Research map / Image generation / Image restoration
Ultra-High-Definition Dynamic Multi-Exposure Image Fusion via Infinite Pixel Learning
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 Infinite Pixel Learning (IPL), a chunk-cache-quantization network for real-time 4K multi-exposure dynamic scene fusion on a single RTX 4090, and introduces the 4K-DMEF benchmark.
Problem
Multi-exposure image fusion for dynamic scenes is designed for low-resolution images and is inefficient or infeasible for generating full-resolution UHD (4K) outputs on a resource-constrained consumer GPU.
Approach
Borrow the LLM long-sequence ideas of chunking and KV caching: slice the image into blocks along channel, width and height, cache local attention features at inference (attention cache), quantize the cache to save memory, and add a dimensional rolling transformation (permute-based global mixing) to recover global context.
Key results
- On 4K-DMEF, IPL reaches PSNR 33.29, SSIM 0.9776, LPIPS 0.0427 at 0.0441s per image, versus BracketIRE at 27.44 PSNR (the second-best PSNR).
- Ablation shows removing DRTM drops PSNR from 34.78 to 32.82 (nearly 2 dB), and the attention cache cuts inference time from 0.1475s to 0.0441s at a cost of about 1.5 dB PSNR.
- On Kalantari, IPL gets PSNR 42.13 / SSIM 0.9902, roughly comparable to baselines (SCTNet 42.49 PSNR); on Mobile-HDR it gets 36.45 PSNR, below AHDRNet and HDR-Transformer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IPL instead of FSHDR | 4K-DMEF fusion · 4K-DMEF | PSNR: 22.85 -> 33.29 | Helped |
| Use IPL instead of FSHDR | 4K-DMEF fusion · 4K-DMEF | SSIM: 0.6595 -> 0.9776 | Helped |
| Use IPL instead of FSHDR | 4K-DMEF fusion · 4K-DMEF | LPIPS: 0.3520 -> 0.0427 | Helped |
| Use IPL instead of SCTNet | image fusion · Kalantari | PSNR: 42.49 -> 42.13 | Hurt |
| Use IPL instead of FSHDR | image fusion · Kalantari | SSIM: 0.9900 -> 0.9902 | Helped |
| Use IPL instead of HDR-Transf | image fusion · Mobile-HDR | PSNR: 36.67 -> 36.45 | Hurt |
| Use IPL instead of HDR-Transf | image fusion · Mobile-HDR | SSIM: 0.9809 -> 0.9902 | Helped |
| Add DAEM to the model (ablation) | IPL · 4K-DMEF | PSNR: 8.92 -> 34.78 | Helped |
| Add DRTM (ablation) | IPL with all DAEM dimensions · 4K-DMEF | PSNR: 32.82 -> 34.78 | Helped |
| Add attention cache (ablation) | IPL · 4K-DMEF | inference time: 0.1475s -> 0.0441s | Helped |
| Add attention cache with quantization compression (ablation) | IPL · 4K-DMEF | PSNR: 34.78 -> 33.29 | Hurt |
| Add 64 × 64 chunking and cache (ablation) | AHDRNet · 4K-DMEF | PSNR: 20.04 -> 22.81 | Helped |
Surprising findings
- Applying the chunk-cache pipeline to AHDRNet raised its PSNR from 20.04 to 22.81, suggesting the chunking itself helps baselines at UHD resolution.
- On the 4K-DMEF benchmark the baselines are run at downsampled maximum resolutions, so the reported PSNR gains partly reflect resolution differences rather than a like-for-like comparison.
Limitations
- Attention cache and quantization reduce PSNR by about 1.5 dB relative to the uncached model.
- Most baselines cannot run at full 4K, so they were evaluated on downsampled inputs and upsampled outputs, making the comparison non-uniform.
- Dataset is small (110 scenes) and synthetic exposure levels are generated by LECARM rather than captured.
Experimental setup
Datasets: 4K-DMEF Kalantari dataset Mobile-HDR dataset
Benchmarks: 4K-DMEF Kalantari Mobile-HDR
Metrics: PSNR SSIM LPIPS inference time MACs params
Baselines: GDC AHDRNet FSHDR HDR-Transformer SCTNet BracketIRE
Builds on: ChunkAttention KV cache MLP-Mixer LECARM
Closest papers by meaning
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.86)
- Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion (AAAI 2024 · similarity 0.85)
- UltraLED: Learning to See Everything in Ultra-High Dynamic Range Scenes (NeurIPS 2025 · similarity 0.83)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.83)
- Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond (CVPR 2025 · similarity 0.83)
- Neural Exposure Fusion for High-Dynamic Range Object Detection (CVPR 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.