PICABench
Models edit images while preserving physical effects in optics, mechanics, and state transitions. A vision-language judge answers region-specific questions to measure physical correctness.
The evaluator also measures similarity outside edited regions. PICA-100K is a separate training dataset.
- Released
- October 2025
- Kind
- Image generation and editing
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision, Image, video and 3D generation
- Scores refreshed
- November 19th, 2026
In the authors’ words
Image editing has achieved remarkable progress recently. Modern editing models could already follow complex instructions to manipulate the original content. However, beyond completing the editing instructions, the accompanying physical effects are the key to the generation realism. For example, removing an object should also remove its shadow, reflections, and interactions with nearby objects. Unfortunately, existing models and benchmarks mainly focus on instruction completion but overlook these physical effects. So, at this moment, how far are we from physically realistic image editing? To answer this, we introduce PICABench, which systematically evaluates physical realism across eight sub-dimension (spanning optics, mechanics, and state transitions) for most of the common editing operations (add, remove, attribute change, etc.). We further propose the PICAEval, a reliable evaluation protocol that uses VLM-as-a-judge with per-case, region-level human annotations and questions. Beyond benchmarking, we also explore effective solutions by learning physics from videos and construct a training dataset PICA-100K. After evaluating most of the mainstream models, we observe that physical realism remains a challenging problem with large rooms to explore. We hope that our benchmark and proposed solutions can serve as a foundation for future work moving from naive content editing toward physically consistent realism.
Seedream 4.0 leads with 61.9% Seedream 4.0 leads with 61.9%
Best score per entry from the benchmark's own leaderboard, read on November 19th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
- 1
Seedream 4.0 61.9%
- 2 GPT-Image-1 61.1%
- 3 Gemini (image model unspecified) 59.9%
- 4
Qwen-Image-Edit 58.3%
- 5 Flux.1 Kontext + SFT 50.6%
- 6 Flux.1 Kontext 48.9%
- 7
Step1X-Edit 48.2%
- 8 Hidream-E1.1 47.9%
- 9 OmniGen2 46.8%
- 10 Bagel-Think 46.5%
- 11
BAGEL 45.1%
- 12 Uniworld-V1 37.7%
- 13 DiMOO 35.7%
Source: Official PICABench leaderboard
- 1
Seedream 4.0 61.9%
- 2 GPT-Image-1 61.1%
- 3
Qwen-Image-Edit 58.3%
- 4
Step1X-Edit 48.2%
- 5
BAGEL 45.1%
Source: Official PICABench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 61.9% | Official leaderboard | ||
| 2 | GPT-Image-1 | 61.1% | Official leaderboard | ||
| 3 | Gemini (image model unspecified) | 59.9% | Official leaderboard | ||
| 4 | | 58.3% | Official leaderboard | ||
| 5 | Flux.1 Kontext + SFT | 50.6% | Official leaderboard | ||
| 6 | Flux.1 Kontext | 48.9% | Official leaderboard | ||
| 7 | | 48.2% | Official leaderboard | ||
| 8 | Hidream-E1.1 | 47.9% | Official leaderboard | ||
| 9 | OmniGen2 | 46.8% | Official leaderboard | ||
| 10 | Bagel-Think | 46.5% | Official leaderboard | ||
| 11 | | 45.1% | Official leaderboard | ||
| 12 | Uniworld-V1 | 37.7% | Official leaderboard | ||
| 13 | DiMOO | 35.7% | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 61.9% | Official leaderboard | ||
| 2 | GPT-Image-1 | 61.1% | Official leaderboard | ||
| 3 | | 58.3% | Official leaderboard | ||
| 4 | | 48.2% | Official leaderboard | ||
| 5 | | 45.1% | Official leaderboard |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
No dated values yet.
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the official leaderboard (13 entries).
- Pass Models scored 5 models we track, plus 8 other entries on its leaderboard.
- Note Newest score No scores yet.
- Note Frontier models None of its models are on the capabilities index.
- Note Headroom No scores yet.
Leaderboard import 3 pass
- Pass Importer We read the board on picabench.github.io: 5 rows for models we track and 8 other entries at the last read.
- Pass Last read November 19th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.