498 research outputs found
DiffExp: Efficient Exploration in Reward Fine-tuning for Text-to-Image Diffusion Models
Fine-tuning text-to-image diffusion models to maximize rewards has proven effective for enhancing model performance. However, reward fine-tuning methods often suffer from slow convergence due to online sample generation. Therefore, obtaining diverse samples with strong reward signals is crucial for improving sample efficiency and overall performance. In this work, we introduce DiffExp, a simple yet effective exploration strategy for reward fine-tuning of text-to-image models. Our approach employs two key strategies: (a) dynamically adjusting the scale of classifier-free guidance to enhance sample diversity, and (b) randomly weighting phrases of the text prompt to exploit high-quality reward signals. We demonstrate that these strategies significantly enhance exploration during online sample generation, improving the sample efficiency of recent reward fine-tuning methods, such as DDPO and AlignProp
Wearable fabric-based hybrid energy harvester from body motion and body heat
We demonstrate a wearable fabric-based energy harvester (Fab-EH) to hybridize a triboelectric generator (TrG) and a thermoelectric generator (ThG). Using a liquid-phase aluminum (Al) coating technique, we fabricate Al-coated fabrics with high electrical and thermal conductivity while maintaining the original texture. By adopting the Al-coated fabric material as both the electrodes of the TrG and the heat-transport layer of the ThG, the fabricated Fab-EH effectively scavenges electrical energy from body motion and body heat. The Fab-EH charges a storage capacitor with capacitance of 3.3 mF to 3 V within 240 sec with the designed transforming system. Moreover, a smartphone can be partially charged using the harvested energy. The Fab-EH thus has the potential to alleviate the recharging issue for portable and wearable devices.
Preparation of cellular samples using graphene cover and air-plasma treatment for time-of-flight secondary ion mass spectrometry imaging
We report on sample preparation methods based on plasma treatment for an improvement of multiple molecular ion images of cellular membranes in the ToF-SIMS method. The air-plasma treatment of fixed cellular samples efficiently removed the organic residues of any solutions used during sample preparation and improved the quality of ToF-SIMS images due to the resulting clean surface. We also studied cell preparation methods that combine single-layer graphene covering with air-plasma treatment to achieve a synergistic effect that eliminates background spectra by organic impurities while minimizing morphological cell deformation in a vacuum environmental analysis. When the cellular sample on the glass substrate is completely covered with the single-layer graphene, the cells trapped between the graphene and the substrate can effectively reduce morphological deformation by slow-dehydration. After slow-dehydration of cells is completed inside the graphene-cover, the covered graphene layer can be peeled off by air-plasma treatment, and the unwanted organic residues on the surface of cells and substrate can also be removed by plasma cleaning, thereby much improving ion imaging of cells with the ToF-SIMS method. It is confirmed that the cell samples in which the graphene-cover was removed by air-plasma treatment maintained their morphology well in comparison with the rapid air-dried cells in atomic force microscopy (AFM) and ToF-SIMS images
Self-Stabilized Soliton Generation in a Microresonator Through Mode- Pulled Brillouin Lasing
Major hurdle for fully chip-compatible soliton microcombs is that complex external systems are required for generating and maintaining the soliton state which stays in thermo-optically unstable region [1]. To overcome this limitation, self-stabilized soliton generation method using Brillouin lasing has been recently suggested [2]. Here we verify its operating mechanism and long-term stability through analyzing RF spectrum
Recommended from our members
Divisions, Areas, and Ensembles - Jazz - Jazz Singers I
Photograph of Stella Lee standing backstage during the Fall 2012 Concert at Voertman Hall. Daewon Lee is visible in the background
Quantum Dot/Siloxane Composite Film Exceptionally Stable against Oxidation under Heat and Moisture
We report on the fabrication of a siloxane-encapsulated quantum dot (QD) film (QD-silox film), which exhibits stable emission intensity for over 1 month even at elevated temperature and humidity. QD-silox films are solidified via free radical addition reaction between oligosiloxane resin and ligand molecules on QDs. We prepare the QD-oligosiloxane resin by sol gel condensation reaction of silane precursors with QDs blended in the precursor solution, forgoing ligand-exchange of QDs. The resulting QD-oligosiloxane resin remains optically clear after 40 days of storage, in contrast to other QD-containing resins which turn turbid and ultimately form sediments. QDs also disperse uniformly in the QD-silox film, whose photoluminescence (PL) quantum yield (QY) remains nearly unaltered under harsh conditions; for example, 85 degrees C/5% relative humidity (RH), 85 degrees C/85% RH, strongly acidic, and strongly basic environments for 40 days. The QD-silox film appears to remain equally emissive even after being immersed into boiling water (100 degrees C). Interestingly, the PL QY of the QD-silox film noticeably increases when the film is exposed to a moist environment, which opens a new, facile avenue to curing dimmed QD-containing films. Given its excellent stability, we envision that the QD-silox film is best suited in display applications, particularly as a PL-type down-conversion layer.
Recommended from our members
Divisions, Areas, and Ensembles - Jazz - Jazz Singers I
Photograph of (l-r) Daewon Lee, Claire Hebert, Tyler Thomas, and Stella Lee performing with the University of North Texas Jazz Singers during the Fall 2012 Concert at Voertman Hall
- …
