Stable Diffusion models deliver photorealistic image synthesis but face computational bottlenecks that hinder deployment in resource-constrained environments. However, most existing distillation methods for diffusion models rely on point-wise feature matching, which often fails to preserve global semantic structure, or require additional and costly teacher finetuning. To address these limitations, we propose StableKOT — a novel knowledge distillation framework that leverages Optimal Transport (OT) theory. Still, direct OT alignment of high-dimensional feature maps is computationally prohibitive and sensitive to spatial noise. To overcome this issues, we transform teacher-student knowledge transfer into a distribution matching problem, applying OT to max-pooled features augmented with positional embeddings across U-Net layers. This captures geometric relationships in latent space while reducing computational overhead. Empirically, our method reduces parameters by 32.6% and accelerates inference 1.5×, while maintaining generative fidelity.
Bich-Nga Pham, Quoc-Truong Truong, Anh-Khoa Nguyen Vu et al.· International Conference on...· 0 citations
Compositional text-to-image generation requires faithful depiction of multiple objects with distinct visual attributes. While recent inference-time optimization methods have substantially improved attribute binding, entity neglect - where specified objects are absent from the generated image - remains an unresolved failure mode. We identify that existing syntax-guided attention objectives optimize for relative modifier-noun alignment but impose no constraint on the absolute activation magnitude of entity tokens, leaving neglected entities uncorrected even when binding objectives are satisfied. To address this, we propose a noun presence loss, a hinge-based objective that penalizes entity tokens with insufficient peak cross-attention activation during inference. A loss-ratio dynamic weight scales the presence loss proportionally to the binding loss magnitude at each timestep, preventing gradient conflict while maintaining stable optimization. Our method is training-free, requires no external components, and integrates with syntax-guided cross-attention optimization frameworks. Experiments on CC-500 and T2I-CompBench across color, shape, and texture categories demonstrate state-of-the-art performance on all metrics and attribute categories, confirming that entity presence optimization consistently improves compositional faithfulness and perceptual quality across diverse prompt types.
Trong-Tai Dam Vu, Vinh-Tiep Nguyen· International Conference on...· 0 citations
Text-guided image editing using rectified flow models such as Multimodal Diffusion Transformer (DiT) has demonstrated impressive generation quality. However, existing methods apply edits globally, inevitably modifying background regions unrelated to the intended semantic change. Our key observation is that the Text-to-Image (T2I) sub-block within the MM-DiT joint self-attention naturally encodes highly discriminative spatial localization signals. In this paper, we propose MaskFlow, a training-free framework for spatially localized image editing in rectified flow models. By strategically extracting and aggregating these attention maps from edit-relevant tokens during the standard ODE forward passes, MaskFlow automatically constructs a soft spatial mask. This mask confines semantic edits to the target region while perfectly preserving the original background. MaskFlow operates as a lightweight, plug-and-play extension without retraining the baseline model. Experiments on the FlowEdit benchmark show that MaskFlow reduces LPIPS by 28.1% relative to the baseline while maintaining competitive semantic alignment.
Trong-Tai Dam Vu, Vinh-Tiep Nguyen· International Conference on...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
