Aesthetic retouching of scientific visualizations: what is allowed for humans and AI ✛ ✎
Context
In a scientific image designer's studio: on the screen — a 3D model of a planetary collision obtained from a simulation. A couple of clicks and harsh shadows soften, noise is smoothed, and the background darkens, directing the viewer's eye to the shockwave. This digital retouching is familiar to anyone who has prepared an illustration for publication or a press release. But what if the same edits are made not by a human with contextual understanding, but by a generative neural network trained on millions of strangers' images? The study, conducted by an interdisciplinary team, poses the question sharply: where does acceptable aesthetic gloss turn into deception, and does that boundary shift depending on who clicks the 'enhance' button.
Methods
To separate the effect of change type and source (human vs AI), the authors prepared two sets of stimuli. First — an abstract scene with a 3D teapot, into which 15 transformations were manually introduced in Houdini: from presentational (camera position, light, color) to deeper ones affecting data (noise removal, gap filling, shape change). Each change was shown as an animation alternating between the original and modified images. Then 95 experts — researchers in visualization, practicing designers, and domain scientists — rated the acceptability of each edit on a 6-point scale, first assuming a human author, then a generative AI (models Nano Banana Pro and Qwen-Image were used). Additionally, participants examined a realistic case — a planetary collision — and answered an open-ended question about what causes comfort or discomfort when viewing an AI-enhanced visualization.
Results
The numbers paint a clear hierarchy. The most 'forgivable' edits — viewpoint change (average score 3,38 out of 5), smoothing (3,27), and lighting adjustments (3,19). At the opposite pole — moving objects (1,71), removing important details (1,62), adding non-existent features (1,61), and shape distortion (1,40). The paradox of 'smoothing vs blurring': the former changes the data itself but is considered acceptable, while the latter — a purely presentational trick — is rated much more strictly. Thus, what matters is not the level of intervention, but its semantic consequences. AI versions of the same edits systematically received lower scores (a downward shift of 0,5-1 point), especially for previously acceptable changes. However, the ranking was hardly broken: both algorithm and human equally 'fail' on meaning distortion. Open responses clarify: people assume humans have intention and responsibility, while AI has a tendency to hallucinate and lacks context. Interestingly, judgment strictness increases sharply if the final image is called a 'visualization' rather than an 'illustration': the label dictates standards.
Implications
For creators of AI visualization assistants, this means the need for built-in safeguards. Operations that change shape or add features should trigger warnings or be blocked by default. Experts agree: transparency of edits (which ones and why) is more important than explaining how the neural network works. In practice, this could be a panel where each change is documented and can be undone. Also, explicit disclosure of AI use is required — participants ranked this condition as the most important (3,44 out of 4).
Future development
Aesthetic retouching is moving along an adoption curve already traversed in photography: once color correction was considered dishonest, but today smartphone filters are taken for granted. The study captures an early stage of this shift: only 14% of experts are categorically against AI tools, and two are already using them in their work. The next step is to test how non-experts perceive such images and how their trust changes after authorship is revealed. Perhaps 'human-verified' certificates for key scientific publications will emerge.
Impact
Affects scientific communication, popular science media, interface design, regulatory policy on synthetic content.
Next steps
Develop a library of permissible transformations with contextual constraints and conduct longitudinal measurements of public opinion evolution.
Key open problems
Directly touches on the problem of trust in automated systems and 'black boxes' of machine learning, as well as the eternal conflict between rigorous accuracy and visual clarity in scientific illustration.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
A survey of 95 visualization experts assessed the acceptability of 15 types of edits to 3D scenes, performed by either humans or AI. The results show that the semantic impact of an edit plays a decisive role, not the level of intervention, and that AI versions of the same edits are systematically judged more harshly, especially for edits that are initially acceptable.
Where its strength is
Separating the effects of edit type and source (human vs. AI) on identical stimuli isolates the influence of trust in the agent, providing a clear foundation for further research on perceptions of AI-enhanced visualizations.
Where it could go next
- It might be interesting to see how automatic image quality control methods, like artifact detection in MRI, could be applied to identify semantic distortions in scientific visualizations.based on: How a neural network spots flaws in MRI scans arXiv:2605.24609
- A related approach involves neural network self-checking without external prompts: it could be adapted to assess how well AI tools can independently recognize unacceptable changes in visualizations.based on: Round Trip: How a Neural Network Checks Itself Without Hints arXiv:2608.00675v1
- In our archive, there is work on digital twins that predict patient anatomy; its approach to creating realistic models could help develop standardized stimuli for future surveys on visualization perception.based on: Digital Twin Predicts Patient Anatomy arXiv:2608.00831v1
Worth trying
- Test whether evaluation strictness changes when using AI tools with built-in restrictions on edit types.
- Adapt neural network self-check metrics for automatic classification of edits by acceptability.
- Compare perceptions of edits in abstract versus realistic scenes, using digital twins as a source of stimuli.
Why this may matter
Understanding how experts perceive AI-enhanced visualizations is crucial for developing guidelines and tools that maintain trust in scientific imagery in the era of generative models.
Nearby works in our archive
- Atomic Nuclei Offer Clues on How to Squeeze a Neutron Starsimilarity 0.60
- A Quantum Compass for 3D Visionsimilarity 0.58
- Seeing Through Bone Without X-Rayssimilarity 0.57
- Round Trip: How a Neural Network Checks Itself Without Hintssimilarity 0.57
- How the Electron's Shape Controls Gamma Lightsimilarity 0.57
- Digital Twin Predicts Patient Anatomysimilarity 0.57
- How a neural network spots flaws in MRI scanssimilarity 0.56
The archive is sparse around this work: the closest review is only 0.60, against a usual 0.64. It looks like few have drilled nearby.
The same visual result — for example, shifting an object away from the center — was rated differently depending on whether it resulted from a camera rotation (acceptable) or moving the object itself (unacceptable): experts judged by intention, not by the image.
Key numbers
- average acceptability of viewpoint change (human): 3,38 out of 5
- average acceptability of shape distortion (human): 1,40 out of 5
- number of experts: 95
- share of opponents of AI tools: 14%
- experts already using AI for visualization: 2