Answers are grounded in this paper and cite their sources.
Camouflaged Object Detection (COD) is a challenging task in computer vision, as camouflaged objects share extremely subtle differences with their surroundings, often leading to missed detections and false alarms. Recently, diffusion models have shown unique advantages for COD. By modeling object distributions in the latent space and progressively refining denoising, they capture subtle foreground–background differences of camouflaged objects. However, most methods rely mainly on visual features and underutilize semantic priors, which can cause ambiguous object semantics and, during iteration, error accumulation leading to blurred boundaries or spatial misalignment. To overcome these limitations, we propose a Text-aware Visual Prompt-driven Diffusion (TVP-Diff) framework, which introduces textual priors to enhance semantic representations and incorporates visual prompts to complement the lack of spatial detail in text guidance. Specifically, TVP-Diff consists of a Text-aware Visual Prompt Generator (TVPG) and a Visual Prompt-driven Diffusion (VP-Diff) network. TVPG leverages contrastive learning to align textual and visual features, generating visual prompts that encode both semantic and spatial information. These prompts are injected into VP-Diff and processed through four hierarchical prompt fusion modules (PromptC1–C4), providing layered guidance from shallow textures to deep semantics, thereby improving the precision of diffusion denoising and the clarity of the generated results. Extensive experimental results on three COD benchmark datasets demonstrate that TVP-Diff consistently surpasses 26 state-of-the-art methods across multiple metrics, with MAE scores of 0.014 on COD10K and 0.029 on CAMO, confirming its robustness and effectiveness in complex camouflaged scenarios. The code of TVP-Diff will be released athttps://github.com/yanjin12321/TVP-Diff
No takes yet. Share an insight, caveat, or question.
Yan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: