AnchorSteer is proposed, a training-free framework that exerts fine-grained control over both initialization and denoising trajectory that consistently outperforms existing baselines in text--image alignment while preserving high visual quality.
Abstract
While text-to-image diffusion models achieve impressive visual quality, they frequently struggle to maintain precise alignment with complex compositional prompts. An effective strategy is to improve the inference process of diffusion models, thereby better leveraging their pretrained priors to address misalignment. Existing training-free methods can be divided into two categories. The first category focuses on improving the randomly sampled initial noise, either performing costly search over noise pools or manipulating sampled noise without ensuring reliable semantic injection. The second category focuses on improving the denoising trajectory, lacking explicit mechanisms to timely diagnose and correct semantic errors. we propose \textbf{AnchorSteer}, a training-free framework that exerts fine-grained control over \textbf{both initialization} and \textbf{the denoising trajectory}. AnchorSteer consists of two synergistic components: \textbf{Semantic Anchoring} replaces uninformative Gaussian noise with text-aligned initializations via CLIP-based prior extraction and a novel Latent-Prior Score Distillation Sampling (LP-SDS) objective. Specifically, LP-SDS distills CLIP visual priors into the knowledge distribution of diffusion models, mitigating the domain gap between CLIP-based priors and diffusion-based priors. \textbf{Reflective Steering} transforms passive denoising with an active Think--Erase--Retouch loop that enables mid-generation self-correction. It leverages VLM-based diagnosis to detect semantic deviations and performs targeted latent refinement to suppress erroneous content and recover missing attributes. Extensive experiments on GenEval and T2I-CompBench++ demonstrate that AnchorSteer consistently outperforms existing baselines in text--image alignment while preserving high visual quality.
Text-to-Image diffusion models are highly effective but remain heavily sensitive to the initial noise. This sensitivity causes significant instability in personalization tasks, where maintaining a specific subject's identity is crucial. While inference-based methods like the W+ Adapter offer efficient alternatives to costly fine-tuning, they suffer from structural conflicts between identity preservation and prompt consistency depending on this noise. In this study, we address this issue by proposing an automated discrete "Latent Space Exploration" framework utilizing random search to optimize seed selection. We compare our discrete seed optimization approach against "Initial Noise Selection," a continuous optimization method that modifies the noise tensor directly via gradient descent. We define a multi-objective scoring function integrating text consistency (CLIP), identity preservation (ArcFace), and structural validity (MTCNN). Quantitative experiments reveal a critical trade-off: while continuous optimization preserves identity competitively, it frequently degrades text consistency by ignoring prompt contexts like clothing or backgrounds. In contrast, our discrete exploration achieves a superior balance, ensuring a 100% face detection rate while maximizing both identity fidelity and text alignment. Furthermore, a subjective evaluation with 151 participants confirms that our method yields significantly higher overall visual quality and prompt fidelity. We conclude that discrete seed optimization offers a robust and practical solution for personalized generation.
Yu Yamamoto, Qiu Chen· International Conference on...· 0 citations
Text-to-image diffusion models can be misused to generate harmful content through adversarial or paraphrased prompts that bypass built-in safety mechanisms. Existing concept erasure methods often suffer from limited robustness against adversarial prompts, degradation of benign generation quality, or reliance on inference-time interventions that introduce persistent computational overhead. To address these limitations, we formulate concept erasure as a domain alignment problem in the text representation space. We propose a lightweight Text Encoder Alignment framework (TEA) that fine-tunes only the text encoder while keeping the generative backbone fully frozen. Given concept--anchor prompt pairs, our method trains a discriminator to distinguish token-level representations of concept-containing prompts from those of safe anchor prompts, while updating the text encoder to make these representations indistinguishable. TEA introduces zero inference-time overhead and requires only a small number of fine-tuning steps, making it highly efficient to deploy at scale. Despite this efficiency, TEA achieves state-of-the-art erasure robustness against black-box and white-box adversarial attacks on Stable Diffusion v1.4, while preserving generation quality on benign prompts. Furthermore, TEA is model-agnostic and achieves the lowest attack success rate on Stable Diffusion v3.5, extending concept erasure to a Rectified Flow Transformer architecture with T5 conditioning where prior methods remain largely unexplored. Code is available at \href{https://github.com/alirezafarashah/TEA.git}{https://github.com/alirezafarashah/TEA.git}
Backbone training-free video editing built on pre-trained text-to-image (T2I) diffusion models enables lightweight, prompt-driven edits without additional finetuning. A critical yet often overlooked factor is cross-frame latent selection during DDIM inversion, which largely determines spatiotemporal coherence in the subsequent denoising process. Existing pipelines typically rely on static, heuristic keyframe policies and temperature-softmax responsibilities, yielding unscalability i.e., numerical sensitivity and scale bias, that degrades generalization across diverse scenes. In this paper, we propose VIVID (Variational Inference for Video editing with Image Diffusion), an uncertainty-aware variational latent anchoring module that dynamically selects informative frames and compresses cross-frame latents into a compact set of semantic anchors. VIVID learns stable assignments via a variational objective with contrastive alignment and prior regularization, producing anchors that preserve spatial details while enforcing temporal continuity, and can be plugged into existing backbone training-free T2I-based video editing frameworks as a drop-in replacement for heuristic selection. Extensive experiments on standard benchmarks and in-the-wild videos demonstrate that VIVID achieves state-of-the-art inversion fidelity, editing quality, and temporal consistency, while reducing memory and runtime compared with prior backbone training-freebaselines. Code is released in: https://github.com/amasawa/VIVID.
Zhangkai Wu, Xuhui Fan, Zhongyuan Xie et al.· Proceedings of the 32nd ACM...· 0 citations
This paper proposes ElasticTTT, a novel framework that preserves the prior generative distribution and rescues generative elasticity in standard TTT, achieving state-of-the-art performance on one-shot video editing.
Diffusion models for text-to-image (T2I) synthesis, e.g. Stable Diffusion, generate visually realistic images, but often fail to capture the fine-grained semantic nuances of complex prompts. Typical failure modes include missing objects, incorrect attribute assignment e.g. wrong colors, distorted spatial relationships and inaccurate object counts. This paper proposes a VQA-Guided Text-to-Image Feedback Framework to address these limitations, substituting the traditional one-shot generation with an iterative self-correcting loop. In our framework, an input prompt is first decomposed into its semantic components (objects, properties, relationships) using natural language processing. Then a set of Visual Question Answering (VQA) questions is automatically generated stepwise to probe object existence, colors, spatial relations and quantities. A VQA model (BLIP) answers the questions on the initially generated image from Stable Diffusion. Whenever it detects a semantic mismatch, it automatically refines the prompt (for example, by adding explicit corrective instructions such as “ make sure the car is red ”), and feeds the revised prompt back into the diffusion model. This cycle is repeated until adequate semantic alignment is obtained, with the framework incorporating various assessment criteria (CLIP Score, VQA Accuracy, LPIPS, picture quality score, and consistency loss) integrated into a unified multi-metric evaluation framework. For our experiments on MS-COCO captions, DrawBench, and a custom set of challenging prompts, the proposed method resulted in the following improvements observed on our test sets: the average CLIP Score improved from 0.311 (baseline Stable Diffusion) to 0.364, a relative increase of 17%; VQA accuracy was above 87% for all prompts evaluated; and in a human preference study of 39 pairwise comparisons, the VQA-guided method was preferred in 76.9% of the cases. Ablation experiments demonstrated the beneficial contribution of many components of the framework. We show that VQA can serve as an effective semantic feedback to significantly enhance prompt-image alignment without retraining the diffusion model, providing a powerful, interpretable and self-correcting strategy for text-to-image production.
Recent advances in preference alignment for diffusion-based video generation, particularly via Direct Preference Optimization (DPO), have significantly improved visual quality. However, temporally sparse artifacts such as motion collapse, object flickering, and color oversaturation remain a major barrier to perceptual realism. Existing methods struggle with these issues due to two key limitations: (1) the preference attribution bottleneck, where offline human annotations are costly and fail to accurately capture learning dynamics, while online reward signals are rollout-aware but often unstable and biased; and (2) temporal credit misallocation, where uniformly applied supervision cannot effectively target the brief segments in which artifacts occur. To address these challenges, we propose concentrated Implicit Preference Optimization (cIPO), a post-training framework for video diffusion models. cIPO derives implicit preference signals directly from the denoising process: given a real video, the model adds forward noise and reconstructs it via iterative denoising, treating the original as the preferred sample and the reconstruction as the dispreferred one. This formulation captures inference-time errors without requiring human annotations or external reward models. Moreover, frame-level discrepancies between original and reconstructed videos reveal when failures occur. cIPO leverages this by computing temporal reconstruction errors and concentrating optimization on high-error segments, enabling more precise correction of failure-prone regions. Extensive experiments demonstrate that cIPO consistently enhances video authenticity and temporal coherence across multiple datasets, highlighting the effectiveness and efficiency of implicit preference with temporally concentrated optimization.
Henglin Liu, Fangyuan Kong, Jing Wang et al.· 0 citations