Cloudy, Snow	Spring, Fall, Winter	USA	404 × 230
	TRI2I	19,768	11,913		Campus, Road	Clear, Cloudy	Spring, Summer	USA	229 × 228
	METU-VisTIR	33	1,052		Campus	Clear		Turkey	632 × 497
	MIRAGE Outdoor	269,694	139,508	84,358
Indoor	Trimodal	4,550	2,653		Room			Austria	640 × 480
	MultiSpectralMotion	11,575	5,777	3,647	Room			China	640 × 480
	OdomBeyondVision	20,904	4,372		Room			UK	328 × 249
	CAMEL (Indoor)	221	117		Hall			USA	404 × 230
	MIRAGE Indoor	37,250	12,919	88,005
Total	MIRAGE	306,944	152,427	88,005
	MIRAGE Raw	278,341	130,491	88,005

• MIRAGE Raw represents the data pairs providing both 8-bit and raw 14-bit TIR

Example Visualizations of each Dataset Sequence in MIRAGE (Left: RGB / Right: TIR)

Comparison with Other Datasets

---

Download

Refer to the link below for dataset download.

Google Drive

---

Dataset Structure

  R2T2
  ├── {$DATASET_NAME}
  |   └── {$SEQUENCE_NAME}
  |       ├── RGB
  |       |   ├── 1.jpg
  |       |   └── ...
  |       └── TIR
  |           ├── 1.jpg
  |           └── ...
  ├── ...
  ├── ViVID
  |   ├── img_campus_day1
  |   |   ├── RGB
  |   |   |   ├── 000001.png
  |   |   |   └── ...
  |   |   └── TIR
  |   |       ├── 000001.png
  |   |       └── ...
  |   ├── ...
  ├── ...

---

Citation

If you found our work useful, please cite

@inproceedings{lee2025thera,
title={TherA: Thermal-Aware Visual-Language Prompting for<br>Controllable RGB-to-Thermal Infrared Translation},
  author={Lee, Dong-Guw and Rhee, TaiHyoung and Jang, Hyunsoo and Young-Sik Shin and Shin, UkCheol and Kim, Ayoung},
  booktitle={CVPR},
  year={2026}
}