conditioning mechanisms

Current text-to-image models suffer from significant limitations in content controllability, safety, and scalability—particularly in robustly suppressing unsafe concepts (e.g., nudity) or performing zero-shot style injection. To address this, we propose the first training-free, interpretable concept modulation framework. It leverages k-sparse autoencoders (k-SAEs) to disentangle and localize unambiguous concepts within the text embedding latent space, enabling bidirectional steering—i.e., suppression or excitation—via Concept Activation Intervention (CAI). Our method requires no fine-tuning, LoRA adaptation, or architectural modification, supporting zero-shot style transfer and adversarially robust intervention. Experiments demonstrate a 20.01% improvement in unsafe content removal rate, preservation of generation quality, inference speed five times faster than state-of-the-art alternatives, and exceptional stability across multi-style transfer and adversarial prompting scenarios.

To address the lack of universality and interpretability in latent variable manipulation for generative models, this paper proposes Linear Latent Composition (LOL). LOL establishes the first modality-agnostic and architecture-agnostic framework enabling arbitrary linear operations—including interpolation, subspace construction, and low-dimensional representation extraction—without additional training or fine-tuning. Its core innovation lies in applying geometrically consistent linear transformations within the latent spaces of mainstream generative paradigms, including diffusion models, flow matching, and continuous normalizing flows. Unlike existing approaches constrained to specific architectures or data modalities, LOL significantly enhances the flexibility and reproducibility of controllable generation. Empirical evaluations demonstrate its strong generalization and practical utility across diverse applications: synthetic data generation, data augmentation, and multimodal experimental design.

Existing generative data augmentation methods struggle to simultaneously preserve image fidelity and diversity under joint vision-language prompting, primarily due to entanglement between image representations and non-essential attributes (e.g., background), causing conflicts with text prompts. To address this, we propose a saliency-aware image generation framework that explicitly disentangles and suppresses interfering visual attributes via a novel embedding model. Our method integrates vision-language conditional generation with a personalized disentanglement architecture to enable fine-grained, semantically consistent, and controllable image synthesis. Evaluated on eight fine-grained visual datasets, our approach improves classification accuracy by 0.73% on average under standard settings and by 6.5% under long-tail scenarios. It also significantly enhances vision-language alignment and preserves discriminative class-specific features.

This work addresses the challenge of detecting generative AI–produced content. We propose an unsupervised, interpretable embedding-space analysis method: semantic embeddings of text or images are extracted using pre-trained large language or multimodal models; subsequently, dimensionality reduction (e.g., PCA) uncovers an intrinsic, low-dimensional distributional shift between AI-generated and human-created samples—rendering them highly separable without supervision. This phenomenon is systematically validated for the first time and endowed with human-interpretable semantic meaning (e.g., topic coherence, syntactic redundancy). Experiments across diverse generative models—including ChatGPT, Gemini, and Stable Diffusion—demonstrate that high-accuracy separation is achieved solely from raw embeddings and unsupervised projection, without fine-tuning, labeled data, or model-specific detectors. Our approach thus significantly enhances both generalizability and interpretability of AI-content detection.

This work addresses the challenge of efficiently generating high-quality training data required for supervised learning in text-to-image generation models. We propose the Guided Adversarial Prompts (GAP) framework—a closed-loop data generation system integrating three core mechanisms: (1) adversarial prompt optimization guided by supervised model loss, (2) target distribution alignment via feature matching or discriminator-based guidance, and (3) online feedback adaptation. GAP is the first method to synergistically couple adversarial generation with explicit distributional constraints, shifting data synthesis from open-loop, static prompting to closed-loop, adaptive refinement. Empirical evaluation across diverse settings—including multi-task learning, heterogeneous model architectures, and distribution shifts (e.g., spurious correlations, unseen domains)—demonstrates substantial improvements in downstream model generalization. Data utilization efficiency increases by up to 3.2× compared to baseline approaches.

This work addresses the challenge of precise control in conditional discrete generative models when confronted with unseen condition combinations. The authors propose a theory-driven, composable discrete generation framework that integrates parallel token prediction with an absorbing diffusion mechanism and a concept-weighted conditional fusion strategy. This approach enables accurate modeling of an arbitrary number and combination of conditions while unifying mask-based generation within the same paradigm. Leveraging compositional vocabularies derived from VQ-VAE/VQ-GAN, the method achieves a 63.4% average reduction in error rate, a 9.58 improvement in FID, and 2.3–12× faster inference across three datasets. Furthermore, it successfully extends to pretrained text-to-image models, enabling fine-grained controllable generation.

Existing generative models lack a unified theoretical framework for concept manipulation and behavioral alignment in intermediate representations. This work proposes MidSteer, a general-purpose steering method based on affine transformations, which unifies concept erasure (e.g., LEACE) and concept switching under a closed-form affine intervention theory for the first time. MidSteer enables minimal-perturbation, targeted modifications to intermediate layer activations while circumventing restrictive assumptions inherent in prior approaches. The method is applicable across both diffusion models and large language models, demonstrating consistent superiority over existing techniques across diverse tasks, modalities, and architectures, thereby validating its effectiveness and broad generality.

Existing text-to-image personalization methods suffer from a trade-off between fidelity and editability due to the shared conditioning pathway for subject identity and scene context. This work proposes the Decoupled Guidance (DeGu) framework, which explicitly identifies and validates this conditioning entanglement problem for the first time. DeGu employs a dual-path architecture to separately process identity and contextual information and introduces a spatial mixing mechanism for semantically aware dynamic fusion. The framework is plug-and-play, requiring no modification to the backbone model, and is compatible with mainstream generative architectures such as DiT. It enables flexible adjustment of the fidelity–editability balance during inference and achieves significant performance gains across diverse personalization settings and models.

This work addresses the high deployment costs of multimodal models and their reliance on scarce, precisely aligned data, compounded by inefficiencies in existing generative data augmentation methods that introduce information redundancy and computational overhead through decode-encode cycles. To overcome these limitations, the authors propose leveraging undecoded latent representations from generative models as privileged synthetic data and introduce a Multi-level Explicit Synesthetic Simulation mechanism (MESSy) alongside Direct Latent Augmentation (DLA) to enable efficient cross-modal knowledge transfer. By circumventing conventional decode-encode pipelines and employing predictive knowledge distillation, the approach endows unimodal visual student models with an intrinsic capacity to align with unseen physical attributes. Experiments demonstrate that the proposed framework substantially outperforms current data augmentation and knowledge distillation techniques, yielding high-accuracy visual models imbued with synesthetic latent structures.

Deep generative models—such as diffusion models and Transformers—are often opaque black boxes, lacking interpretability and controllability. Method: We propose the first theoretical framework grounded in the principle of causal minimality to establish a principled foundation for generative model interpretability. Our approach constructs a hierarchical selection model and rigorously proves, under sparsity/compression constraints, that latent variables are causally identifiable and recover the true generative mechanism. Integrating causal inference, sparse coding, and interpretability-aware training, we achieve the first causal identification and decomposition of latent representations in mainstream vision and language generative models. Results: The extracted concept graphs enable fine-grained semantic manipulation. Extensive experiments across multiple benchmarks validate both the causal identifiability of latent variables and their cross-task generalizability for controllable generation.