Search

Diffusion models have shown excellent performance on various image generation tasks, but the substantial computational costs and huge memory footprint hinder their low-latency applications in real-world scenarios. Quantization is a promising way to compress and accelerate models. Nevertheless, due to the wide range and time-varying activations in diffusion models, existing methods cannot maintain both accuracy and efficiency simultaneously for low-bit quantization. To tackle this issue, we propose DilateQuant, a novel quantization framework for diffusion models that offers comparable accuracy and high efficiency. Specifically, we keenly aware of numerous unsaturated in-channel weights, which can be cleverly exploited to reduce the range of activations without additional computation cost. Based on this insight, we propose Weight Dilation (WD) that maximally dilates the unsaturated in-channel weights to a constrained range through a mathematically equivalent scaling. WD costlessly absorbs the activation quantization errors into weight quantization. The range of activations decreases, which makes activations quantization easy. The range of weights remains constant, which makes model easy to converge in training stage. Considering the temporal network leads to time-varying activations, we design a Temporal Parallel Quantizer (TPQ), which sets time-step quantization parameters and supports parallel quantization for different time steps, significantly improving the performance and reducing time cost. To further enhance performance while preserving efficiency, we introduce a Block-wise Knowledge Distillation (BKD) to align the quantized models with the full-precision models at a block level. The simultaneous training of time-step quantization parameters and weights minimizes the time required, and the shorter backpropagation paths decreases the memory footprint of the quantization process., Comment: Code: http://github.com/BienLuky/DilateQuant

The rapid advancement of visual generative models necessitates efficient and reliable evaluation methods. Arena platform, which gathers user votes on model comparisons, can rank models with human preferences. However, traditional Arena methods, while established, require an excessive number of comparisons for ranking to converge and are vulnerable to preference noise in voting, suggesting the need for better approaches tailored to contemporary evaluation challenges. In this paper, we introduce K-Sort Arena, an efficient and reliable platform based on a key insight: images and videos possess higher perceptual intuitiveness than texts, enabling rapid evaluation of multiple samples simultaneously. Consequently, K-Sort Arena employs K-wise comparisons, allowing K models to engage in free-for-all competitions, which yield much richer information than pairwise comparisons. To enhance the robustness of the system, we leverage probabilistic modeling and Bayesian updating techniques. We propose an exploration-exploitation-based matchmaking strategy to facilitate more informative comparisons. In our experiments, K-Sort Arena exhibits 16.3x faster convergence compared to the widely used ELO algorithm. To further validate the superiority and obtain a comprehensive leaderboard, we collect human feedback via crowdsourced evaluations of numerous cutting-edge text-to-image and text-to-video models. Thanks to its high efficiency, K-Sort Arena can continuously incorporate emerging models and update the leaderboard with minimal votes. Our project has undergone several months of internal testing and is now available at https://huggingface.co/spaces/ksort/K-Sort-Arena, Comment: Project page: https://huggingface.co/spaces/ksort/K-Sort-Arena

Large transformer models have demonstrated remarkable success. Post-training quantization (PTQ), which requires only a small dataset for calibration and avoids end-to-end retraining, is a promising solution for compressing these large models. Regrettably, existing PTQ methods typically exhibit non-trivial performance loss. We find that the performance bottleneck stems from over-consideration of hardware compatibility in the quantization process, compelling them to reluctantly employ simple quantizers, albeit at the expense of accuracy. With the above insights, we propose RepQuant, a novel PTQ framework with quantization-inference decoupling paradigm to address the above issues. RepQuant employs complex quantizers in the quantization process and simplified quantizers in the inference process, and performs mathematically equivalent transformations between the two through quantization scale reparameterization, thus ensuring both accurate quantization and efficient inference. More specifically, we focus on two components with extreme distributions: LayerNorm activations and Softmax activations. Initially, we apply channel-wise quantization and log$\sqrt{2}$ quantization, respectively, which are tailored to their distributions. In particular, for the former, we introduce a learnable per-channel dual clipping scheme, which is designed to efficiently identify outliers in the unbalanced activations with fine granularity. Then, we reparameterize the scales to hardware-friendly layer-wise quantization and log2 quantization for inference. Moreover, quantized weight reconstruction is seamlessly integrated into the above procedure to further push the performance limits. Extensive experiments are performed on different large-scale transformer variants on multiple tasks, including vision, language, and multi-modal transformers, and RepQuant encouragingly demonstrates significant performance advantages.

Diffusion models have achieved great success in image generation tasks through iterative noise estimation. However, the heavy denoising process and complex neural networks hinder their low-latency applications in real-world scenarios. Quantization can effectively reduce model complexity, and post-training quantization (PTQ), which does not require fine-tuning, is highly promising for compressing and accelerating diffusion models. Unfortunately, we find that due to the highly dynamic distribution of activations in different denoising steps, existing PTQ methods for diffusion models suffer from distribution mismatch issues at both calibration sample level and reconstruction output level, which makes the performance far from satisfactory, especially in low-bit cases. In this paper, we propose Enhanced Distribution Alignment for Post-Training Quantization of Diffusion Models (EDA-DM) to address the above issues. Specifically, at the calibration sample level, we select calibration samples based on the density and variety in the latent space, thus facilitating the alignment of their distribution with the overall samples; and at the reconstruction output level, we modify the loss of block reconstruction with the losses of layers, aligning the outputs of quantized model and full-precision model at different network granularity. Extensive experiments demonstrate that EDA-DM significantly outperforms the existing PTQ methods across various models (DDIM, LDM-4, LDM-8, Stable-Diffusion) and different datasets (CIFAR-10, LSUN-Bedroom, LSUN-Church, ImageNet, MS-COCO)., Comment: Code: http://github.com/BienLuky/EDA-DM

We investigate the temperature-dependent production of feebly interacting massive dark matter particle (FIMP DM) within a $Z_2$ model, incorporating two $Z_2$-odd scalar fields. In specific parameter regions, three distinct mechanisms individually dominate the production of the FIMP DM. These mechanisms include the semi-production process, commonly known as the ``exponential growth'' process, the production from pair annihilations of the bath particles, and the three-body decay process. It is crucial to consider the thermal history during the evolution of FIMPs, as it involves multiple phase transitions occurring prior to the freeze-in of dark matter. Consequently, the scalar masses experience thermal variations, leading to a distinctive evolutionary trajectory for FIMPs when compared to scenarios without accounting for the thermal effects. Notably, the unique pattern of FIMP evolution is accompanied by the production of gravitational waves, presenting promising opportunities for detection using forthcoming interferometers., Comment: Published in EPJC

We investigate the inelastic signatures of dark matter-nucleus interactions, explicitly focusing on the ramifications of polarization, dark matter splitting, and the Migdal effect. Direct detection experiments, crucial for testing the existence of dark matter, encounter formidable obstacles such as indomitable neutrino backgrounds and the elusive determination of dark matter spin. To overcome these challenges, we explore the potential of polarized-target dark matter scattering, examining the impact of nonvanishing mass splitting and the role of the Migdal effect in detecting light dark matter. Our analysis demonstrates the valuable utility of the polarized triple-differential event rate as an effective tool for studying inelastic dark matter. It enables us to investigate both angular and energy dependencies, providing valuable insights into the scattering process., Comment: 12 pages, 7 figures, reference updated

Several Pulsar Timing Array (PTA) collaborations have recently reported the evidence for a stochastic gravitational-wave background (SGWB), which can unveil the formation of primordial seeds of inhomogeneities in the early universe. With the SGWB parameters inferred from PTAs data, we can make a prediction of the seeds for early galaxy formation from the domain walls in the axion-like particles (ALPs) field distribution. This also naturally provides a solution to the observation of high redshifts by the James Webb Space Telescope. The predicted photon coupling of the ALP is within the reach of future experimental searches., Comment: 9 pages, 4 figures. version published in Sci. China-Phys. Mech. Astron. November 2024 Vol.67 No.11: 111011

The early kinetic decoupling (eKD) effect is an inevitable ingredient in calculating the relic density of dark matter (DM) for various well-motivated scenarios. It appears naturally in forbidden dark matter annihilation, the main focus of this work, which contains fermionic DM and a light singlet scalar that connects the DM and standard model (SM) leptons. The strong suppression of the scattering between DM and SM particles happens quite early in the DM depletion history, where the DM temperature drops away from the thermal equilibrium, $T_\chi < T_{\rm SM}$, leading to the decreased kinetic energy of DM. The forbidden annihilation thus becomes inefficient since small kinetic energy cannot help exceed the annihilation threshold, naturally leading to a larger abundance. To show the eKD discrepancy, we numerically solve the coupled Boltzmann equations that govern the evolution of DM number density and temperature. It is found that eKD significantly affects the DM abundance, resulting in almost an order of magnitude higher than that by the traditional calculation. We also discuss the constraints from experimental searches on the model parameters, where the viable parameter space shrinks when considering the eKD effect., Comment: 10 pages, matches published version

The $\rm{keV}$ scale sterile neutrino was a qualified candidate for dark matter particles in the Dodelson-Widrow mechanism. But the mixing angle, needed to provide enough amount of dark matter, is in contradiction with the astrophysical observations. To alleviate such tension, we introduce an effective interaction, i.e. $g_a (\phi/\Lambda)\partial_{\mu}a \overline{\nu_\alpha}\gamma^{\mu} \gamma_5 \nu_\alpha$, among Standard Model neutrino $\nu_\alpha$, axion $a$, and singlet $\phi$. The axial-vector interaction form is determined by the axion shift symmetry, and the singlet $\phi$ with dynamically varied vacuum expectation value is introduced to reinforce the axial-vector coupling strength and evade the stringent neutrino oscillation constraints. The effective potential generated by the new interaction {could cancel} the SM counterpart, resulting in an {enhanced converting} probability between SM neutrino and sterile neutrino. Hence, the production rate of sterile neutrinos can be substantially enlarged with smaller mixing compared to the DW mechanism., Comment: 6 pages, 2 figures; publish version for EPJC

The supersymmetry (SUSY) breaking mechanism generally predicts the existence of the sgoldstinos, which can play the role of wave-like dark matter. Due to the ubiquitous coupling to the electromagnetic fields, the light scalar sgoldstino dark matter can lead to the variance of the fine-structure constant. With the precise atomic clock data, we find the SUSY breaking scale $\sqrt{F}$ can be probed up to the GUT scale in the sgoldstino mass range of $10^{-22}$ eV $

We study the cosmological inflation and dark matter (DM) in a unified way within a $Z_3$ complex scalar model. The real and imaginary parts of the complex scalar act as the inflaton and DM respectively. The slow-rolling inflation with non-minimal coupling in both the metric and Palatini formalisms can be realized. We examine the whole parameters space by fully considering the theoretical and experimental constraints. We find that in the low-energy scale, the DM relic density and the DM-nucleon direct scattering experiments favor the mixing angle $|\theta| \lesssim 0.25$, the DM mass $m_\chi \gtrsim 80\rm{GeV}$, and the mass of Higgs-like scalar $m_{h_2} \gtrsim 300\rm{GeV}$. In the high-energy scale, after further considering the cosmological constraints of the scalar spectral index and the tensor-to-scalar ratio for the two forms of inflation, the scalar spectral indices are both $\sim 0.965$, the non-minimum coupling coefficients are $\sim 10^4$ and $\sim 10^9$, and the tensor-to-scalar ratios are $\sim 10^{-3}$ and $\lesssim 10^{-11}$ respectively, which suggests that the inflation under the two formalisms can be distinguished by measuring the tensor-to-scalar ratio with higher precision., Comment: 16 pages, 5 figures

By assuming the photosphere located at the outmost edge of the ejecta, Arnett et al. (1980, 1982, 1989) presented the light curves of homologous explosions in supernovae analytically and numerically to include recombination effects. Actually as homologous expansion proceeds, the photosphere recedes deeper into the ejecta. In this situation, the photosphere radius increases at early times and decreases later on which can be described by a simple method proposed by Liu et al. (2018). To study how the photosphere recession effect the luminosity evolution, we impose a boundary condition on the photosphere to determine the spatial and time distribution of the temperature of the ejecta which is clarified to be reasonable. We find that the photosphere recession reduce the luminosity compared with the previous result without the recession, which can be tested with observations of Type-IIP supernovae., Comment: 9 pages, 3 figures

We study a minimal extension of the Standard Model by introducing three right-handed neutrinos and a new scotogenic scalar doublet, in which the mass splittings between neutral and charged components are responsible for the $W$-boson mass newly measured by the CDF collaboration. This model can not only generate non-vanishing Majorana neutrino masses via the interaction of right-handed neutrinos and scotogenic scalars, but also explain the Universe's missing matter in the form of FIMP dark matter. We also study the influence of the mass splitting on the first order electroweak phase transition, and find that it can further enhance the transition strength and thus induce gravitational waves during the phase transition, which may be detected in the forthcoming detectors such as U-DECIGO., Comment: References updated, accepted for publication in Science Bulletin