Text-to-image models have recently made significant advances in generating realistic and semantically coherent images, driven by advanced diffusion models and large-scale web-crawled datasets. However, these datasets often contain inappropriate or biased content, raising concerns about the generation of harmful outputs when provided with malicious text prompts. We propose Safe Text embedding Guidance (STG), a training-free approach to improve the safety of diffusion models by guiding the text embeddings during sampling. STG adjusts the text embeddings based on a safety function evaluated on the expected final denoised image, allowing the model to generate safer outputs without additional training. Theoretically, we show that STG aligns the underlying model distribution with safety constraints, thereby achieving safer outputs while minimally affecting generation quality. Experiments on various safety scenarios, including nudity, violence, and artist-style removal, show that STG consistently outperforms both training-based and training-free baselines in removing unsafe content while preserving the core semantic intent of input prompts. Our code is available at https://github.com/aailab-kaist/STG.

Text-to-image diffusion models rely on text embeddings from a pre-trained text encoder, but these embeddings remain fixed across all diffusion timesteps, limiting their adaptability to the generative process. We propose Diffusion Adaptive Text Embedding (DATE), which dynamically updates text embeddings at each diffusion timestep based on intermediate perturbed data. We formulate an optimization problem and derive an update rule that refines the text embeddings at each sampling step to improve alignment and preference between the mean predicted image and the text. This allows DATE to dynamically adapts the text conditions to the reverse-diffused images throughout diffusion sampling without requiring additional model training. Through theoretical analysis and empirical results, we show that DATE maintains the generative capability of the model while providing superior text-image alignment over fixed text embeddings across various tasks, including multi-concept generation and text-guided image editing. Our code is available at https://github.com/aailab-kaist/DATE.

The Geostationary Environment Monitoring Spectrometer (GEMS), the first ultraviolet-visible hyperspectral imager in geostationary orbit, has been operating since its launch in February 2020. This paper evaluates and discusses the on-orbit performance of GEMS using 4.5 years of irradiance measurements (300-500 nm), with particular emphasis on the long-term changes in its radiometric accuracy and spectral stability. As part of this assessment, we examine various super-Gaussian forms used for deriving on-orbit slit functions and conclude that applying symmetry constraints to both shape and width parameters enhances fitting stability and accuracy. During the initial phase of operation, the full width at half maximum (FWHM) of the derived slit functions ranged from 0.56 to 0.58 nm across spatial and spectral dimensions. The wavelength shift remains spatially uniform and varies smoothly within ± 0.05 nm, showing a consistent, repeating pattern in 300-400 nm and 400-500 nm. However, over time, instrument degradation has led to pronounced spectral and spatial variations, particularly at shorter wavelengths. Slit widths (FWHM) at the center pixels fluctuate between 0.54 and 0.60 nm, and a “frown-shaped” spatial pattern emerges in the wavelength shift. Radiometric offsets exhibit a steady increase—approximately 5 % per year at shorter wavelengths below 330 nm and 0.2 % above 460 nm. By 2024, irradiance values are up to 35 % lower near 325 nm and about 10 % lower above 460 nm compared to the solar reference. To isolate fine-scale uncertainties, we further evaluate the fitting residuals between the measured irradiance and the radiometrically adjusted solar reference. As the degradation progressed, residuals in 310-330 nm increased sharply from 0.4 % in 2020 to 4 % in 2024, while fitting uncertainties nearly doubled in 330-360 nm. This degradation feature indicates the emergence of instrumental artifacts in spectral structures that are unrelated to physical solar absorption features. Such artifacts are most pronounced at the central pixels and have gradually extended toward the southern pixels and adjacent spectral regions throughout the mission. Beyond 400 nm, radiometric offsets occur on a broadband scale, with no noticeable degradation in the fine spectral structures.

Geostationary Environment Monitoring Spectrometer (GEMS), the first UV-Vis hyperspectral imaging spectrometer onboard a geostationary satellite launched in February 2020, is working with overall performances as well as characteristics aligned with ground-based characterizations. However, there are noticeable issues, especially in the solar irradiances which show a significant discrepancy compared to reference datasets, the focus of current study. The key discrepancy is the variation of measured solar irradiance along the time as well as space of which the root causes are traced back to the angular dependence of the diffuser transmittance and its degradation, both of which critically impact the accuracy of the GEMS Level-2 data products. To mitigate the discrepancy, the current study introduces an empirical correction approach that uses the correlation between the azimuth angle and the measured daily irradiance using 3.5 years of data. With the correction, the spatial and seasonal discrepancies in both irradiance and Earth reflectance disappeared almost completely. Furthermore, the mean bias and root-mean-square deviation (RMSD) against the solar reference spectrum decreased by 12% and 5%, respectively. However, the corrected irradiance values are still lower than those from reference data and other satellites, indicating the potential need for future updates to the radiometric calibration coefficients.

The Geostationary Environment Monitoring Spectrometer (GEMS), an ultraviolet and visible imaging spectrometer, provides air-quality information over a large area of the Asia Pacific region with a high spatiotemporal resolution. To assure the reliability of trace gas retrieval, accurate knowledge of the spectral response function (SRF) is critical for spectral calibration as well as retrieval algorithms. Here, we characterize the GEMS SRF using prelaunch SRFs obtained with the monochromatic laser measurements during the ground test and inflight SRFs retrieved using the solar irradiance measurements after the launch. The prelaunch SRFs are analyzed in terms of shape (skewness and kurtosis), width, and under-sampling and show that the full-width at half-maximum is smaller than 0.6 nm with a maximum of 0.589 nm. The variations along both the spectral and spatial directions are smooth and within 3.65%, indicating a highly homogenous and stable optical system of GEMS. To characterize the prelaunch SRFs and monitor the behavior of inflight SRFs, we applied several analytical functions including asymmetric super Gaussian (ASG) and hybrid Gaussians to the prelaunch SRFs. The spectral fitting of the measured GEMS irradiance with a reference spectrum shows that the ASG to be the best representative of the GEMS SRFs. The inflight SRFs, retrieved with the GEMS irradiances and the ASG, agree well with the prelaunch SRFs, suggesting that the inflight spectral performance and characteristics of GEMS are similar to those investigated from the on-ground characterization.