Research map / Image generation / Image restoration
Generative Quanta Color Imaging
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 AtomODE, a neural-ODE-driven filter-atom exposure synthesis module that corrects exposure of a single 1-bit image so that existing colorizers (Pix2Pix-based) can produce color images from single binary frames or bursts.
Problem
Colorizing a single binary frame from a single-photon camera is hard because binary images suffer from large, unknown exposure variation (over/under-exposure), which degrades standard colorizers trained on well-exposed data.
Approach
Model the continuum of exposures by synthesizing convolutional filters via atom-coefficient decomposition whose atoms evolve under a neural ODE over the exposure interval, so one observation can be corrected to a target exposure and fed to a fixed-exposure colorizer (single-image or burst).
Key results
- On AFHQ cat, AtomODE-CycleGAN reaches FID 60.54 and AtomODE-Pix2Pix 57.88 versus 61.89 for Pix2Pix-DNI in exposure-corrected colorization.
- AtomODE-Pix2Pix attains relative linearity of ~0.9995-0.9998 across datasets, far above DLOW (~0.95-0.98) and SAVI2I.
- On real data, fine-tuning the colorizer on 100 real CMOS image pairs significantly improves colorization quality, per qualitative results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AtomODE-CycleGAN instead of DLOW for exposure synthesis | AFHQ Cat · AFHQ Cat | MSE: 2.44 -> 1.15 | Helped |
| use AtomODE-Pix2Pix instead of Pix2Pix-DNI for exposure synthesis | AFHQ Cat · AFHQ Cat | RL: 0.9816 -> 0.9997 | Helped |
| use AtomODE-CycleGAN instead of DLOW for exposure synthesis | AFHQ Dog · AFHQ Dog | MSE: 2.32 -> 1.18 | Helped |
| use AtomODE-CycleGAN instead of Pix2Pix-DNI for exposure synthesis | AFHQ Dog · AFHQ Dog | FID: 95.43 -> 87.43 | Helped |
| use AtomODE-CycleGAN instead of DLOW for exposure synthesis | AFHQ Wild · AFHQ Wild | MSE: 2.75 -> 1.23 | Helped |
| use AtomODE-CycleGAN instead of Pix2Pix-DNI for exposure synthesis | AFHQ Wild · AFHQ Wild | FID: 39.29 -> 35.75 | Helped |
| use AtomODE-CycleGAN instead of DLOW for exposure synthesis | CelebA-HQ Male · CelebA-HQ Male | MSE: 1.87 -> 1.58 | Helped |
| use AtomODE-Pix2Pix instead of Pix2Pix-DNI for exposure synthesis | CelebA-HQ Male · CelebA-HQ Male | RL: 0.9563 -> 0.9995 | Helped |
| use AtomODE-CycleGAN instead of DLOW for exposure synthesis | CelebA-HQ Female · CelebA-HQ Female | MSE: 1.37 -> 1.87 | Hurt |
| increase standard colorization training time and dataset size | Standard colorization · Colorization training comparison | FID: 62.34 -> 54.25 | Helped |
| increase standard colorization generator size | Standard colorization · Colorization training comparison | FID: 62.34 -> 74.58 | Hurt |
| use the proposed method instead of standard colorization | Colorization training comparison · Colorization training comparison | FID: 62.34 -> 49.88 | Helped |
Surprising findings
- Simply increasing colorizer model size or training time with exposure augmentation did not match the proposed exposure correction (Table 1: FID 74.58 for larger model vs 49.88 proposed).
Limitations
- Trained mostly on synthetic data; real-sensor performance relies on a small fine-tuning set of 100 pairs.
- Colorization quality on CelebA-HQ FID remains high for some AtomODE variants (e.g., 140.23 for Pix2Pix on dog-like/female sets).
Experimental setup
Datasets: AFHQ CelebA-HQ
Benchmarks: AFHQ CelebA-HQ
Metrics: MSE Relative Linearity FID
Baselines: Pix2Pix-DNI CycleGAN-DNI SAVI2I DLOW standard colorization with exposure augmentation
Builds on: Pix2Pix CycleGAN neural ODE atom-coefficient filter decomposition DNI Cross Non-Local Fusion
Closest papers by meaning
- Color Alignment in Diffusion (CVPR 2025 · similarity 0.85)
- bit2bit: 1-bit quanta video reconstruction via self-supervised photon prediction (NeurIPS 2024 · similarity 0.83)
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.83)
- GCC: Generative Color Constancy via Diffusing a Color Checker (CVPR 2025 · similarity 0.83)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.83)
- Image Inpainting via Tractable Steering of Diffusion Models (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.