Rethinking Camouflage Image Generation towards a Training-Free Paradigm

Haodong Yang; Zhongling Huang; Gong Cheng

cs.CVSep 13, 2026

Rethinking Camouflage Image Generation towards a Training-Free Paradigm

Authors: Haodong Yang, Zhongling Huang, Gong Cheng

Organizations: School of Automation, Northwestern Polytechnical University, Xi’an, China.

Abstract

Camouflage image generation (CIG) aims to synthesize realistic camouflaged images by blending foreground objects into concealment-compatible background contexts. Achieving this objective requires jointly satisfying three coupled requirements: foreground preservation to retain target integrity, semantic compatibility to select plausible concealment contexts, and appearance assimilation to reduce visual discrepancies. Recent approaches predominantly rely on task-specific training on camouflage datasets to address these requirements, incurring substantial computational cost and limiting generalization beyond the training domain. To address these limitations, we formulate training-free CIG as a concealment-oriented paradigm that preserves the target while reducing its perceptual separability from the synthesized surroundings, rather than maintaining its visual prominence, without parameter updates. We instantiate this paradigm with FreeCam based on a frozen inpainting diffusion framework to preserve the foreground. Within this framework, a Contextual Reasoning Module exploits frozen multimodal priors to infer an environment favorable to concealment, thereby promoting semantic compatibility, while an Intrinsic Appearance Module extracts low-level color and texture cues from the foreground to guide background synthesis toward appearance assimilation. Extensive experiments demonstrate that FreeCam achieves state-of-the-art generation quality and camouflage effectiveness without task-specific training, while its generated images provide synthetic supervision for camouflaged object detection and reduce target detectability under general object detectors.

Figures & tables

Figure 1: (a) Illustration of the proposed training-free CIG paradigm, in which contextual reasoning and intrinsic appearance conditioning guide a frozen diffusion model toward semantic compatibility and appearance assimilation. (b) Comparison with CNN-trained and diffusion-trained CIG methods across seven normalized axes: efficiency; inverted KID on COD, general data, and the overall set; and $S_{\mathrm{b}}$ on the same three evaluation sets. All axes are oriented such that larger values are better. (c–d) The generated results on (c) COD and (d) general data validate the capability of the proposed paradigm.

Figure 2: (a–b) Training-based CNN and diffusion methods rely on COD-specific optimization. (c) Generic training-free generation steers frozen priors and avoids task-specific optimization costs but is not concealment-oriented.

Figure 3: Framework of FreeCam. Background-only inpainting preserves the original foreground. CRM addresses semantic compatibility by constructing a concealment-oriented semantic condition, whereas IAM addresses appearance assimilation by extracting foreground color and texture cues as an appearance condition. The two conditions jointly guide the frozen diffusion model during background synthesis.

Figure 4: Visual comparison of images generated with and without the environment set. The text above each image denotes the inferred environment.

Figure 5: Visual comparison of images generated with and without IAM. The text above each image denotes the inferred environment.

			Camouflaged Objects		Salient Objects		General Objects		Overall
Method	Paradigm	Training	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
LCGNet Li et al. (2022) [TMM]	CNN	Yes	129.80	0.0504	136.24	0.0597	132.64	0.0548	129.88	0.0550
TFill Zheng et al. (2022) [CVPR]	CNN	Yes	63.74	0.0336	96.91	0.0453	122.44	0.0747	80.39	0.0438
LDM Rombach et al. (2022) [CVPR]	Diffusion	Yes	58.65	0.0380	107.38	0.0524	129.04	0.0748	84.48	0.0488
LAKE-RED Zhao et al. (2024) [CVPR]	Diffusion	Yes	39.55	0.0212	88.70	0.0428	102.67	0.0555	64.27	0.0355
FACIG Chen et al. (2025) [ICME]	Diffusion	Yes	27.61	0.0099	82.23	0.0326	96.94	0.0503	52.87	0.0229

Table 1: Quantitative comparison on LAKE-RED using FID and KID across the camouflaged objects, salient objects, general objects, and overall subsets. Lower scores indicate better generation quality. Bold and underlined values indicate the best and second-best results in each column, respectively.

			Camouflaged Objects		Salient Objects		General Objects		Overall
Method	Paradigm	Training	$S_{\mathrm{Rf}}$ $\uparrow$	$S_{\mathrm{b}}$ $\uparrow$	$S_{\mathrm{Rf}}$ $\uparrow$	$S_{\mathrm{b}}$ $\uparrow$	$S_{\mathrm{Rf}}$ $\uparrow$	$S_{\mathrm{b}}$ $\uparrow$	$S_{\mathrm{Rf}}$ $\uparrow$	$S_{\mathrm{b}}$ $\uparrow$
LCGNet Li et al. (2022) [TMM]	CNN	Yes	0.2724	0.4223	0.2845	0.3923	0.2212	0.4161	0.2618	0.4030
TFill Zheng et al. (2022) [CVPR]	CNN	Yes	0.2862	0.2771	0.2604	0.2968	0.2271	0.3130	0.2273	0.3164
LDM Rombach et al. (2022) [CVPR]	Diffusion	Yes	0.3361	0.3480	0.4012	0.4539	0.3692	0.4317	0.3804	0.4576
LAKE-RED Zhao et al. (2024) [CVPR]	Diffusion	Yes	0.3715	0.4208	0.3941	0.4764	0.3209	0.4621	0.3533	0.4436
FACIG Chen et al. (2025) [ICME]	Diffusion	Yes	0.5725	0.4515	0.3992	0.4843	0.3575	0.4520	0.4057	0.4586

Table 2: Quantitative comparison of camouflage effectiveness on the LAKE-RED dataset using

S_{\mathrm{Rf}}

and

S_{\mathrm{b}}

. Since these two metrics are not reported in the original papers of the compared CIG methods, all results are re-evaluated by us.

\uparrow

indicates higher is better. Bold and underlined values indicate the best and second-best results in each column, respectively.

Figure 6: Qualitative comparison with representative CIG methods. The first two columns show the input images and masks.

Method	Preprocessing	Inference	Total
LCGNet Li et al. (2022)	70 min	0.05 s	70 min
LAKE-RED Zhao et al. (2024)	5 h	9.8 s	5 h
FACIG Chen et al. (2025)	6 h	3.7 s	6 h
CamAny Das and Gopalakrishnan (2025)	$>$ 1 day	13.2 s	$>$ 1 day
CT-CIG Qian et al. (2026)	8 h	23.1 s	8 h
FreeCam SD1.5	0 s	8.3 s	8.3 s

Table 3: Comparison of preprocessing time and inference time. Preprocessing time includes the time required for fine-tuning or re-training. Bold indicates the lowest total time.

CRM	IAM-Color	IAM-Texture	KID $\downarrow$	$S_{\mathrm{b}}\uparrow$
✗	✗	✗	13.326	0.2421
IAM-based ablation without CRM
✗	✓	✗	8.981	0.7096
✗	✗	✓	11.424	0.6818
✗	✓	✓	8.655	0.7631
CRM-based ablation

Table 4: Main ablation of FreeCam. The table summarizes the route from a frozen prior to FreeCam by isolating CRM and IAM. KID is scaled by

10^{3}

. Lower KID and higher

S_{\mathrm{b}}

are better. Bold indicates the best value for each metric.

Figure 7: Qualitative component ablation showing the progression from the frozen baseline to the complete FreeCam framework.

Prompt	KID $\downarrow$	$S_{\mathrm{b}}\uparrow$
Base (IAM)	8.655	0.7631
+Category	8.135	0.7697
+Description	7.315	0.7846
+Environment w/o Env.Set	6.859	0.7959
+Env.Set	6.926	0.8029

Table 5: Ablation analysis of CRM. (a) evaluates prompt composition, and (b) compares environment-selection strategies in the CRM scorer. KID is scaled by

10^{3}

. Lower KID and higher

S_{\mathrm{b}}

are better. Bold indicates the best value in each subtable.

Filter	KID $\downarrow$	$S_{\mathrm{b}}\uparrow$
Sobel	7.099	0.7994
Laplacian	7.203	0.8011
Canny	7.144	0.8002
DoG	7.065	0.8026
High-pass	6.926	0.8029

Table 6: Ablation analysis of IAM. (a) compares filter choices for texture extraction, and (b) evaluates cumulative IAM components. KID is scaled by

10^{3}

. Lower KID and higher

S_{\mathrm{b}}

are better. Bold indicates the best value in each subtable.

Figure 8: Sensitivity analysis of hyperparameters and architectural choices in terms of KID ( $\downarrow$ ) and $S_{\mathrm{b}}$ ( $\uparrow$ ).

Figure 9: Effect of the CLIP feature layer used for intrinsic appearance extraction. Intermediate features provide the best balance between texture preservation and semantic abstraction. The labels 5–12 denote the corresponding layers of CLIP.

Figure 10: Fine-grained analysis of the starting timestep for intrinsic appearance injection. Starting at an intermediate denoising stage improves appearance consistency while preserving the global scene layout. The labels denote the injection timesteps.

Model	KID $\downarrow$	$S_{\mathrm{b}}\uparrow$
CLIP (ours)	6.926	0.8029
MAE He et al. (2022)	8.058	0.7567
SigLIP Zhai et al. (2023)	6.676	0.8093
DINOv3 Siméoni et al. (2025)	6.663	0.8159

Table 7: Model-choice sensitivity of FreeCam. (a) evaluates vision encoders used by IAM, and (b) evaluates MLLMs used by CRM. KID is scaled by

10^{3}

. Lower KID and higher

S_{\mathrm{b}}

are better. Bold indicates the best value in each subtable.

	CHAMELEON				CAMO				COD10K				NC4K
Methods	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$M\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$M\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$M\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$M\downarrow$
SINetv2 Fan et al. (2021)
Base	0.881	0.933	0.796	0.035	0.806	0.858	0.708	0.081	0.802	0.876	0.656	0.040	0.837	0.887	0.746	0.053
+LAKE-RED 100%	0.886	0.934	0.807	0.033	0.811	0.869	0.727	0.078	0.806	0.881	0.661	0.038	0.842	0.901	0.763	0.049
+FreeCam 10%	0.890	0.940	0.811	0.029	0.818	0.875	0.728	0.070	0.812	0.885	0.676	0.031	0.848	0.903	0.767	0.042
+FreeCam 30%	0.894	0.952	0.822	0.027	0.829	0.884	0.739	0.067	0.818	0.890	0.684	0.027	0.856	0.909	0.772	0.036

Table 8: Camouflaged object detection results on four COD benchmarks under different synthetic-data augmentation ratios. Higher

S_{\alpha}

E_{\phi}

, and

F_{\beta}^{w}

and lower

M

are better. Bold and underlined values denote the best and second-best results within each backbone, respectively.

	Average Precision			Recall		Overlap
Method	$\mathrm{AP}_{50}\uparrow$	$\mathrm{AP}_{75}\uparrow$	$\mathrm{mAP}\uparrow$	$\mathrm{R}_{50}\uparrow$	$\mathrm{R}_{75}\uparrow$	$\mathrm{IoU}\uparrow$
Overall
YOLOv10-X	0.393	0.348	0.332	0.680	0.614	0.638
+FreeCam SD1.5	0.372	0.224	0.224	0.533	0.383	0.473
degradation	$\downarrow 0.021$	$\downarrow 0.124$	$\downarrow 0.108$	$\downarrow 0.147$	$\downarrow 0.231$	$\downarrow 0.165$
G-DINO-B	0.394	0.331	0.311	0.785	0.712	0.738

Table 9: Detection degradation on the overall set and three subsets using FreeCam SD1.5 . The degradation rows report changes from Original to FreeCam for each evaluated detector. Red decreases and blue increases indicate better and worse camouflage, respectively.

Figure 11: Effect of the CRM environment-set size across five normalized metrics. KID axes are inverse-normalized so that higher values are better.

Figure 12: Generation results across different frozen diffusion backbones. FreeCam is applied to (a) SD2.0 and (b) SDXL using the same training-free formulation, illustrating that its semantic and appearance conditioning can be transferred across pretrained generative models.

Figure 13: Controllability analysis under user-designated environments. (a) Semantic-match settings use environments compatible with the foreground object, whereas (b) semantic-mismatch settings force less compatible contexts to test controllability.

Figure 14: Cross-scenario generalization of FreeCam. The examples cover (a) military imagery and (b) underwater imagery, showing camouflage synthesis beyond the COD-style evaluation setting.

Figure 15: Random-seed sensitivity of FreeCam. For each input, five outputs are generated under fixed semantic and appearance conditions using different random seeds.

Figure 16: Feasibility of FreeCam on larger image-editing models. The comparison shows the examples generated with FLUX.1 Fill Labs (2024) and Qwen-Image-Edit Wu et al. (2025) .

Figure 17: Limitation analysis. The examples show typical failure cases, mainly involving highly salient targets such as human targets and aircraft in remote sensing imagery.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 474 papers

To Blend In, First Decouple: Rethinking Camouflage Image Generation via Context-Decoupled Representations
Jul 20, 2026Wenzhuang Wang, Yifan Zhao, Mingcan Ma +4Disentangled Representation Learning Conditional Image Generation
State Key Laboratory of Virtual Reality Technology and Systems, SCSE & QRI, Beihang University · Geely Automobile Research Institute (Ningbo) Co., Ltd, AI Center, Geely
SeamCam: Quantifying Seamless Camouflage via Multi-Cue Visual Detectability
May 15, 2026Amin Karimi Monsefi, Abolfazl Meyarian, Mridul Khurana +6Camouflaged Object Detection
The Ohio State University · Path Robotics, USA · Virginia Tech +2
Concentration, Not Uncertainty: Why Targeted Synthetic Data Doesn't Help Camouflaged Object Detection
Oct 7, 2026Akshat Dobhal, Sanjay SinghSynthetic Data Selection Synthetic Data Generation
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Jul 20, 2026J/K move · Enter open · S save

To Blend In, First Decouple: Rethinking Camouflage Image Generation via Context-Decoupled Representations

Wenzhuang Wang, Yifan Zhao, Mingcan Ma, Yunlong Che+3

State Key Laboratory of Virtual Reality Technology and Systems, SCSE & QRI, Beihang University · Geely Automobile Research Institute (Ningbo) Co., Ltd, AI Center, Geely

Disentangled Representation Learning Conditional Image Generation Image Generation Controllable Image Generation

Open paper AskPDF arXiv

Camouflage image generation (CIG) focuses on generating visually concealed objects that seamlessly blend into their backgrounds. Existing methods typically follow either background-guided paradigms that adapt object appearance via style transfer, or foreground-guided strategies that outpaint surrounding regions conditioned on object features. However, they still suffer from appearance discrepancy and background artifacts. We attribute these limitations to cross-context representation leakage, where object and background cues are entangled in a coupled conditional space, resulting in ambiguous control and degraded camouflage fidelity. To tackle this, we propose a new context-decoupled generative paradigm, termed CamoDreamer, which aims to isolate contextual conditional guidance and explicitly decouple latent camouflage features into coordinated object and background control streams. First, a Contrast-aware Contextual Bridge is designed to model cross-context discrepancies and construct contrast-aware dual conditional guidance. Second, Context-Decoupled Assimilation Streams are employed to separate generative interactions conditioned on the dual guidance, while facilitating background rendering with target-aware cues in the latent space. Finally, a Frequency-Adaptive Contextual Blend module integrates complementary high-frequency textures and low-frequency structures from decoupled features to improve holistic coherence. Extensive experiments demonstrate that CamoDreamer consistently outperforms existing methods with a substantial margin, while maintaining a relatively lightweight design.

Abstract

Image Generation

Explore similar work

To Blend In, First Decouple: Rethinking Camouflage Image Generation via Context-Decoupled Representations

SeamCam: Quantifying Seamless Camouflage via Multi-Cue Visual Detectability

Concentration, Not Uncertainty: Why Targeted Synthetic Data Doesn't Help Camouflaged Object Detection