Computational imaging benchmark · 2026

Imaging
Bench

Can AI Solve Imaging Problems? Benchmarking Multimodal AI for Image Formation and Reconstruction

Ethan Chung1* Chuanjun Zheng1* Jiongze Yu2 Jasper Tan3 Jingxi Li3 Haopeng Zhang4 Zhengzhong Tu2 Huaijin Chen1

1CIRP Lab, University of Hawaiʻi at Mānoa 2Texas A&M University 3Glass Imaging 4University of North Carolina at Charlotte

* Equal contribution

Leaderboard Paper SOON Code SOON Dataset SOON

Overview

Has AI learned how images are formed?

Vision–language models and image-editing models have strong semantic visual skills. Whether they have learned the image formation behind scientific and computational imaging, and can invert it, is largely unexplored.

ImagingBench gives a model a description of the image-formation process, the input measurements and the task context, and asks it to produce the target image directly: no external code execution and no task-specific optimization. That isolates how much of the physics the model has actually internalized.

22
Imaging tasks
5
Categories across the pipeline
3
Protocols: Expert · Planner · Forward
6
Image models + specialized baselines
~100
Instances per subtask
2,665
Human judgments calibrating the score
ImagingBench overview: (a) task space across five categories with example inputs, (b) radar plots of per-task scores under Expert, Planner and Forward, (c) aggregate model performance and score versus release date.
(a) Diverse task space across five computational imaging categories. (b) Per-task unified scores for each protocol. (c) Aggregate performance by model, and score against model release date.

Tasks

The full imaging pipeline

Tasks follow the imaging pipeline, from the optics that form an image to the sensor, the processing chain and the reconstruction that recovers a scene. Twenty tasks produce images and feed the leaderboard; lens design and camera calibration produce text and are scored separately.

01

Ray & Wave Optics

Designing and inverting the optical front end: how light propagates through lenses and diffractive elements.

  • Computer-generated holography Image
  • Lens design Text
02

Image Signal Processing

The camera's processing chain: from raw sensor values to a clean, correctly coloured photograph.

  • Demosaicking · Hexadeca / Quad / Regular Image
  • HDR reconstruction Image
  • White balance Image
  • Denoising Image
03

Inverse Reconstruction

Undoing a known degradation: blur, aberration, sub-sampling and missing pixels.

  • Spherical & chromatic aberration Image
  • Motion deblurring Image
  • Random-mask sampling Image
  • Block compressive sensing Image
  • Super-resolution · Inpainting Image
04

Computational Sensing

Non-conventional cameras whose measurements look nothing like the scene until they are inverted.

  • Lensless imaging Image
  • Light-field extrapolation & interpolation Image
  • Event-based reconstruction Image
  • Time-of-flight depth Image
05

Calibration

Recovering a camera's geometry: intrinsics from checkerboards and correction of lens distortion.

  • Lens distortion correction Image
  • Camera calibration Text
→

How models rank

Per-category and per-task scores for every model and protocol, with specialized neural-network baselines alongside.

Open leaderboard

Protocols

Three ways to probe a model

InverseE

Expert

Inverse reconstruction with a fixed, domain-expert prompt for each task. The editor sees the measurement and the expert instruction, and returns the reconstruction.

measurement + expert prompt → editor → image

InverseP

Planner

A frontier VLM first writes an image-specific restoration plan; the same editor then executes it. Comparing with Expert isolates the effect of planning.

measurement → VLM planner → adaptive prompt → editor → image

ForwardF

Forward

The reverse direction: given a clean image, simulate the measurement a physical system would record. Tests forward–inverse consistency on five tasks.

clean image → VLM planner → forward prompt → editor → measurement

Pipeline diagram of the Expert, Planner and Forward protocols: input, guidance, planner, prompt, execution and output stages.
Expert and Planner share the editor, so their difference isolates planning. Editors that share a planner execute identical plans.

Results

Leaderboard

#ModelAccess ExpertPlannerForward
Loading…
Overall = mean unified score (0–100) over a protocol's image tasks; failed or refused jobs score 0. Ranked by Expert. Bold blue = best in column.

The unified score normalizes each signal to [0, 1] (e.g. PSNR / 40, 1 − LPIPS, VLM-judge score / 100) and takes a weighted mean per instance. The weights were fitted to 2,665 blind pairwise judgments by 28 human raters. CGH, ToF depth and Forward outputs, which are not natural photographs, use the unweighted mean of four full-reference metrics.

Per-category & per-task results →

Findings

What we found

01

Specialized models still win on physics

Where a specialized baseline exists (9 of 20 image tasks) it usually leads, most clearly in PSNR: by at least 15 dB on demosaicking, 11 dB on lensless imaging, 8 dB on holography and 6.5 dB on denoising. Super-resolution is the exception, where Gemini beats an ESRGAN reference.

02

Computational sensing is the hardest

Tasks that require inverting a sensing operator fail worst. Under Expert, lensless imaging, event reconstruction and holography average only 7.6, 9.0 and 9.4 dB PSNR; on those inputs models return textures or a different scene altogether.

03

Planning adds cost, not accuracy

Planner guidance changes Overall by −3.2 to +2.3 points, and five of six models lose slightly. The editor matters more than the plan; on the cost frontier Qwen reaches 42.7 at no API cost and Gemini 61.4 at $0.066 per image.

04

Plausible is not faithful

Pixel metrics can hide large failures. On demosaicking, GPT-Image-2 returns grayscale images for 59% of Planner inputs: its judge score halves (68.3 → 32.2) while its PSNR does not move. The planner's step-by-step recipes often never say the output should be in colour.

05

Forward simulation reorders models

Simulating a measurement from a clean image is not easy either. GPT-Image-2 (42.9) and Qwen (42.1) lead, ahead of Gemini (35.8); all six models score between 28.4 and 42.9.

Citation

BibTeX

Citation details will appear here once the preprint is public.