Research map / Image generation / Autoregressive and latent generation

Latent Space Imaging

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 2 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 Latent Space Imaging, which co-designs a binary single-pixel optical encoder with a digital encoder that maps measurements into the latent space of a frozen StyleGAN-XL face generator, enabling reconstruction and linear downstream tasks at extreme compression.

Problem

Conventional cameras capture dense pixel grids, which is bandwidth- and memory-heavy; the paper asks whether images can be encoded directly into a generative model's latent space with very few optical measurements, supporting reconstruction and downstream tasks.

Approach

Co-design a linear, quantized binary optical encoder (single-pixel camera masks) with a nonlinear digital encoder that targets the latent code of a frozen StyleGAN-XL, so that downstream tasks (attributes, landmarks, segmentation) can be solved by linear projections of that latent without full image reconstruction.

Key results

• Real single-pixel prototype achieves 1:512 reconstruction with 90.98% VGGFace match at 1:256 in simulation and 40-attribute accuracy of 80% at 1:16384 compression (4 scalars).
• At 1:256 compression, LSI reaches VGGFace 90.98%, Dlib 92.68%, FID 26.62, outperforming FSI-DL (3.30%, 13.15%, FID 123.5) and SAUNet (17.84%, 50.12%, FID 75.53).
• Landmark NME stays at 1.48 at 1:128 and degrades to 2.47 at 1:16384; segmentation F1 remains around 70% from 1:128 to 1:2048.

Measured effects reported in the paper

Surprising findings

• Linear probes on the latent space retain 80% attribute accuracy at 1:16384 compression, where reconstruction is no longer recognizable by face-recognition metrics.

Limitations

• Reconstruction is limited by the StyleGAN-XL face domain and fails on features absent in training such as hats, glasses, or makeup.
• Prototype uses a single-pixel DMD setup with time-multiplexed measurements, which limits frame rate; alternative hardware is not validated.
• Segmentation accuracy suffers for fine details such as earrings and accessories.

Experimental setup

Datasets: FFHQ CelebAHQ AFHQ

Benchmarks: CelebAHQ CelebAMask-style face segmentation via FaRL pseudo-labels CelebAHQ 40-attribute labels

Metrics: VGGFace similarity match rate Dlib match rate FID attribute accuracy segmentation F1 NME PSNR SSIM

Baselines: FSI-DL SAUNet Fourier single-pixel deep learning deep unfolding networks GAN-prior optimization

Builds on: StyleGAN-XL single-pixel imaging compressive sensing VQ-VAE straight-through estimator linear latent boundaries in GANs

Closest papers by meaning

• L2P: Unlocking Latent Potential for Pixel Generation (NeurIPS 2026 · similarity 0.86)
• Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.86)
• Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.85)
• PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.85)
• Global Latent Neural Rendering (CVPR 2024 · similarity 0.85)
• Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation (NeurIPS 2025 · similarity 0.85)