图像生成

The Frechet Inception Distance (FID) is the de facto arbiter of image generation, yet most papers report just a single number from a single trained model using a single sampling seed. How reproducible is that number if we retrain the model, or merely resample from it? In this paper, we treat FID as a random variable on a two-axis panel of training and generation seeds, and measure its variance directly on several hundred SiT networks trained on class-conditional ImageNet 256x256. We report surprising findings: (a) Retraining the model using the same recipe with a different seed moves FID 3.2x more (in Inception feature space) than redrawing samples from a fixed network. (b) That gap is driven by three factors: random initialisation, data ordering, and the per-step Gaussian noise of the flow-matching loss. (c) Increasing compute or model size barely tightens the spread, holding the FID coefficient of variation (CoV) inside a 1-2% band. (d) Per-cell classifier-free-guidance tuning halves the spread but reshuffles which seeds work best, and a lucky training seed reaches the same FID with up to 2x less compute than an unlucky one. Based on these findings, we recommend a new FID evaluation protocol: evaluate under per-cell optimal guidance, treat any FID gap below the empirically measured ~1.3% CoV as inconclusive, and report an error bar over several training seeds rather than a single FID number.

Training-free detectors of AI-generated images promise generator-agnostic deployment without classifier training, yet their reported numbers are rarely compared under a single controlled protocol. We audit two representative training-free scores -- an autoencoder-reconstruction score (AEROBLADE-style) and a noise-perturbation feature-similarity score (RIGID-style) -- plus a naive feature-kNN control, on a common 1,500-image GenImage-derived benchmark spanning seven generators and JPEG compression at quality 70 and 50. The audit yields three cautionary findings. (i) Implementation details masquerade as method differences: replacing the LPIPS backbone (AlexNet -> VGG-16) changes overall AUROC by +0.085, and switching between resize-to-512 and native-resolution preprocessing flips per-generator conclusions by up to 0.38 AUROC. (ii) Score direction is not a property of the method but of its hyperparameters: the RIGID-style score is inverted (AUROC < 0.5) on SD1.5 and Wukong at noise level sigma=0.05, recovers to >0.5 for every generator at sigma=0.01, and collapses to 0.15 at sigma=0.3. (iii) Dataset format bias inflates robustness claims: without unified re-encoding, AUROC under JPEG-50 exceeds the clean condition for the AlexNet-backbone reconstruction score; after bias correction the residual anomaly localizes to a single generator (BigGAN). The audited scores have complementary per-generator failure sets, but naive z-score fusion does not beat the best single score, indicating that exploiting complementarity requires direction-aware combination.

Creating 3D visual illusions, a single 3D mesh that reveals entirely different semantics from various viewing angles, is a fascinating but tough challenge. Existing optimization-based methods are slow and can produce oversaturated colors. In contrast, naive stitching approaches fail to produce geometrically coherent objects. This results in visible unnatural seams and semantic leaks. In this paper, we present a fast and training-free framework for generating text-driven 3D visual illusions. Our approach decouples the generation into two stages. First, we propose a cross-space dual-branch denoising process. This process dynamically decodes 3D latents into voxel space for CLIP-guided orientation alignment and Signed Distance Field (SDF) blending, which ensures seamless geometric fusion. Second, we introduce a view-conditioned texture synthesis module that projects and aggregates view-specific 2D diffusion priors onto the fused geometry. Extensive experiments demonstrate that our method generates highly realistic, dual-semantic 3D illusions in just 3-5 minutes. It significantly outperforms existing methods in geometric integrity, semantic recognizability, and efficiency. Project page: https://siang1105.github.io/JanusMesh.github.io/

Accent text-to-speech (TTS) aims to synthesize speech with target accents. Existing accent TTS systems typically rely on a two-stage pipeline that first converts standard phone sequences into accented phone sequences and then synthesizes accented speech. However, such approaches suffer from error accumulation and require paired standard-accented phone sequence data, which is often limited in practice. Moreover, text-based accented phone representations are insufficient to model acoustic accent characteristics such as prosody and rhythm. In this work, we propose Joycent, a diffusion-based accent TTS model that synthesizes accented speech directly from standard phone sequences and speech references without accented phone prediction. Joycent integrates accent and speaker representations through conditional layer normalization (CLN) in the text encoder. We introduce WhisAID, a Mandarin accent identification model trained on accented Mandarin speech to extract accent representations. Experimental results show that Joycent improves accentedness while preserving speaker identity compared with baseline systems. We release our code and demos at: https://github.com/oshindow/Joycent-code.

Machine learning is often touted to improve the efficiency of ICT, but that small gain is overwhelmed by the enormous carbon, water, and land footprints of data centers and ML-ready devices. We survey the electricity consumption of ML applications in training and inference, focusing on electricity-intensive image generation. Our team of a computer engineer, a media scholar, and an artist explore solutions including inexact computing; tiny language models; low-precision hardware architectures; hardware with limited capacity; and anticipating and mitigating energy demands at the design phase. We will sketch our work in progress of an ethical and aesthetically sophisticated tiny image generator using non-scraped data. Looking to the economic context, we will propose a true-cost accounting for the environmental impact of machine learning and suggest that the criterion of efficiency is driven by the shareholder-capitalist framing of ICT.

Measurements of cosmic-ray proton and helium spectra below the knee show deviations from simple power laws, including multi-TeV structures. A possible explanation is that one or a few nearby sources contribute an additional component to the local spectrum. However, previous study shows that a dominant local contribution is statistically unlikely under a homogeneous diffusion model. In this work, we investigate how this probability changes if cosmic rays experience inefficient transport near their sources, motivated by observations of extended gamma-ray emission around Galactic accelerators. We derive a series Green's function that enables fast calculation of the particle distribution in this scenario, making Monte Carlo calculations for Galactic source populations feasible. The inner slow-diffusion region delays escape and redistributes the arriving particles in time and energy. In Monte Carlo realizations, the probability that the strongest local source becomes comparable to the background at $10\,\rm{TeV}$ increases from about $0.4\%$ in homogeneous diffusion to $1.7$--$2.2\%$ in the two-zone models. Thus inhibited near-source transport weakens, but does not remove, the statistical difficulty. We then examine cataloged nearby candidate supernova remnants and show that a $10\,\rm{TeV}$ feature can be reproduced only with additional assumptions, especially a harder local injection spectrum and a favorable diffusion coefficient. The predicted contribution of a given source changes strongly among different particle transport model. Therefore, the local source interpretations are plausible but highly model dependent, and require independent constraints on source injection history, particle transport mechanisms, and local interstellar turbulence.