SearcharxivSearch

arXiv subjects

Seungjoo Baek

Publications and source records attributed to Seungjoo Baek.

4 recordsLinked to original sources

Setting-Matched and Semantics-Scaled Benchmarking of One-Step Generative Models Against Multistep Diffusion and Flow Models

State-of-the-art text-to-image models produce high-quality images, but inference remains expensive as generation requires several sequential ODE or denoising steps. Native one-step models aim to reduce this cost by mapping noise to an image in a single step, yet fair comparisons to multi-step systems are difficult because studies use mismatched sampling steps and different classifier-free guidance (CFG) settings, where CFG can shift FID, Inception Score, and CLIP-based alignment in opposing directions. It is also unclear how well one-step models scale to multi-step inference, and there is limited standardized out-of-distribution evaluation for label-ID-conditioned generators beyond ImageNet. To address this, we benchmark eight models spanning one-step flows (MeanFlow, Improved MeanFlow, SoFlow), multi-step baselines (RAE, Scale-RAE), and established systems (SiT, Stable Diffusion 3.5, FLUX.1) under a class-conditional protocol on ImageNet validation, ImageNetV2, and reLAIONet, our new proofread out-of-distribution dataset aligned to ImageNet label IDs. Using FID, Inception Score, CLIP Score, and Pick Score, we show that FID-focused model development and CFG selection can be misleading in few-step regimes, where guidance changes can improve FID while degrading text-image alignment and human preference signals, worsening visual quality. To make these tradeoffs explicit, we introduce CLIP-scaled and PickScore-scaled variants of FID (csFID, psFID) and Inception Score (csIS, psIS) to serve as a diagnostic for semantically aligned image generation.

cs.CV

Non-hyperbolic 3-manifolds and bulk field theories for supersymmetric/$W_N$ minimal models

Building on the work of Gang, Kang, and Kim arXiv:2405.16377, we propose 3D bulk dual field theories for 2D $\mathcal{N}=1$ supersymmetric minimal models $SM(P, Q)$ and $W_{N}$ algebra minimal models $W_{N}(P, Q)$. We associate to $SM(P, Q)$ a Seifert fibered space $S^2((P,P-R),(Q,S),(3,1))$ with $PS-QR=2$, and for $W_{N}(P, Q)$ a Seifert fibered space $S^2((P,P{-}R),(Q,S),(N{+}1,-2N{-}1))$ with $PS-QR=1$, and realize the bulk theory via the 3D-3D correspondence. For the unitary series, the bulk theory flows in the IR to a gapped phase which, under suitable boundary conditions, supports the unitary chiral minimal model on the boundary. For the non-unitary series, the bulk theory flows to the 3D $\mathcal{N}=4$ superconformal field theory whose topological twist yields a non-unitary topological field theory supporting the non-unitary chiral minimal model on the boundary under appropriate boundary conditions. We also propose UV gauge theory descriptions of the bulk theories obtained by gluing $T[SU(n)]$ building blocks. For $SM(P, Q)$, we provide non-trivial consistency checks -- matching between various bulk partition functions and boundary conformal data -- while for $W_N(P, Q)$, we present preliminary checks and leave further consistency checks for future work.

hep-th

3D bulk field theories for 2D non-unitary N=1 supersymmetric minimal models

We propose bulk 3D N=4 rank-0 superconformal field theories, which are related to 2D N=1 supersymmetric minimal models, SM(2, ...) and SM(3, ...), via recently discovered non-unitary bulk-boundary correspondence. The correspondence relates a 3D N=4 rank-0 superconformal field theory to 2D chiral rational conformal field theories. A topologically twisted theory of the rank-0 SCFT supports the rational chiral algebra at the boundary upon a proper choice of boundary condition. We test the proposal by checking several non-trivial dictionaries of the correspondence.

hep-th

Krylov complexity in inverted harmonic oscillator

Recently, the out-of-time-ordered correlator(OTOC) and Krylov complexity have been studied actively as a measure of operator growth. OTOC is known to exhibit exponential growth in chaotic systems, which was confirmed in many previous works. However, in some non-chaotic systems, it was observed that OTOC shows chaotic behavior and cannot distinguish saddle-dominated scrambling from chaotic systems. For K-complexity, in the universal operator growth hypothesis, it was stated that Lanczos coefficients show linear growth in chaotic systems, which is the fastest. But recently, it appeared that Lanczos coefficients and K-complexity show chaotic behavior in the LMG model and cannot distinguish saddle-dominated scrambling from chaos. In this paper, we compute Lanczos coefficients and K-complexity in an inverted harmonic oscillator. We find that they exhibit chaotic behavior, which agrees with the case of the LMG model. We also analyze bounds on the quantum Lyapunov coefficient and the growth rate of Lanczos coefficients and find that there is a difference with the chaotic system. Microcanonical K-complexity is also analyzed and compared with the OTOC case.

quant-ph