RealCamo: Boosting Real Camouflage Synthesis with Layout Controls and Textual-Visual Guidance

TL;DR

RealCamo tackles the data bottleneck in camouflaged object detection by introducing a controllable camouflaged image synthesis framework that couples explicit layout controls with a textual-visual conditioning strategy. It leverages latent diffusion and ControlNet to enforce structural and semantic coherence, while a texture-oriented background retrieval module and a unified fine-grained textual prompt guide realistic camouflage. The paper introduces the $\mathrm{KL}_{BF}$ metric to quantify camouflage effectiveness and demonstrates downstream benefits by constructing SynCOD12K for COD training. Empirical results show competitive generation quality, improved semantic consistency, and tangible gains in COD performance, validating the approach for data augmentation and realism-focused CAM tasks.

Abstract

Camouflaged image generation (CIG) has recently emerged as an efficient alternative for acquiring high-quality training data for camouflaged object detection (COD). However, existing CIG methods still suffer from a substantial gap to real camouflaged imagery: generated images either lack sufficient camouflage due to weak visual similarity, or exhibit cluttered backgrounds that are semantically inconsistent with foreground targets. To address these limitations, we propose ReamCamo, a unified out-painting based framework for realistic camouflaged image generation. ReamCamo explicitly introduces additional layout controls to regulate global image structure, thereby improving semantic coherence between foreground objects and generated backgrounds. Moreover, we construct a multi-modal textual-visual condition by combining a unified fine-grained textual task description with texture-oriented background retrieval, which jointly guides the generation process to enhance visual fidelity and realism. To quantitatively assess camouflage quality, we further introduce a background-foreground distribution divergence metric that measures the effectiveness of camouflage in generated images. Extensive experiments and visualizations demonstrate the effectiveness of our proposed framework.

Figures (6)

• Figure 1: Motivation and qualitative comparison of camouflaged image synthesis.(a) LAKE-RED CVPR2024LAKERED fails to achieve effective camouflage, with targets remaining visually distinguishable, whereas our method produces images with well-camouflaged targets. (b) CamoAny CVPR2025CamoAny achieves visual camouflage but suffers from weak semantic consistency between the target and the generated background, in contrast, our method preserves semantic coherence, resulting in more realistic scenes. (c) Our approach is capable of generating diverse and realistic camouflaged images while jointly satisfying visual similarity and semantic consistency.
• Figure 2: The proposed KL divergence based metric. Existing generative evaluation metrics, such as FID and KID, are insufficient to measure the similarity between foreground and background, while our proposed $\mathrm{KL}_{BF}$ provides a quantitative assessment of camouflage. Top: Image with a higher $\mathrm{KL}_{BF}$. Bottom: Our generated image with a lower $\mathrm{KL}_{BF}$, more camouflaged.
• Figure 3: The pipeline of our camouflaged image generation framework RealCamo. Explicit layout controls (i.e., the contrast, depth, and HED controls) are extracted by the Layout Controls Generator (LCG) and injected to ControlNet, while the multi-modal condition produced by the Textual-Visual Condition Generator (TVCG) is used as a guidance in refining the camouflage generation process.
• Figure 4: Qualitative comparison with existing SOTA methods including style transfer based (column 3 to 4), out-painting based (column 5 to 9), and text-driven (column 10) approaches. Our results are demonstrated in the rightmost column with red box.
• Figure 5: Qualitative results of ablation study on main components of our proposed RealCamo framework.