Research map / Image generation / Autoregressive and latent generation

LAKE-RED: Camouflaged Images Generation by Latent Background Knowledge Retrieval-Augmented Diffusion

CVPR 2024 · Image generation › Autoregressive and latent generation · approach: Classical generative image synthesis · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LAKE-RED, a latent diffusion inpainting model that generates camouflaged images from a foreground object alone by retrieving background knowledge from the VQ-VAE codebook and reconstructing background features, without manual background input.

Problem

Camouflaged image generation methods require a manually specified background, limiting diversity and scale and making it costly to create camouflaged training data for camouflaged vision perception.

Approach

Generate the background automatically from the foreground: foreground features (localized masked pooling via SLIC superpixels) query a pretrained VQ-VAE codebook through multi-head attention to retrieve background-aligned features, which are then used to reconstruct background features and enhance the diffusion condition, with an added background reconstruction loss.

Key results

• Overall FID 64.27 and KID 0.0355, versus 84.48 / 0.0488 for LDM and 80.39 / 0.0438 for TFill.
• Ablation: the three modules improve FID from 96.14 to 64.27 (about 33% relative) with about 0.01M extra parameters.
• User study with 20 participants ranked the method as the most natural and closest to real camouflaged data.

Measured effects reported in the paper

Surprising findings

• LCGNet obtains the lowest KID on camouflaged objects but has outlier FID on the general objects subset, which the authors attribute to discarding foreground features and relying on random backgrounds.

Limitations

• Relies on manually specified backgrounds for comparison baselines, and evaluation uses FID/KID against small real datasets rather than downstream task gains.
• Evaluation is generative quality only; no downstream camouflaged object detection results are reported in the text.
• Compute stated as GeForce RTX 3090 GPUs without an exact count.

Experimental setup

Datasets: COD10K CAMO NC4K DUTS DUT-OMRON COCO2017 Places365

Benchmarks: COD test sets (CO subset from CAMO, COD10K, NC4K) salient object datasets (DUTS, DUT-OMRON) COCO2017

Metrics: FID KID

Baselines: AB CI AdaIN DCI LCGNet TFill LDM RePaint-L

Builds on: Latent Diffusion Models RePaint VQ-VAE Masked Averaged Pooling

Closest papers by meaning

• Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering (CVPR 2025 · similarity 0.85)
• RealRAG: Retrieval-augmented Realistic Image Generation via Self-reflective Contrastive Learning (ICML 2025 · similarity 0.85)
• Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.85)
• INQUIRE: A Natural World Text-to-Image Retrieval Benchmark (NeurIPS 2024 · similarity 0.84)
• WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models (NeurIPS 2024 · similarity 0.84)
• GenIR: Generative Visual Feedback for Mental Image Retrieval (NeurIPS 2025 · similarity 0.84)